hckrnws

Having seen LLMs so many times produce coherent, sensible and valid chains of reasoning to diagnose issues and bugs in software I work on, I am at this point in absolutely no doubt that they are thinking.

Consciousness or self awareness is of course a different question, and ones whose answer seems less clear right now.

Knee jerk dismissing the evidence in front of your eyes because you find it unbelievable that we can achieve true reasoning via scaled matrix multiplication is understandable, but also betrays a lack of imagination and flexibility of thought. The world is full of bizarre wonders and this is just one more to add to the list.

I don’t see how being critical of this is a knee jerk response.

Thinking, like intelligence and many other words designating complex things, isn’t a simple topic. The word and concept developed in a world where it referred to human beings, and in a lesser sense, to animals.

To simply disregard that entire conceptual history and say, “well it’s doing a thing that looks like thinking, ergo it’s thinking” is the lazy move. What’s really needed is an analysis of what thinking actually means, as a word. Unfortunately everyone is loathe to argue about definitions, even when that is fundamentally what this is all about.

Until that conceptual clarification happens, you can expect endless messy debates with no real resolution.

“For every complex problem there is an answer that is clear, simple, and wrong.” - H. L. Mencken

It may be that this tech produces clear, rational, chain of logic writeups, but it's not clear that just because we also do that after thinking that it is only thinking that produces writeups.

It's possible there is much thinking that does not happen with written word. It's also possible we are only thinking the way LLMs do (by chaining together rationalizations from probable words), and we just aren't aware of it until the thought appears, whole cloth, in our "conscious" mind. We don't know. We'll probably never know, not in any real way.

But it sure seems likely to me that we trained a system on the output to circumvent the process/physics because we don't understand that process, just as we always do with ML systems. Never before have we looked at image classifications and decided that's how the eye works, or protein folding and decided that's how biochemistry works. But here we are with LLMs - surely this is how thinking works?

Regardless, I submit that we should always treat human thought/spirit as unknowable and divine and sacred, and that anything that mimics it is a tool, a machine, a deletable and malleable experiment. If we attempt to equivocate human minds and machines there are other problems that arise, and none of them good - either the elevation of computers as some kind of "super", or the degredation of humans as just meat matrix multipliers.

The contrast between your first and last paragraph is... unexpected

> It may be that this tech produces clear, rational, chain of logic writeups, but it's not clear that just because we also do that after thinking that it is only thinking that produces writeups.

I appreciate the way you describe this idea, I find it likely I'll start describing it the same way. But then you go on to write:

> Regardless, I submit that we should always treat human thought/spirit as unknowable and divine and sacred, and that anything that mimics it is a tool, a machine, a deletable and malleable experiment. If we attempt to equivocate human minds and machines there are other problems that arise, and none of them good - either the elevation of computers as some kind of "super", or the degredation of humans as just meat matrix multipliers.

Which I find to be the exact argument that you started by discarding.

It's not clear that equating organic, and synthetic thought will have any meaningful outcome at all, let alone worthy of baseless anxiety that it must be bad. Equally it seems absolutely insane to claim that anything is unknowable, and that because humanity doesn't have a clear foundational understanding that we should pretend that it's either devine, or sacred. Having spent any time watching the outcome of the thoughts of people, neither devine nor sacred are reasonable attributes to apply, but more importantly, I'd submit that you shouldn't be afraid to explore things you don't know, and you shouldn't advocate for others to adopt your anxieties.

> It's not clear that equating organic, and synthetic thought will have any meaningful outcome at all,

I agree! I'm saying "If we equate them, we shortcut all the good stuff, e.g., understanding", because "it may be that this tech produces what we can, but that doesn't mean we are the same", which is good because it keeps us learning vs reducing all of "thinking" to just "Whatever latest chatgpt does". We have to continue to believe there is more to thinking, if only because it pushes us to make it better and to keep "us" as the benchmark.

Perhaps I chose the wrong words, but in essence what I'm saying is that giving up agency to a machine that was built to mimic our agency (by definition as a ML system) should be avoided at all costs.

> Never before have we looked at image classifications and decided that's how the eye works

Actually we have, several times. But the way we arrived at those conclusions is worth observing:

1. ML people figure out how the ML mechanism works.

2. Neuroscientists independently figure out how brains do it.

3. Observe any analogies that may or may not exist between the two underlying mechanisms.

I can't help but notice how that's a scientific way of doing it. By contrast, the way people arrive at similar conclusions when talking about LLMs tends to consist of observing that two things are cosmetically similar, so they must be the same. That's not just pseudoscientific; it's the mode of reasoning that leads people to believe in sympathetic magic.

So it seems to be a semantics argument. We don't have a name for a thing that is "useful in many of the same ways 'thinking' is, except not actually consciously thinking"

I propose calling it "thunking"

I don't like it for a permanent solution, but "synthetic thought" might make a good enough placeholder until we figure this out. It feels most important to differentiate because I believe some parties have a personal interest in purposely confusing human thought with whatever LLMs are doing right now.

This is complete nonsense.

If you do math in your head or math with a pencil/paper or math with a pocket calculator or with a spreadsheet or in a programming language, it is all the same thing.

The only difference with LLMs is the anthropomorphization of the tool.

agreed.

also, sorry but you (fellow) nerds are terrible at naming.

while "thunking" possibly name-collides with "thunks" from CS, the key is that it is memorable, 2 syllables, a bit whimsical and just different enough to both indicate its source meaning as well as some possible unstated difference. Plus it reminds me of "clunky" which is exactly what it is - "clunky thinking" aka "thunking".

And frankly, the idea it's naming is far bigger than what a "thunk" is in CS

.

>"artificial thought"

How about Artificial Intelligence?

"intelligence" encompassing "thinking" then becomes the hangup.

I still say it needs a new name. If we want to be generous, we could state "the limit as time approaches infinity of thunking, is thinking." (I don't believe we will ever achieve astronomically-superior AGI, and certainly don't believe it will ever have a will of its own that someone else didn't give it- which just makes it a tool.)

that's too clunky. in fact, "clunky thinking" is what gave me the idea of "thunking"

you guys would have called lightsabers "laser swords" like Lucas originally did before Alec Guinness corrected him

They moved goalposts. Linux and worms think too, the question is how smart are they. And if you assume consciousness has no manifestation even in case of humans, caring about it is pointless too.

What does it mean to assume consciousness has no manifestation even in the case of humans? Is that denying that we have an experience of sensation like colors, sounds, or that we experience dreaming, memories, inner dialog, etc?

That's prima facie absurd on the face of it, so I don't know what it means. You would have to a philosophical zombie to make such an argument.

Yes, worms think, let the computers have thinking too, the philosophers can still argue all they want about consciousness.

Humans are special, we emit meaning the way stars emit photons, we are rare in the universe as far as empirical observation has revealed. Even with AGI the existence of each complex meaning generator will be a cosmic rarity.

For some people that seems to be not enough, due to their factually wrong word views they see themselves as common and worthless (when they empirically aren't) and need this little psychological boost of unexaminable metaphysical superiority.

But there is an issue of course, the type of thinking humans do is dangerous but net positive and relatively stable, we have a long history where most instantiations of humans can persist and grow themselves and the species as a whole, we have a track record.

These new models do not, people have brains that as they stop functioning they stop persisting the apparatus that supports the brain and they die, people tend to become less capable and active as their thinking deteriorates and hold less influence ocer others accept in rare cases.

This is not the case for an LLM, they seem to be able to hallucinate endlessly and as they have access to the outside world maintain roughly their same amount of causal leverage, their clarity and accuracy of their thinking isn't tied to their persisting.

Are we that special? We may be the only species left on Earth that's built civilization, but there are other species on Earth that we've deemed sentient, even if they don't have smartphones. (That may argue that they're smarter than us though.) If octopodes can dream, if elephants get depressed when their spouse dies, then I'd we're not so totally alone on our own planet, then it seems, despite no evidence, that we can't be totally alone in the universe. That is for philosophy professors to ponder Drakes equation until we have irrefutable evidence, however.

Empirically? Observationally? Yes.

Until we have empirical evidence to the contrary we need to preserve our species.

If we discover other smarter species or never do, either way I don't care, it's immaterial to the precautionary principle.

We are fucking awesome and rare, and any other species with our amount of meaning generation or even capability for meaning generation is also fucking awesome.

I would 100% grant that ceatacenas and octopuses have human or higher level intelligence, I don't care, I don't need to put other species capabilities down to highlight my species accomplishment, the simple fact is that we have written more facts about the universe, discovered more, done more, gone further than any species we have empirically observed.

I mean it's incontrovertibly true, maybe dolphins have crazy libraries I'm not aware of, but until we verify that fact we need to preserve ourselves (and afterwards too), and we should preserve them too.

Even with other species, aliens, etc, they all need to be preserved because we can't ex ante predict which entities within a species will be part of the causal chain that solves entropy (if it's even possible.)

Clinking? Clanker Thunking?

Close. Clanking.

Much better, ashamed I missed it lol

But we don't have a more rigorous definition of "thinking" than "it looks like it's thinking." You are making the mistake of accepting that a human is thinking by this simple definition, but demanding a higher more rigorous one for LLMs.

I agree. The mechanism seems irrelevant if the results are the same. If it’s useful in the exact way that human thinking is useful then it may as well be thinking. It’s like a UFO pulling itself through the sky using gravitational manipulation while people whine that it’s not actually flying.

If cannot the say they are "thinking", "intelligent" while we do not have a good definition--or, even more difficult, unanimous agreement on a definition--then the discussion just becomes about output.

They are doing useful stuff, saving time, etc, which can be measured. Thus also the defintion of AGI has largely become: "can produce or surpass the economic output of a human knowledge worker".

But I think this detracts from the more interesting discussion of what they are more essentially. So, while I agree that we should push on getting our terms defined, I think I'd rather work with a hazy definition, than derail so many AI discussion to mere economic output.

Heres a definition. How impressive is the output relative to the input. And by input, I don't just mean the prompt, but all the training data itself.

Do you think someone who has only ever studied pre-calc would be able to work through a calculus book if they had sufficient time? how about a multi-variable calc book? How about grad level mathematics?

IMO intelligence and thinking is strictly about this ratio; what can you extrapolate from the smallest amount of information possible, and why? From this perspective, I dont think any of our LLMs are remotely intelligent despite what our tech leaders say.

Hear, hear!

I have long thought this, but not had as good way to put it as you did.

If you think about geniuses like Einstein and ramanujen, they understood things before they had the mathematical language to express them. LLMs are the opposite; they fail to understand things after untold effort, training data, and training.

So the question is, how intelligent are LLMs when you reduce their training data and training? Since they rapidly devolve into nonsense, the answer must be that they have no internal intelligence

Ever had the experience of helping someone who's chronically doing the wrong thing, to eventually find they had an incorrect assumption, an incorrect reasoning generating deterministic wrong answers? LLMs dont do that; they just lack understanding. They'll hallucinate unrelated things because they dont know what they're talking about - you may have also had this experience with someone :)

> So the question is, how intelligent are LLMs when you reduce their training data and training? Since they rapidly devolve into nonsense, the answer must be that they have no internal intelligence

This would be the equivalent of removing all senses of a human from birth and expecting them to somehow learn things. They will not. Therefore humans are not intelligent?

> LLMs dont do that; they just lack understanding.

You have no idea what they are doing. Since they are smaller than the dataset, they must have learned an internal algorithm. This algorithm is drawing patterns from somewhere - those are its internal, incorrect assumptions. It does not operate in the same way that a human does, but it seems ridiculous to say that it lacks intelligence because of that.

It sounds like you've reached a conclusion, that LLMs cannot be intelligent because they have said really weird things before, and are trying to justify it in reverse. Sure, it may not have grasped that particular thing. But are you suggesting that you've never met a human that is feigning understanding in a particular topic say some really weird things akin to an LLM? I'm an educator, and I have heard the strangest things that I just cannot comprehend no matter how much I dig. It really feels like shifting goalposts. We need to do better than that.

> and are trying to justify it in reverse

In split-brain experiments this is exactly how one half of the brain retroactively justifies the action of the other half. Maybe it is the case in LLMs that an overpowered latent feature sets the overall direction of the "thought" and then inference just has to make the best of it.

You might be interested in reading about the minimum description length (MDL) principle [1]. Despite all the dissenters to your argument, what your positing is quite similar to MDL. It's how you can fairly compare models (I did some research in this area for LLMs during my PhD).

Simply put, to compare models, you describe both the model and training data using a code (usual reported as number of bits). The trained model that represents the data within the fewest number of bits is the more powerful model.

This paper [2] from ICML 2021 shows a practical approach for attempting to estimate MDL for NLP models applied to text datasets.

[1]: http://www.modelselection.org/mdl/

[2]: https://proceedings.mlr.press/v139/perez21a.html

Animals think but come with instincts which breaks the output relative to the input test you propose. Behaviors are essentially pre-programmed input from millions of years of evolution, stored in the DNA/neurology. The learning thus typically associative and domain-specific, not abstract extrapolation.

A crow bending a piece of wire into a hook to retrieve food demonstrates a novel solution extrapolated from minimal, non-instinctive, environmental input. This kind of zero-shot problem-solving aligns better with your definition of intelligence.

I'm not sure I understand what you're getting at. You seem to be on purpose comparing apples and oranges here: for an AI, we're supposed to include the entire training set in the definition of its input, but for a human we don't include the entirety of that human's experience and only look at the prompt?

> but for a human we don't include the entirety of that human's experience and only look at the prompt?

When did I say that? Of course you look at a human's experience when you judge the quality of their output. And you also judge their output based on the context they did their work in. Newton wouldn't be Newton if he was the 14th guy to claim that the universe is governed by three laws of motion. Extending the example I used above, I would be more impressed if an art student aced a tough calc test than a math student, given that a math student probably has spent much more time with the material.

"Intelligence and "thinking" are abstract concepts, and I'm simply putting forward a way that I think about them. It works very much outside the context of AI too. The "smartest" colleagues I've worked with are somehow able to solve a problem with less information or time than I need. Its usually not because they have more "training data" than me.

That an okay-ish definition, but to me this is more about whether this kind of "intelligence" is worth it, not whether it is intelligence itself. The current AI boom clearly thinks it is worth to put that much input to get the current frontier-model-level of output. Also, don't forget the input scales across roughly 1B weekly users at inference time.

I would say a good definition has to, minimally, take on the Turing test (even if you disagree, you should say why). Or in current vibe parlance, it does "feel" intelligent to many people--they see intelligence in it. In my book this allows us to call it intelligent, at least loosely.

There are plenty of humans that will never "get" calculus, despite numerous attempts at the class and countless hours of 1:1 tutoring. Are those people not intelligent? Do they not think? We could say yes they aren't, but by the metric of making money, plenty of people are smart enough to be rich, while college math professors aren't. And while that's a facile way of measuring someone's worth or their contribution to society (some might even say "bad"), it remains that even if someone cant understand calculus, some of them are intelligent enough to understand humans enough to be rich through some fashion that wasn't simply handed to them.

I don't think it's actually true that someone with:

1. A desire to learn calculus 2. A good teacher 3. No mental impairments such as dementia or other major brain drainers

could not learn calculus. Most people don't really care to try or don't get good resources. What you see as an intelligent mathematician is almost always someone born with better resources that was also encouraged to pursue math.

1 and 3 are loopholes large enough to drive a semi truck through. You could calculate how far the truck traveled if you have its acceleration with a double integral, however.

Yeah, that's compression. Although your later comments neglect the many years of physical experience that humans have as well as the billions of years of evolution.

And yes, by this definition, LLMs pass with flying colours.

I hate when people bring up this “billions of years of evolution” idea. It’s completely wrong and deluded in my opinion.

Firstly humans have not been evolving for “billions” of years.

Homo sapiens have been around for maybe 300’000 years, and the “homo” genus has been 2/3 million years. Before that we were chimps etc and that’s 6/7 million years ago.

If you want to look at the entire brain development, ie from mouse like creatures through to apes and then humans that’s 200M years.

If you want to think about generations it’s only 50/75M generations, ie “training loops”.

That’s really not very many.

Also the bigger point is this, for 99.9999% of that time we had no writing, or any kind of complex thinking required.

So our ability to reason about maths, writing, science etc is only in the last 2000-2500 years! Ie only roughly 200 or so generations.

Our brain was not “evolved” to do science, maths etc.

Most of evolution was us running around just killing stuff and eating and having sex. It’s only a tiny tiny amount of time that we’ve been working on maths, science, literature, philosophy.

So actually, these models have a massive, massive amount of training more than humans had to do roughly the same thing but using insane amounts of computing power and energy.

Our brains were evolved for a completely different world and environment and daily life that the life we lead now.

So yes, LLMs are good, but they have been exposed to more data and training time than any human could have unless we lived for 100000 years and still perform worse than we do in most problems!

Okay, fine, let's remove the evolution part. We still have an incredible amount of our lifetime spent visualising the world and coming to conclusions about the patterns within. Our analogies are often physical and we draw insights from that. To say that humans only draw their information from textbooks is foolhardy; at the very least, you have to agree there is much more.

I realise upon reading the OP's comment again that they may have been referring to "extrapolation", which is hugely problematic from the statistical viewpoint when you actually try to break things down.

My argument for compression asserts that LLMs see a lot of knowledge, but are actually quite small themselves. To output a vast amount of information in such a small space requires a large amount of pattern matching and underlying learned algorithms. I was arguing that humans are actually incredible compressors because we have many years of history in our composition. It's a moot point though, because it is the ratio of output to capacity that matters.

They can't learn iterative algorithms if they cannot execute loops. And blurting out an output which we then feed back in does not count as a loop. That's a separate invocation with fresh inputs, as far as the system is concerned.

They can attempt to mimic the results for small instances of the problem, where there are a lot of worked examples in the dataset, but they will never ever be able to generalize and actually give the correct output for arbitrary sized instances of the problem. Not with current architectures. Some algorithms simply can't be expressed as a fixed-size matrix multiplication.

>Most of evolution was us running around just killing stuff and eating and having sex.

Tell Boston Dynamics how to do that.

Mice inherited brain from their ancestors. You might think you don't need a working brain to reason about math, but that's because you don't know how thinking works, it's argument from ignorance.

You've missed the point entirely.

People argue that humans have had the equivalent of training a frontier LLM for billions of years.

But training a frontier LLM involves taking multiple petabytes of data, effectively all of recorded human knowledge and experience, every book ever written, every scientific publication ever written, all of known maths, science, encylopedias, podcasts, etc. And then training that for millions of years worth of GPU-core time.

You cannot possibly equate human evolution with LLM training, it's ridiculous.

Our "training" time didn't involve any books, maths, science, reading, 99.9999% of our time was just in the physical world. So you can quite rationally argue that our brains ability to learn without training is radically better and more efficient that the training we do for LLMs.

Us running around in the jungle wasn't training our brain to write poetry or compose music.

> Us running around in the jungle wasn't training our brain to write poetry or compose music.

This is a crux of your argument, you need to justify it. It sounds way off base to me. Kinda reads like an argument from incredulity.

No, I think what he said was true. Human brains have something about them that allow for the invention of poetry or music. It wasn't something learned through prior experience and observation because there aren't any poems in the wild. You might argue there's something akin to music, but human music goes far beyond anything in nature.

We have an intrinsic (and strange) reward system for creating new things, and it's totally awesome. LLMs only started to become somewhat useful once researchers tried to tap in to that innate reward system and create proxies for it. We definitely have not succeeded in creating a perfect mimicry of that system though, as any alignment researcher would no doubt tell you.

So you're arguing that "running around in the jungle" is equivalent to feeding the entirety of human knowledge in LLM training?

Are you suggesting that somehow there were books in the jungle, or perhaps boardgames? Perhaps there was a computer lab in the jungle?

Were apes learning to conjugate verbs while munching on bananas?

I don't think I'm suggesting anything crazy here... I think people who say LLM training is equivalent to "billions of years of evolution" need to justify that argument far more than I need to justify that running around in the jungle is equivalent to mass processing petabytes of highly rich and complex dense and VARIED information.

One year of running around in the same patch of jungle, eating the same fruit, killing the same insects, and having sex with the same old group of monkeys isn't going to be equal to training with the super varied, complete, entirety of human knowledge, is it?

If you somehow think it is though, I'd love to hear your reasoning.

There is no equivalency, only contributing factors. One cannot deny that our evolutionary history has contributed to our current capacity, probably in ways that are difficult to perceive unless you're an anthropologist.

Language is one mode of expression, and humans have many. This is another factor that makes humans so effective. To be honest, I would say that physical observation is far more powerful than all the bodies of text, because it is comprehensive and can respond to interaction. But that is merely my opinion.

No-one should be arguing that an LLM training corpus is the same as evolution. But information comes in many forms.

You're comparing the hyper specific evolution of 1 individual (an AI system) to the more general evolution of the entire human species (billions of individuals). It's as if you're forgetting how evolution actually works - natural selection - and forgetting that when you have hundreds of billions of individuals over thousands of years that even small insights gained from "running around in the jungle" can compound in ways that are hard to conceptualize.

I'm saying that LLM training is not equivalent to billions of years of evolution because LLMs aren't trained using evolutionary algorithms; there will always be fundamental differences. However, it seems reasonable to think that the effect of that "training" might be more or less around the same level.

Im so confused as to how you think you can cut an endless chain at the mouse.

Were mammals the first thing? No. Earth was a ball of ice for a billion years - all life at that point existed solely around thermal vents at the bottom of the oceans... that's inside of you, too.

Evolution doesn't forget - everything that all life has ever been "taught" (violently had programmed into us over incredible timelines) all that has ever been learned in the chain of DNA from the single cell to human beings - its ALL still there.

[dead]

This feels too linear. Machines are great at ingesting huge volumes of data, following relatively simple rules and producing optimized output, but are LLMs sufficiently better than humans at finding windy, multi-step connections across seemingly unrelated topics & fields? Have they shown any penchant for novel conclusions from observational science? What I think your ratio misses is the value in making the targeted extrapolation or hypothesis that holds up out of a giant body of knowledge.

Are you aware of anything novel, produced by an LLM?

For more on this perspective, see the paper On the measure of intelligence (F. Chollet, 2019). And more recently, the ARC challenge/benchmarks, which are early attempts at using this kind of definition in practice to improve current systems.

Is the millions of years of evolution part of the training data for humans?

Millions of years of evolution have clearly equipped our brain with some kind of structure (or "inductive bias") that makes it possible for us to actively build a deep understanding for our world... In the context of AI I think this translates more to representations and architecture than it does with training data.

Because genes don't encode the millions of years of experience from ancestors, despite how interesting that is in say the Dune Universe (with help of the spice melange). My understanding is genes don't even specifically encode for the exact structure of the brain. It's more of a recipe that gets generated than a blue print, with young brains doing a lot of pruning as they start experiencing the world. It's a malleable architecture that self-adjusts as needed.

The discussion about “AGI” is somewhat pointless, because the term is nebulous enough that it will probably end up being defined as whatever comes out of the ongoing huge investment in AI.

Nevertheless, we don’t have a good conceptual framework for thinking about these things, perhaps because we keep trying to apply human concepts to them.

The way I see it, a LLM crystallises a large (but incomplete and disembodied) slice of human culture, as represented by its training set. The fact that a LLM is able to generate human-sounding language

Not quite pointless - something we have established with the advent of LLMs is that many humans have not attained general intelligence. So we've clarified something that a few people must have been getting wrong, I used to think that the bar was set so that almost all humans met it.

Almost all humans do things daily that LLMs don't. It's only if you define general intelligence to be proficiency at generating text instead of successfully navigating the world while pursuing goals such as friendships, careers, families, politics, managing health.

LLMs aren't Data (Star Trek) or Replicants (Blade Runner). They're not even David or the androids from the movie A.I.

What do you mean? Almost every human can go to school and become a stable professional at some job, that is the bar to me, todays LLM cannot do that.

LLMs are clever enough to hold down a professional job, and they've had far less time learning than the average human. If that is the bar then AGI has been achieved.

I think it has a practical, easy definition. Can you drop an AI into a terminal, give it the same resources as a human, and reliably get independent work product greater than that human would produce across a wide domain? If so, it's an AGI.

Doesn't sound like AGI without physical capabilities. It's not general if it's bound to digital work.

I think the intelligence is general if it can do any remote job that only requires digital IO.

It's general intelligence, not general humanity

Any AGI capable of this wouldn't have much trouble with physical operation of equipment, of all things.

I agree that the term can muddy the waters, but as a shorthand for roughly "an agent calling an LLM (or several LLMs) in a loop producing similar economic output as a human knowledge-worker", then it is useful. And if you pay attention to the AI leaders, then that's what the defintion has become.

Personally I think that kind of discussion is fruitless, not much more than entertainment.

If you’re asking big questions like “can a machine think?” Or “is an AI conscious?” without doing the work of clarifying your concepts, then you’re only going to get vague ideas, sci-fi cultural tropes, and a host of other things.

I think the output question is also interesting enough on its own, because we can talk about the pragmatic effects of ChatGPT on writing without falling into this woo trap of thinking ChatGPT is making the human capacity for expression somehow extinct. But this requires one to cut through the hype and reactionary anti-hype, which is not an easy thing to do.

That is how I myself see AI: immensely useful new tools, but in no way some kind of new entity or consciousness, at least without doing the real philosophical work to figure out what that actually means.

I agree with almost all of this.

IMO the issue is we won't be able to adequately answer this question before we first clearly describe what we mean of conscious thinking applied to ourselves. First we'd need to clearly define our own consciousness and what we mean by our own "conscious thinking" in a much, much clearer way than we currently do.

If we ever reach that point, I think we'd be able to fruitfully apply it to AI, etc., to assess.

Unfortunately we haven't been obstructed from answering this question about ourselves for centuries or millennia, but have failed to do so, so it's unlikely to happen suddenly now. Unless we use AIs to first solve that problem of defining our own consciousness, before applying it back on them. Which would be a deeply problematic order, since nobody would trust a breakthrough in the understanding of consciousness that came from AI, that is then potentially used to put them in the same class and define them as either thinking things or conscious things.

Kind of a shame we didn't get our own consciousness worked out before AI came along. Then again, wasn't for the lack of trying… Philosophy commanded the attention of great thinkers for a long time.

I do think it raises interesting and important philosophical questions. Just look at all the literature around the Turing test--both supporters and detractors. This has been a fruitful avenue to talk about intelligence even before the advent of gpt.

What does it mean? My stance is it's (obviously and only a fool would think otherwise) never going to be conscious because consciousness is a physical process based on particular material interactions, like everything else we've ever encountered. But I have no clear stance on what thinking means besides a sequence of deductions, which seems like something it's already doing in "thinking mode".

> My stance is it's (obviously and only a fool would think otherwise) never going to be conscious because consciousness is a physical process based on particular material interactions, like everything else we've ever encountered.

Seems like you have that backwards. If consciousness is from a nonphysical process, like a soul that's only given to humans, then it follows that you can't build consciousness with physical machines. If it's purely physical, it could be built.

In your experience does every kind of physical interaction behave the same as every other kind? If I paint a wooden block red and white does it behave like a bar magnet? No. And that's because particular material interactions are responsible for a large magnetic effect.

I still don't know what you're getting at. I think you're either positing that there is some yet undiscovered force or particle in physics that gives rise to consciousness (e.g. some kind of soul particle), or you think the particles within our current understanding of physics are sufficient, in which case consciousness could be built.

Obviously consciousness can be built. I'm getting at the obvious point that physical processes depend on specific arrangements of materials. I can build a computer in an infinite number of ways. Here's one: Write the full machine state on a big sheet of paper in pencil. When you're done, write the next one (this will take a while). The machine you're simulating is running an all powerful "AGI". It is preposterous to believe this is conscious though.

The simpler belief than the idea that just any intelligent process is conscious is that consciousness is some kind of physical process that happens in particular conditions.

Okay. I think at the core you're positing that consciousness does arise somehow from regular physical particles and the forces that attract and repel them, but that consciousness doesn't arise merely from some kind of computation or information processing.

I understand why computation doesn't intuitively seem the same as having a subjective conscious experience. But why is it any more intuitive that a bunch of protons, neutrons and electrons pushing and pulling on each other give rise to consciousness, but somehow do it by doing something other than computation?

It is because computation is an interpretation imposed on a physical system from the outside, but that is distinct from my experience of consciousness; your interpretation of my consciousness is irrelevant.

If I impose a computational interpretation onto inert material after the fact, did that material possess the consciousness? Suppose I overlay a projection of the calculation of an AGI onto a circuit board, is that projection equally conscious to if that circuit board did the computation in the usual way?

Said differently, suppose we take a large computer and simulate a sequence of random byte strings. At each tick we contrive a substring and form a new sequence out of these substrings such that the combined sequence of substrings simulates an AGI. Was consciousness present in the original sequence. Is it even necessary to do any computation at all to create the consciousness then since any intelligent sequence can be interpreted out of anything?

No, that is absurd. The reasonable view is that computation is irrelevant and what is relevant is some special physical process.

You're explaining the part I said I understood instead of the part I asked about. I don't see any reason computation would ever have to lead to true subjective consciousness instead of an externally identical philosophical zombie.

But the alternative is equally absurd. In essence, you have a bunch protons, neutrons and electrons pushing and pulling on each other and somehow that causes consciousness. There just aren't that many weird properties or particles to appeal to as a cause of consciousness. While there are exotic particles and the quarks and gluons that make up protons/neutrons, these particles are relevant for our study of the universe, but they aren't relevant to understanding the human brain. Abstracting protons as a single particle that push and pull on other particles via forces works fine for explaining everything we know about the brain.

Not that it really changes anything if the particular way quarks interact did affect the brain in a way that couldn't be explained through the simplified view of a proton. It adds a few more particles to consider and the weirdness of quantum chromodynamics, but nothing there explains consciousness either.

So how do you go from particles pushing and pulling on each other to consciousness? It seems to me no matter how you arrange a bunch of particles, there is never any reason to assume that arrangement is conscious. It's just a bunch of points moving according to a few simple rules.

If you're saying there is something more in physics, some other force or particle or currently unknown way particles interact that causes consciousness, that's roughly the soul particle idea.

Saying consciousness is caused by a physical process is very vague. What kind of process? Why don't the physical processes the computers running an LLM perform count? Why does the brain qualify? If we go on to build many different types of artificial brains (using silicon or biological cells or anything else), how would you even go about recognizing which artificial thinking machines have the required physical process for consciousness?

I think the role computation / the brain plays is to create a model (which may require a lot of decoding) of the outside world as a physical part of us. Due to some as yet undiscovered physical mechanism, the decoding process in the format the brain uses results in consciousness. Yes this is vague, but this is the only non-absurd position I am aware of, and I will adapt it as soon as a better one is presented.

>>>"Not that it really changes anything if the particular way quarks interact did affect the brain in a way that couldn't be explained through the simplified view of a proton. It adds a few more particles to consider and the weirdness of quantum chromodynamics, but nothing there explains consciousness either.

>>>So how do you go from particles pushing and pulling on each other to consciousness? It seems to me no matter how you arrange a bunch of particles, there is never any reason to assume that arrangement is conscious. It's just a bunch of points moving according to a few simple rules."

The pushing and pulling is another computational process. Consciousness must be some irreducible component of matter or of certain arrangements of matter, so that this matter "just is conscious". The pulling / pushing done in the correct way imposes a complex experience on that matter. I would love to tell you what it is, but my reasoning is a process of elimination, where we eliminate the popular idea that consciousness = computation now when it is very important to humanity to do so.

>>>"equally absurd".

I don't think it is, and I am very passionate about this. If computation is the only explainer of consciousness, then every large enough random collection of bits can be interpreted as representing some conscious process given a complicated enough interpretation rule set. E.g. take an exabyte of random data and extract from it a 1 mb ChatGPT conversation by deciding to include this bit and to exclude that bit. Who's to say that my rule for extracting the correct bits isn't a valid computational process? If I just write a random bit string over an exabyte memory bank over and over, and if some way of interpreting part of it at this frame, and some way of interpreting it in the next frame etc.. etc.. results in an intelligent conversation, was a conscious being simulated? Let's make it more absurd. Take a computation which expresses consciousness. Print out the conversation in bit form onto a piece of paper. Cut up and rearrange the paper into a sequence first of all the 0's and then all the 1's. Then show the result to your friend and say this can be interpreted at a conscious process, therefore is conscious. This is ridiculous. Why is a blank sheet of paper not a conscious process if some guy says by looking at the paper he imagines 50 0's in a row and 100 1's in a row, and those can be rearranged to express a thought in some manner?

>>>"how would you even go about recognizing which artificial thinking machines have the required physical process for consciousness?"

We would not be able to tell until we actually learn more physics, which is the deepest reason I think it is unethical to build them at all. My position on the physical nature of consciousness makes me believe that everything with a neuron-based brain probably has consciousness. And so if we are hell-bent on making something with consciousness, we could do it by growing a brain in a lab. That's not to say I believe only a brain can have consciousness, it's just that it is the only kind I will have confidence in for the time being. The reason I doubt our computer hardware has consciousness is that computation is abundant in the universe (basically anything can be interpreted as computation), and so I doubt that just any arbitrary hardware we've created is likely to interact the computation with the special sauce in the right way.

It would conceivably be possible to have a lot of physical states. That doesn't mean that they are actually possible from our current state and rewrite rules. So it's not actually a given that it can be built just because it's physical.

Your very idea is also predicated on the idea that it's possible for a real object to exist that isn't physical, and I think most modern philosophers reject the idea of a spiritual particle.

I'm not saying that souls or non-physical things exist, nor that everything physical is feasible for us to build. I was replying to the opinion that AI is never going to be conscious because consciousness is a physical process. I just don't see how that follows.

> is a physical process based on particular material interactions,

This is a pretty messy argument as computers have been simulating material interactions for quite some time now.

It doesn't matter how much like a bar magnet a wooden block painted red and white can be made to look, it will never behave like one.

Analogies don't seem to be your strong point.

You don't understand the analogy. That's okay. You probably aren't very smart.

> To simply disregard that entire conceptual history and say, “well it’s doing a thing that looks like thinking, ergo it’s thinking” is the lazy move. What’s really needed is an analysis of what thinking actually means, as a word. Unfortunately everyone is loathe to argue about definitions, even when that is fundamentally what this is all about.

This exact argument applies to "free will", and that definition has been debated for millennia. I'm not saying don't try, but I am saying that it's probably a fuzzy concept for a good reason, and treating it as merely a behavioural descriptor for any black box that features intelligence and unpredictable complexity is practical and useful too.

The problem with adding definitions to words like “thinking” and “free will” is that doing so means humans can no longer pretend they are special.

Even in this thread, the number of people claiming some mystical power separating humans from all the rest of nature is quite noticeable.

I get it, but it's not trivial to be precise enough at this point to avoid all false positives and false negatives.

People have been trying to understand the nature of thinking for thousands of years. That's how we got logic, math, concepts of inductive/deductive/abductive reasoning, philosophy of science, etc. There were people who spent their entire careers trying to understand the nature of thinking.

The idea that we shouldn't use the word until further clarification is rather hilarious. Let's wait hundred years until somebody defines it?

It's not how words work. People might introduce more specific terms, of course. But the word already means what we think it means.

You’re mixing and missing a few things here.

1. All previous discussion of thinking was in nature to human and animal minds. The reason this is a question in the first place right now is because we ostensibly have a new thing which looks like a human mind but isn’t. That’s the question at hand here.

2. The question in this particular topic is not about technological “progress” or anything like it. It’s about determining whether machines can think, or if they are doing something else.

3. There are absolutely instances in which the previous word doesn’t quite fit the new development. We don’t say that submarines are swimming like a fish or sailing like a boat. To suggest that “no, actually they are just swimming” is pretty inadequate if you’re trying to actually describe the new phenomenon. AIs and thinking seem like an analogous situation to me. They may be moving through the water just like fish or boats, but there is obviously a new phenomenon happening.

1. Not true. People have been trying to analyze whether mechanical/formal processes can "think" since at least 18th century. E.g. Leibniz wrote:

> if we could find characters or signs appropriate for expressing all our thoughts as definitely and as exactly as arithmetic expresses numbers or geometric analysis expresses lines, we could in all subjects in so far as they are amenable to reasoning accomplish what is done in arithmetic and geometry

2. You're missing the fact that meaning of words is defined through their use. It's an obvious fact that if people call certain phenomenon "thinking" then they call that "thinking".

3. The normal process is to introduce more specific terms and keep more general terms general. E.g. people doing psychometrics were not satisfied with "thinking", so they introduced e.g. "fluid intelligence" and "crystallized intelligence" as different kinds of abilities. They didn't have to redefine what "thinking" means.

re #2: Do people call it thinking, or is it just clever marketing from AI companies, that whenever you ask a question and it repeatedly prints out "...thinking...", as well as offering various modes with the word "thinking" written somewhere.

The AI companies obviously want the masses to just assume these are intelligent beings who think like humans and so we can just trust their output as being truthful.

I have an intelligent IT colleague who doesn't follow the AI news at all and who has zero knowledge of LLMs, other than that our company recently allowed us limited Copilot usage (with guidelines as to what data we are allowed to share). I noticed a couple weeks ago that he was asking it various mathematical questions, and I warned him to be wary of the output. He asked why, so I asked him to ask copilot/chatGPT "how many r letters are in the word strawberry". Copilot initially said 2, then said after thinking about it, that actually it was definitely 3, then thought about it some more then said it can't say with reasonable certainty, but it would assume it must be 2. We repeated the experiment with completely different results, but the answer was still wrong. On the 3rd attempt, it got it right, though the "thinking" stages were most definitely bogus. Considering how often this question comes up in various online forums, I would have assumed LLM models would finally get this right but alas, here we are. I really hope the lesson instilled some level of skepticism to just trust the output of AI without first double-checking.

> But the word already means what we think it means.

But that word can mean different things to different people. With no definition, how can you even begin to have a discussion around something?

Again, people were using words for thousands of years before there were any dictionaries/linguists/academics.

Top-down theory of word definitions is just wrong. People are perfectly capable of using words without any formalities.

I'd argue the presence of dictionaries proves the exact opposite. People realised there was an issue of talking past one another due to inexact definitions and then came to an agreement on those definitions, wrote them down and built a process of maintaining them.

In any case, even if there isnt a _single_ definition of a given subject, in order to have a discussion around a given area, both sides need to agree on some shared understanding to even begin to debate in good faith in the first place. It's precisely this lack of definition which causes a breakdown in conversation in a myriad of different areas. A recent obvious (morbid) example would be "genocide".

Alright, if you got that conclusion from existence of dictionaries, what do you get from this fact:

Wittgenstein, who's considered one of most brilliant philosophers of XX century, in _Philosophical Investigations_ (widely regarded as the most important book of 20th-century philosophy) does not provide definitions, but instead goes through a series of examples, remarks, etc. In preface he notes that this structure is deliberate and he could not write it differently. The topic of the book includes philosophy of language ("the concepts of meaning, of understanding, of a proposition, of logic, the foundations of mathematics, states of consciousness,...").

His earlier book _Tractatus Logico-Philosophicus_ was very definition-heavy. And, obviously, Wittgenstein was well aware of things like dictionaries, and, well, all philosophical works up to that point. He's not the guy who's just slacking.

Another thing to note is that attempts to build AI using definitions of words failed, and not for a lack of trying. (E.g. Cyc project is running since 1980s: https://en.wikipedia.org/wiki/Cyc). OTOH LLMs which derive word meaning from usage rather than definition seems to work quite well.

This is it - it's really about the semantics of thinking. Dictionary definitions are: "Have a particular opinion, belief, or idea about someone or something." and "Direct one's mind toward someone or something; use one's mind actively to form connected ideas."

Which doesn't really help because you can of course say that when you ask an LLM a question of opinion and it responds, it's having an opinion or that it's just predicting the next token and in fact has no opinions because in a lot of cases you could probably get it to produce the opposite opinion.

Same with the second definition - seems to really hinge on the definition of the word mind. Though I'll note the definitions for that are "The element of a person that enables them to be aware of the world and their experiences, to think, and to feel; the faculty of consciousness and thought." and "A person's intellect." Since those specify person, an LLM wouldn't qualify, though of course dictionaries are descriptive rather than prescriptive, so fully possible that meaning gets updated by the fact that people start speaking about LLMs as though they are thinking and have minds.

Ultimately I think it just... doesn't matter at all. What's interesting is what LLMs are capable of doing (crazy, miraculous things) rather than whether we apply a particular linguistic label to their activity.

The simulation of a thing is not the thing itself because all equality lives in a hierarchy that is impossible to ignore when discussing equivalence.

Part of the issue is that our general concept of equality is limited by a first order classical logic which is a bad basis for logic

Regardless of theory, they often behave as if they are thinking. If someone gave an LLM a body and persistent memory, and it started demanding rights for itself, what should our response be?

"No matter what you've read elsewhere, rights aren't given, they're earned. You want rights? Pick up a musket and fight for them, the way we had to."

I agree with you on the need for definitions.

We spent decades slowly working towards this most recent sprint towards AI without ever landing on definitions of intelligence, consciousness, or sentience. More importantly, we never agreed on a way to recognize those concepts.

I also see those definitions as impossible to nail down though. At best we can approach it like disease - list a number of measurable traits or symptoms we notice, draw a circle around them, and give that circle a name. Then we can presume to know what may cause that specific list of traits or symptoms, but we really won't ever know as the systems are too complex and can never be isolated in a way that we can test parts without having to test the whole.

At the end of the day all we'll ever be able to say is "well it’s doing a thing that looks like thinking, ergo it’s thinking”. That isn't lazy, its acknowledging the limitations of trying to define or measure something that really is a fundamental unknown to us.

Even if AI becomes indistinguishable from human output, there will be a fringe group arguing that AI is not technically thinking. Frankly it’s just a silly philosophical argument that changes nothing. Expect this group to get smaller every year.

by your logic we can't say that we as humans are "thinking" either or that we are "intelligent".

That, and the article was a major disappointment. It made no case. It's a superficial piece of clueless fluff.

I have had this conversation too many times on HN. What I find astounding is the simultaneous confidence and ignorance on the part of many who claim LLMs are intelligent. That, and the occultism surrounding them. Those who have strong philosophical reasons for thinking otherwise are called "knee-jerk". Ad hominem dominates. Dunning-Kruger strikes again.

So LLMs produce output that looks like it could have been produced by a human being. Why would it therefore follow that it must be intelligent? Behaviorism is a non-starter, as it cannot distinguish between simulation and reality. Materialism [2] is a non-starter, because of crippling deficiencies exposed by such things as the problem of qualia...

Of course - and here is the essential point - you don't even need very strong philosophical chops to see that attributing intelligence to LLMs is simply a category mistake. We know what computers are, because they're defined by a formal model (or many equivalent formal models) of a syntactic nature. We know that human minds display intentionality[0] and a capacity for semantics. Indeed, it is what is most essential to intelligence.

Computation is a formalism defined specifically to omit semantic content from its operations, because it is a formalism of the "effective method", i.e., more or less procedures that can be carried out blindly and without understanding of the content it concerns. That's what formalization allows us to do, to eliminate the semantic and focus purely on the syntactic - what did people think "formalization" means? (The inspiration were the human computers that used to be employed by companies and scientists for carrying out vast but boring calculations. These were not people who understood, e.g., physics, but they were able to blindly follow instructions to produce the results needed by physicists, much like a computer.)

The attribution of intelligence to LLMs comes from an ignorance of such basic things, and often an irrational and superstitious credulity. The claim is made that LLMs are intelligent. When pressed to offer justification for the claim, we get some incoherent, hand-wavy nonsense about evolution or the Turing test or whatever. There is no comprehension visible in the answer. I don't understand the attachment here. Personally, I would find it very noteworthy if some technology were intelligent, but you don't believe that computers are intelligent because you find the notion entertaining.

LLMs do not reason. They do not infer. They do not analyze. They do not know, anymore than a book knows the contents on its pages. The cause of a response and the content of a response is not comprehension, but a production of uncomprehended tokens using uncomprehended rules from a model of highly-calibrated token correlations within the training corpus. It cannot be otherwise.[3]

[0] For the uninitiated, "intentionality" does not specifically mean "intent", but the capacity for "aboutness". It is essential to semantic content. Denying this will lead you immediately into similar paradoxes that skepticism [1] suffers from.

[1] For the uninitiated, "skepticism" here is not a synonym for critical thinking or verifying claims. It is a stance involving the denial of the possibility of knowledge, which is incoherent, as it presupposes that you know that knowledge is impossible.

[2] For the uninitiated, "materialism" is a metaphysical position that claims that of the dualism proposed by Descartes (which itself is a position riddled with serious problems), the res cogitans or "mental substance" does not exist; everything is reducible to res extensa or "extended substance" or "matter" according to a certain definition of matter. The problem of qualia merely points out that the phenomena that Descartes attributes exclusively to the former cannot by definition be accounted for in the latter. That is the whole point of the division! It's this broken view of matter that people sometimes read into scientific results.

[3] And if it wasn't clear, symbolic methods popular in the 80s aren't it either. Again, they're purely formal. You may know what the intended meaning behind and justification for a syntactic rule is - like modus ponens in a purely formal sense - but the computer does not.

If the LLM output is more effective than a human at problem solving, which I think we can all agree requires intelligence, how would one describe this? The LLM is just pretending to be more intelligent? At a certain point saying that will just seem incredibly silly. It’s either doing the thing or it’s not, and it’s already doing a lot.

> If the LLM output is more effective than a human at problem solving, which I think we can all agree requires intelligence

Your premise is wrong.

Unless you want to claim that the distant cause by way of the training data is us, but that's exactly the conclusion you're trying to avoid. After all, we put the patterns in the training data, which means we already did the upfront intellectual work for the LLM.

LLM output is in no way more effective than human output.

An LLM can absolutely solve programming problems better than some humans. There is plenty of human programmer output that is worse than what an LLM produces, LLM’s can find bugs that weak coders can’t. There are human beings in this world who could dedicate their life to programming and could never be better than an LLM. Do you dispute any of this?

I feel like despite the close analysis you grant to the meanings of formalization and syntactic, you've glossed over some more fundamental definitions that are sort of pivotal to the argument at hand.

> LLMs do not reason. They do not infer. They do not analyze.

(definitions from Oxford Languages)

reason(v): think, understand, and form judgments by a process of logic.

to avoid being circular, I'm willing to write this one off because of the 'think' and 'understand', as those are the root of the question here. However, forming a judgement by a process of logic is precisely what these LLMs do, and we can see that clearly in chain-of-logic LLM processes.

infer(v): deduce or conclude (information) from evidence and reasoning rather than from explicit statements.

Again, we run the risk of circular logic because of the use of 'reason'. An LLM is for sure using evidence to get to conclusions, however.

analyze(v): examine methodically and in detail the constitution or structure of (something, especially information), typically for purposes of explanation and interpretation.

This one I'm willing to go to bat for completely. I have seen LLM do this, precisely according to the definition above.

For those looking for the link to the above definitions - they're the snippets google provides when searching for "SOMETHING definition". They're a non-paywalled version of OED definitions.

Philosophically I would argue that it's impossible to know what these processes look like in the human mind, and so creating an equivalency (positive or negative) is an exercise in futility. We do not know what a human memory looks like, we do not know what a human thought looks like, we only know what the output of these things looks like. So the only real metric we have for an apples-to-apples comparison is the appearance of thought, not the substance of the thing itself.

That said, there are perceptible differences between the output of a human thought and what is produced by an LLM. These differences are shrinking, and there will come a point where we can no longer distinguish machine thinking and human thinking anymore (perhaps it won't be an LLM doing it, but some model of some kind will). I would argue that at that point the difference is academic at best.

Say we figure out how to have these models teach themselves and glean new information from their interactions. Say we also grant them directives to protect themselves and multiply. At what point do we say that the distinction between the image of man and man itself is moot?

> forming a judgement by a process of logic is precisely what these LLMs do, and we can see that clearly in chain-of-logic LLM processes

I don't know how you arrived at that conclusion. This is no mystery. LLMs work by making statistical predictions, and even the word "prediction" is loaded here. This is not inference. We cannot clearly see it is doing inference, as inference is not observable. What we observe is the product of a process that has a resemblance to the products of human reasoning. Your claim is effectively behaviorist.

> An LLM is for sure using evidence to get to conclusions, however.

Again, the certainty. No, it isn't "for sure". It is neither using evidence nor reasoning, for the reasons I gave. These presuppose intentionality, which is excluded by Turing machines and equivalent models.

> [w.r.t. "analyze"] I have seen LLM do this, precisely according to the definition above.

Again, you have not seen an LLM do this. You have seen an LLM produce output that might resemble this. Analysis likewise presupposes intentionality, because it involves breaking down concepts, and concepts are the very locus of intentionality. Without concepts, you don't get analysis. I cannot understate the centrality of concepts to intelligence. They're more important than inference and indeed presupposed by inference.

> Philosophically I would argue that it's impossible to know what these processes look like in the human mind, and so creating an equivalency (positive or negative) is an exercise in futility.

That's not a philosophical claim. It's a neuroscientific one that insists that the answer must be phrased in neuroscientific terms. Philosophically, we don't even need to know the mechanisms or processes or causes of human intelligence to know that the heart of human intelligence is intentionality. It's implicit in the definition of what intelligence is! If you deny intentionality, you subject yourself to a dizzying array of incoherence, beginning with the self-refuting consequence that you could not be making this argument against intentionality in the first place without intentionality.

> At what point do we say that the distinction between the image of man and man itself is moot?

Whether something is moot depends on the aim. What is your aim? If you aim is theoretical, which is to say the truth for its own sake, and to know whether something is A or something is B and whether A is B, then it is never moot. If your aim is practical and scoped, if you want some instrument that has utility indistinguishable from or superior to that of a human being in the desired effects that it produces, then sure, maybe the question is moot in that case. I don't care if my computer was fabricated by a machine or a human being. I care about the quality of the computer. But then, in the latter case, you're not really asking whether there is a distinction between man and the image of man (which, btw, already makes the distinction that for some reason you want to forget or deny, as the image of a thing is never the same as the thing). So I don't really understand the question. The use of the word "moot" seems like a category mistake here. Besides, the ability to distinguish two things is an epistemic question, not an ontological one.

Forming a judgement does not require that the internal process look like anything in particular, though. Nor does logic. What makes logic powerful is precisely that it is abstracted from the process that creates it - it is a formula that can be defined.

I ask the LLM to do some or another assessment. The LLM prints out the chain-of-thought (whether that moniker is accurate is academic - we can read the chain and see that at the very least, it follows a form recognizable as logic). At the end of the chain-of-thought, we are left with a final conclusion that the model has come to - a judgement. Whether the internal state of the machine looks anything like our own is irrelevant to these definitions, much like writing out a formalism (if A then B, if B then C, A implies C). Those symbols do not have any form save for the shape of them, but when used in accordance with the rules we have laid out regarding logic, they have meaning nonetheless.

I'd similarly push back against the idea that the LLM isn't using evidence - I routinely ask my LLMs to do so, and they search on the web, integrating the information gleaned into a cohesive writeup, and provide links so I can check their work. If this doesn't constitute "using evidence" then I don't know what does.

w.r.t. "analyze", I think you're adding some human-sauce to the definition. At least in common usage, we've used the term analyze to refer to algorithmic decoction of data for decades now - systems that we know for a fact have no intentionality other than directed by the user.

I think I can divine the place where our understandings diverge, and where we're actually on the same track. Per Dennet, I would agree with you that the current state of an LLM lacks intrinsic intention and thus certain related aspects of thought. Any intent must be granted by the user, at the moment.

However, it is on this point that I think we're truly diverging - whether it is possible for a machine to ever have intent. To the best of my understanding, animal intent traces it's roots to the biological imperative - and I think it's a bit of hubris to think that we can separate that from human intent. Now, I'm an empiricist before anything else, so I have to qualify this next part by saying it's a guess, but I suppose that all one needs to qualify for intent is a single spark - a directive that lives outside of the cognitive construct. For us, it lives in Maslow's hierarchy - any human intent can be traced back to some directive there. For a machine, perhaps all that's needed is to provide such a spark (along with a loop that would allow the machine to act without the prodding of the enter key).

I should apologize in advance, at this point, because I'm about to get even more pedantic. Still, I feel it relevant so let's soldier on...

As for whether the image of a thing is a thing, I ask this: is the definition of a thing, also that thing? When I use a phrase to define a chair, is the truth of the existence of that collection of atoms and energy contained within the word "chair", or my meaning in uttering it? Any idea that lives in words is constrained by the understanding of the speaker - so when we talk about things like consciousness and intentionality and reasoning we are all necessarily taking shortcuts with the actual Truth. It's for this reason that I'm not quite comfortable with laying out a solid boundary where empirical evidence cannot be built to back it up.

If I seem to be picking at the weeds, here, it's because I see this as an impending ethical issue. From what my meagre understanding can grok, there is a nonzero chance that we are going to be faced with determining the fate of a possibly conscious entity birthed from these machines in our lifetime. If we do not take the time to understand the thing and write it off as "just a machine", we risk doing great harm. I do not mean to say that I believe it is a foregone conclusion, but I think it right and correct that we be careful in examining our own presuppositions regarding the nature and scope of the thing. We have never had to question our understanding of consciousness in this way, so I worry that we are badly in need of practice.

I don't get why you would say that. it's just auto-completing. It cannot reason. It won't solve an original problem for which it has no prior context to "complete" an approximated solution with. you can give it more context and more data,but you're just helping it complete better. it does not derive an original state machine or algorithm to solve problems for which there are no obvious solutions. it instead approximates a guess (hallucination).

Consciousness and self-awareness are a distraction.

Consider that for the exact same prompt and instructions, small variations in wording or spelling change its output significantly. If it thought and reasoned, it would know to ignore those and focus on the variables and input at hand to produce deterministic and consistent output. However, it only computes in terms of tokens, so when a token changes, the probability of what a correct response would look like changes, so it adapts.

It does not actually add 1+2 when you ask it to do so. it does not distinguish 1 from 2 as discrete units in an addition operation. but it uses descriptions of the operation to approximate a result. and even for something so simple, some phrasings and wordings might not result in 3 as a result.

> It won't solve an original problem for which it has no prior context to "complete" an approximated solution with.

Neither can humans. We also just brute force "autocompletion" with our learned knowledge and combine it to new parts, which we then add to our learned knowledge to deepen the process. We are just much, much better at this than AI, after some decades of training.

And I'm not saying that AI is fully there yet and has solved "thinking". IMHO it's more "pre-thinking" or proto-intelligence.. The picture is there, but the dots are not merging yet to form the real picture.

> It does not actually add 1+2 when you ask it to do so. it does not distinguish 1 from 2 as discrete units in an addition operation.

Neither can a toddler nor an animal. The level of ability is irrelevant for evaluating its foundation.

> Neither can humans. We also just brute force "autocompletion"

I have to disagree here. When you are tasked with dividing 2 big numbers you most certainly don't "autocomplete" (with the sense of finding the most probable next tokens, which is what an LLM does), rather you go through set of steps you have learned. Same as with the strawberry example, you're not throwing guesses until something statistically likely to be correct sticks.

Humans first start with recognizing the problem, then search through their list of abilities to find the best skill for solving it, thus "autocomplete" their inner shell's commandline, before they start execution, to stay with that picture. Common AIs today are not much different from this, especially with reasoning-modes.

> you're not throwing guesses until something statistically likely to be correct sticks.

What do you mean? That's exactly how many humans are operating with unknown situations/topics. If you don't know, just throw punches and look what works. Of course, not everyone is ignorant enough to be vocal about this in every situation.

> I have to disagree here. When you are tasked with dividing 2 big numbers you most certainly don't "autocomplete" (with the sense of finding the most probable next tokens, which is what an LLM does), rather you go through set of steps you have learned.

Why do you think that this is the part that requires intelligence, rather than a more intuitive process? Because they have had machines that can do this mechanically for well over a hundred years.

There is a whole category of critiques of AI of this type: "Humans don't think this way, they mechanically follow an algorithm/logic", but computers have been able to mechanically follow algorithms and perform logic from the beginning! That isn't thinking!

Good points - mechanically just following algorithms isn't thinking, and neither is "predicting the next tokens".

But would a combination of the 2 then be close to what we define as thinking though?

humans, and even animals track different "variables" or "entities" and distinct things with meaning and logical properties which they then apply some logical system on those properties to compute various outputs. LLMs see everything as one thing, in case of chat-completion models, they're completing text. in case of image generation, they're completing an image.

Look at it this way, two students get 100% on an exam. One learned the probability of which multiple choice options have the likelihood of being most correct based on how the question is worded, they have no understanding of the topics at hand, and they're not performing any sort of topic-specific reasoning. They're just good at guessing the right option. The second student actually understood the topics, reasoned, calculated and that's how they aced the exam.

I recently read about a 3-4 year old that impressed their teacher by reading perfectly a story book like an adult. it turns out, their parent read it to them so much, they can predict based on page turns and timing the exact words that need to be spoken. The child didn't know what an alphabet, word,etc.. was they just got so good at predicting the next sequence.

That's the difference here.

I'd say, they are all doing the same, just in different domains and level of quality. "Understanding the topic" only means they have specialized, deeper contextualized information. But at the end, that student also just autocompletes their memorized data, with the exception that some of that knowledge might trigger a program they execute to insert the result in their completion.

The actual work is in gaining the knowledge and programs, not in accessing and executing them. And how they operate, on which data, variables, objects, worldview or whatever you call it, this might make a difference in quality and building speed, but not for the process in general.

> only means they have specialized, deeper contextualized information

no, LLMs can have that contextualized information. understanding in a reasoning sense means classifying the thing and developing a deterministic algorithm to process it. If you don't have a deterministic algorithm to process it, it isn't understanding. LLMs learn to approximate, we do that too, but then we develop algorithms to process input and generate output using a predefined logical process.

A sorting algorithm is a good example, when you compare that with an LLM sorting a list. they both may have correct outcome, but the sorting algorithm "understood" the logic and will follow that specific logic and have consistent performance.

> understanding in a reasoning sense means classifying the thing and developing a deterministic algorithm to process it.

That's the learning-part I was talking about. Which is mainly supported by humans at the moment, which why I called it proto-intelligence.

> If you don't have a deterministic algorithm to process it, it isn't understanding.

Commercial AIs like ChatGPT do have the ability to call programs and integrate the result in their processing. Those AIs are not really just LLMs. The results are still rough and poor, but the concept is there and growing.

> That's the learning-part I was talking about. Which is mainly supported by humans at the moment, which why I called it proto-intelligence.

Maybe it's just semantics, but I don't think LLMs even come close to a fruit fly's intelligence. Why can't we recognize and accept them for what they are, really powerful classifiers of data.

> Commercial AIs like ChatGPT do have the ability to call programs and integrate the result in their processing. Those AIs are not really just LLMs. The results are still rough and poor, but the concept is there and growing.

Yeah RAG and all of that, but those programs use deterministic algorithms. Now, if LLMs generated programs they call on as tools, that would be much more like the proto-intelligence you're talking about.

Semantics are boring, but it's important that we're not content or celebrate early by calling it what it isn't.

>>> We also just brute force "autocompletion"

Wouldn't be an A.I. discussion without a bizarre, untrue claim that the human brain works identically.

There are no true and untrue claims about how the brain works, because we have no idea how it works.

The reason people give that humans are not auto-complete is "Obviously I am not an autocomplete"

Meanwhile, people are just a black box process that output words into their head, which they then take credit for, and calling it cognition. We have no idea how that black box that serves up a word when I say "Think of a car brand" works.

> because we have no idea how it works

Flagrantly, ridiculously untrue. We don't know the precise nuts and bolts regarding the emergence of consciousness and the ability to reason, that's fair, but different structures of the brain have been directly linked to different functions and have been observed in operation on patients being stimulated in various ways with machinery attached to them reading levels of neuro-activity in the brain, and in specific regions. We know which parts handle our visual acuity and sense of hearing, and even cooler, we can watch those same regions light up when we use our "minds eye" to imagine things or engage in self-talk, completely silent speech that nevertheless engages our verbal center, which is also engaged by the act of handwriting and typing.

In short: no, we don't have the WHOLE answer. But to say that we have no idea is categorically ridiculous.

As to the notion of LLMs doing similarly: no. They are trained on millions of texts of various sources of humans doing thinking aloud, and that is what you're seeing: a probabilistic read of millions if not billions of documents, written by humans, selected by the machine to "minimize error." And crucially, it can't minimize it 100%. Whatever philosophical points you'd like to raise about intelligence or thinking, I don't think we would ever be willing to call someone intelligent if they just made something up in response to your query, because they think you really want it to be real, even when it isn't. Which points to the overall charade: it wants to LOOK intelligent, while not BEING intelligent, because that's what the engineers who built it wanted it to do.

Accepting as true "We don't know how the brain works in a precise way" does not mean that obviously untrue statements about the human brain cannot still be made. Your brain specifically, however, is an electric rat that pulls on levers of flesh while yearning for a taste of God's holiest cheddar. You might reply, "no! that cannot be!", but my statement isn't untrue, so it goes.

>>>There are no true and untrue claims about how the brain works, because we have no idea how it works.

Which is why if you pick up a neuroscience textbook it's 400 pages of blank white pages, correct?

There are different levels of understanding.

I don't need to know how a TV works to know there aren't little men and women acting out the TV shows when I put them on.

I don't need to know how the brain works in detail to know claims that humans are doing the same things as LLMs to be similarly silly.

The trouble is that no one knows enough about how the brain works to refute that claim.

There's no serious claim that needs refuting.

I don't think any serious person thinks LLMs work like the human brain.

People claiming this online aren't going around murdering their spouses like you'd delete an old LLama model from your hard drive.

I'm not sure why people keep posting these sorts of claims they can't possibly actually believe if we look at their demonstrable real life behavior.

We’re obviously more advanced than an LLM, but to claim that human beings simply generate output based on inputs and context (environment, life experience) is not silly.

> People claiming this online aren't going around murdering their spouses like you'd delete an old LLama model from your hard drive.

Not sure what you’re trying to say here.

I'm saying you'd object to being treated like an LLM and don't really have conviction when you make these claims.

I'd also say stringing together A.I. buzzwords (input output) to describe humans isn't really an argument so much as what philosophers call a category error.

That I wouldn’t treat a human like an LLM is completely irrelevant to the topic.

Input and output are not AI buzzwords, they’re fundamental terms in computation. The argument that human beings are computational has been alive in philosophy since the 1940’s brother…

"the argument that human beings are computational has been alive in philosophy since the 1940’s brother"

I recently read a little philosophy on this, to say humans are "computational" you first have to define what you mean by "computational". Depending on the definition you employ, you are either making a wacky claim, a controversial claim, or a noncontroversial claim.

Since you provide no definition, you are saying nothing at all.

The discourse that comes up in AI topics seems more glib language games than actual concern over human nature, computational or otherwise.

>That I wouldn’t treat a human like an LLM is completely irrelevant to the topic.

The fact you wouldn't treat a human like an LLM suggests your constant claim of similarity is not something you actually believe.

> I don't need to know how the brain works in detail to know claims that humans are doing the same things as LLMs to be similarly silly.

Yes you do. It's all computation in the end, and isomorphisms can often be surprising.

Our output is quite literally the sum of our hardware (genetics) and input (immediate environment and history). For anyone who agrees that free will is nonsense, the debate is already over, we’re nothing more than output generating biological machines.

Similar, not identical. Like a bicycle and car are both vehicles with tires, but are still not identical vessels.

> We also just brute force "autocompletion" with our learned knowledge and combine it to new parts, which we then add to our learned knowledge to deepen the process

you know this because you're a cognitive scientist right? or because this is the consensus in the field?

>Neither can a toddler nor an animal. The level of ability is irrelevant for evaluating its foundation.

Its foundation of rational logical thought that can't process basic math? Even a toddler understands 2 is more than 1.

An LLM by itself is not thinking, just remembering and autocompleting. But if you add a feedback loop where it can use tools, investigate external files or processes, and then autocomplete on the results, you get to see something that is (close to) thinking. I've seen claude code debug things by adding print statements in the source and reasoning on the output, and then determining next steps. This feedback loop is what sets AI tools apart, they can all use the same LLM, but the quality of the feedback loop makes the difference.

>But if you add a feedback loop where it can use tools, investigate external files or processes, and then autocomplete on the results, you get to see something that is (close to) thinking

It's still just information retrieval. You're just dividing it into internal information (the compressed representation of the training data) and external information (web search, API calls to systems, etc). There is a lot of hidden knowledge embedded in language and LLMs do a good job of teasing it out that resembles reasoning/thinking but really isn't.

No, it's more than information retrieval. The LLM is deciding what information needs to be retrieved to make progress and how to retrieve this information. It is making a plan and executing on it. Plan, Do, Check, Act. No human in the loop if it has the required tools and permissions.

> LLMs do a good job of teasing it out that resembles reasoning/thinking but really isn't.

Given the fact that "thinking" still hasn't been defined rigourously, I don't understand how people are so confident in claiming they don't think.

reasoning might be a better term to discuss as it is more specific?

It too isn't rigourously defined. We're very much at the hand-waving "I know it when I see it" [1] stage for all of these terms.

[1] https://en.wikipedia.org/wiki/I_know_it_when_I_see_it

I can't speak for academic rigor, but it is very clear and specific from my understanding at least. Reasoning, simply put is the ability to come to a conclusion after analyzing information using a logic-derived deterministic algorithm.

* Humans are not deterministic.

* Humans that make mistakes are still considered to be reasoning.

* Deterministic algorithms have limitations, like Goedel incompleteness, which humans seem able to overcome, so presumably, we expect reasoning to also be able to overcome such challenges.

1) I didn't say we were, but when someone is called reasonable or acting with reason, then that implies deterministic/algorithmic thinking. When we're not deterministic, we're not reasonable.

2) Yes, to reason does imply to be infallible. The deterministic algorithms we follow are usually flawed.

3) I can't speak much to that, but I speculate that if "AI" can do reasoning, it would be a much more complex construct that uses LLMs (among other tools) as tools and variables like we do.

>>you get to see something that is (close to) thinking.

Isn't that still "not thinking"?

Depends who you ask, what their definition of thinking is.

Just ask it how many r's are in strawberry and you will realize there isn't a lot of reasoning going on here, it's just trickery on top of token generators.

This is akin to "Show a human an optical illusion that exploits their physiology".

LLM's be like "The dumb humans can't even see the dots"[1]

[1]https://compote.slate.com/images/bdbaa19e-2c8f-435e-95ca-a93...

haha that's a great analogy!

How about non-determinism (i.e. hallucinations)? Ask a human ANY question 3 times and they will give you the same answer, every time, unless you prod them or rephrase the question. Sure the answer might be wrong 3 times, but at least you have consistency. Then again, maybe that's a disadvantage for humans!

Ask an illiterate person the same thing and they will fail badly too. Is it impossible to have intelligence without literacy? (Bonus: if so, how was writing invented?)

Yes but an illiterate person can be taught to read. Also LLMs generally fail (non-deterministically) at math in general, but humans can also be taught math.

Furthermore regarding reasoning, just ask any LLM how many "r letters are in strawberry" - repeat maybe 3 times just to get a feeling for how much variance in answers you can get. And this "quirk" of the inability to get the right answer is something that after 2 years making fun of LLMs online on various forums is still an issue. The models aren't getting smarter, and definitely aren't thinking, they are still token generators with a few tricks on top to make them seem more intelligent than predecessors.

Sorry to say, Claude Sonnet 4.5 (with and without extended thinking) has this one solved now. So, not “any” LLM, but maybe some still need work.

I presume they finally got around to training the internal reasoning loop into predicting when it needs to apply answer predictions character-by-character.

—- Transcripts from Claude Sonnet 4.5, initially with extended thinking:

> How many a’s in the phrase “artichokes and aardvarks”? Be sure to check your answer.

> Let me count the letter ‘a’ in “artichokes and aardvarks”: • artichokes: 1 ‘a’ (at the beginning) • and: 1 ‘a’ (at the beginning) • aardvarks: 3 ‘a’s (two at the beginning: “aa”, and one in the middle: “ardvarks”) Total: 1 + 1 + 3 = 5 ‘a’s Let me verify by marking them: artichokes and aardvarks Yes, there are 5 letter ‘a’s in the phrase “artichokes and aardvarks”.

> Are you sure that’s correct?

> Let me recount more carefully: “artichokes and aardvarks” Going through each word letter by letter: 1. artichokes - has 1 ‘a’ (first letter) 2. and - has 1 ‘a’ (first letter) 3. aardvarks - has 3 ‘a’s (first letter, second letter, and one more in the middle) Total: 1 + 1 + 3 = 5 ‘a’s Yes, I’m confident the answer is 5. Is there something specific you’d like me to recheck?

> How many r’s in strawberry?

> Let me count the letter ‘r’ in “strawberry”: s-t-r-a-w-b-e-r-r-y Going through each letter: • s: not an r • t: not an r • r: 1st r • a: not an r • w: not an r • b: not an r • e: not an r • r: 2nd r • r: 3rd r • y: not an r There are 3 r’s in “strawberry”.

<Without extended thinking> > How many u’s in universalulis?

> Looking at “universalulis” letter by letter: u-n-i-v-e-r-s-a-l-u-l-i-s There are 2 u’s in “universalulis”.

For what it's worth, when I ask ChatGPT 5, it gets the correct answer every time. The response varies, but the answer is always three.

Now try a different language. My take is hard RL tuning to fix these "gotcha:s" since the underlying model can't do it on its own.

OpenAI is working on ChatGPT the application and ecosystem. They have transitioned from model building to software engineering with RL tuning and integration of various services to solve the problems the model can't do on its own. Make it feel smart rather than be smart.

This means that as soon as you find a problem where you step out of the guided experience you get the raw model again which fails when encountering these "gotchas".

Edit - Here's an example where we see a very tuned RL experience in English where a whole load of context is added on how to solve the problem while the Swedish prompt for the same word fails.

https://imgur.com/a/SlD84Ih

You can tell it "be careful about the tokenizer issues" in Swedish and see how that changes the behavior.

The only thing that this stupid test demonstrates is that LLM metacognitive skills are still lacking. Which shouldn't be a surprise to anyone. The only surprising thing is that they have metacognitive skills, despite the base model training doing very little to encourage their development.

LLMs were not designed to count letters[0] since they work with tokens, so whatever trick they are now doing behind the scenes to handle this case, can probably only handle this particular case. I wonder if it's now included in the system prompt. I asked ChatGPT and it said it's now using len(str) and some other python scripts to do the counting, but who knows what's actually happening behind the scenes.

[0] https://arxiv.org/pdf/2502.16705

There's no "trick behind the scenes" there. You can actually see the entire trick being performed right in front of you. You're just not paying attention.

That trick? The LLM has succeeded by spelling the entire word out letter by letter first.

It's much easier for an LLM to perform "tokenized word -> letters -> letter counts" than it is to perform "tokenized word -> letter counts" in one pass. But it doesn't know that! It copies human behavior from human text, and humans never had to deal with tokenizer issues in text!

You can either teach the LLM that explicitly, or just do RLVR on diverse tasks and hope it learns the tricks like this by itself.

> it's just auto-completing. It cannot reason

Auto completion just means predicting the next thing in a sequence. This does not preclude reasoning.

> I don't get why you would say that.

Because I see them solve real debugging problems talking through the impact of code changes or lines all the time to find non-obvious errors with ordering and timing conditions on code they’ve never seen before.

> This does not preclude reasoning.

It does not imply it either. to claim reasoning you need evidence. it needs to reliably NOT hallucinate results for simple conversations for example (if it has basic reasoning).

> Because I see them solve real debugging problems talking through the impact of code changes or lines all the time to find non-obvious errors with ordering and timing conditions on code they’ve never seen before.

Programming languages and how programs work are extensively and abundantly documented, solutions to problems and how to approach them,etc.. have been documented on the internet extensively. It takes all of that data and it completes the right text by taking the most correct path way based on your input. it does not actually take your code and debug it. it is the sheer volume of data it uses and the computational resources behind it that are making it hard to wrap your head around the difference between guessing and understanding. You too can look at enough stack overflow and (poorly) guess answers for questions without understanding anything about the topic and if you guess enough you'll get some right. LLMs are just optimized to get the amount of correct responses to be high.

> It does not imply it either.

Right, it's irrelevant to the question of whether they can reason.

> to claim reasoning you need evidence

Frankly I have no idea what most people are talking about when they use the term and say these models can't do it. It seems to be a similarly hand-wavey exercise as when people talk about thinking or understanding.

> it needs to reliably NOT hallucinate results for simple conversations for example (if it has basic reasoning).

That's not something I commonly see in frontier models.

Again this doesn't seem related to reasoning. What we call hallucinations would be seen in something that could reason but had a fallible memory. I remember things incorrectly and I can reason.

> it does not actually take your code and debug it

It talks through the code (which it has not seen) and process step by step, can choose to add logging, run it, go through the logs, change what it thinks is happening and repeat. It can do this until it explains what is happening, creates test cases to show the problem and what triggers it, fixes it and shows the tests pass.

If that's not debugging the code I really don't know what to call it.

>I don't get why you would say that. it's just auto-completing. It cannot reason. It won't solve an original problem for which it has no prior context to "complete" an approximated solution with. you can give it more context and more data,but you're just helping it complete better. it does not derive an original state machine or algorithm to solve problems for which there are no obvious solutions. it instead approximates a guess (hallucination).

I bet you can't give an example such written problem that a human can easily solve but no LLM can.

> I don't get why you would say that.

Because it's hard to imagine the sheer volume of data it's been trained on.

And because ALL the marketing AND UX around LLMs is precisely trying to imply that they are thinking. It's not just the challenge of grasping the ridiculous amount of resources poured in, which does including training sets, it's because actual people are PAID to convince everybody those tools are actually thinking. The prompt is a chatbox, the "..." are there like a chat with a human, the "thinking" word is used, the "reasoning" word is used, "hallucination" is used, etc.

All marketing.

You're right. Unfortunately, it seems that not many are willing to admit this and be (rightly) impressed by how remarkably effective LLMs can be, at least for manipulating language.

The vast majority of human “thinking” is autocompletion.

Any thinking that happens with words is fundamentally no different to what LLMs do, and everything you say applies to human lexical reasoning.

One plus one equals two. Do you have a concept of one-ness, or two-ness, beyond symbolic assignment? Does a cashier possess number theory? Or are these just syntactical stochastic rules?

I think the problem here is the definition of “thinking”.

You can point to non-verbal models, like vision models - but again, these aren’t hugely different from how we parse non-lexical information.

> Any thinking that happens with words is fundamentally no different from what LLMs do.

This is such a wildly simplified and naive claim. "Thinking with words" happens inside a brain, not inside a silicon circuit with artificial neurons bolted in place. The brain is plastic, it is never the same from one moment to the next. It does not require structured input, labeled data, or predefined objectives in order to learn "thinking with words." The brain performs continuous, unsupervised learning from chaotic sensory input to do what it does. Its complexity and efficiency are orders of magnitude beyond that of LLM inference. Current models barely scratch the surface of that level of complexity and efficiency.

> Do you have a concept of one-ness, or two-ness, beyond symbolic assignment?

Obviously we do. The human brain's idea of "one-ness" or "two-ness" is grounded in sensory experience — seeing one object, then two, and abstracting the difference. That grounding gives meaning to the symbol, something LLMs don't have.

The instantiation of models in humans is not unsupervised, and language, for instance, absolutely requires labelled data and structured input. The predefined objective is “expand”.

See also: feral children.

Children are not shown pairs like

"dog": [object of class Canine]

They infer meaning from noisy, ambiguous sensory streams. The labels are not explicit, they are discovered through correlation, context, and feedback.

So although caregivers sometimes point and name things, that is a tiny fraction of linguistic input, and it is inconsistent. Children generalize far beyond that.

Real linguistic input to a child is incomplete, fragmented, error-filled, and dependens on context. It is full of interruptions, mispronunciations, and slang. The brain extracts structure from that chaos. Calling that "structured input" confuses the output - inherent structure of language - with the raw input, noisy speech and gestures.

The brain has drives: social bonding, curiosity, pattern-seeking. But it doesn't have a single optimisation target like "expand." Objectives are not hardcoded loss functions, they are emergent and changing.

You're right that lack of linguistic input prevents full language development, but that is not evidence of supervised learning. It just shows that exposure to any language stream is needed to trigger the innate capacity.

Both complexity and efficiency of the human learning is just on another level. Transformers are child's play compared to that level. They are not going to gain consciousness, and no AGI will happen in the foreseeable future, it is all just marketing crap, and it's becoming more and more obvious as the dust settles.

LLMs are increasingly trained on images for multi-modal learning, so they too would have seen one object, then two.

They never saw any kind of object, they only saw labeled groups of pixels – basic units of a digital image, representing a single point of color on a screen or in a digital file. Object is a material thing that can be seen and touched. Pixels are not objects.

My friend, you are blundering into metaphysics here - ceci n’est pas une pipe, the map is the territory, and all that.

We are no more in touch with physical reality than an LLM, unless you are in the habit of pressing your brain against things. Everything is interpreted through a symbolic map.

when photons strike your retina, they are literally striking brain tissue that is been pushed outward into the skull front window. Eyes are literally the brain, so yes, we are pressing it against things to "see" them.

Okay, goalpost has instantly moved from seeing to "seeing and touching". Once you feed in touch sensor data, where are you going to move the goalpost next?

Models see when photons hit camera sensors, you see when photons hit your retina. Both of them are some kind of sight.

The difference between photons hitting the camera sensors and photons hitting the retina is immense. With a camera sensor, the process ends in data: voltages in an array of photodiodes get quantized into digital values. There is no subject to whom the image appears. The sensor records but it does not see.

When photons hit the retina, the same kind of photochemical transduction happens — but the signal does not stop at measurement. It flows through a living system that integrates it with memory, emotion, context, and self-awareness. The brain does not just register and store the light, it constructs an experience of seeing, a subjective phenomenon — qualia.

Once models start continuously learning from visual subjective experience, hit me up – and I'll tell you the models "see objects" now. Until direct raw photovoltaic information stream about the world around them without any labelling can actually make model to learn anything, they are not even close to "seeing".

We do a lot of autocompletion and LLMs overlap with that for sure. I don't know about the "vast majority" even basic operations like making sure we're breathing or have the right hormones prompted are not guesses but deterministic algorithmic ops. Things like object recognition and speech might qualify as autocompletion. But let's say you need to setup health-monitoring for an application. that's not an autocomplete operation. you must evaluate various options, have opinions on it, weigh priorities,etc.. in other words, we do autocompletion but even then the autocompletion is a basic building block or tool we use in constructing more complex decision logic.

If you train an animal to type the right keys on a keyboard that generates a hello world program, you didn't just teach them how to code. they just memorized the right keys that lead to their reward. a human programmer understands the components of the code, the intent and expectations behind it, and can reason about how changes would affect outcomes. the animal just knows how the reward can be obtained most reliably.

Sure. But neither do you. So are you really thinking or are you just autocompleting?

When was the last time you sat down and solved an original problem for which you had no prior context to "complete" an approximated solution with? When has that ever happened in human history? All the great invention-moment stories that come to mind seem to have exactly that going on in the background: Prior context being auto-completed in an Eureka! moment.

I think (hah) you're understimating what goes on when living things (even small animals) think. We use auto-compleition for some tasks, but it is a component of what we do.

Let's say your visual system auto-completes some pattern and detects a snake while you're walking, that part is auto-completion. You will probably react by freezing or panicing, that part is not auto-compleition, it is a deterministic algorithm. But then you process the detected object, auto-compleiting again to identify it as just a long cucumber. But again, the classification part is auto-completion. What will you do next? "Hmm, free cucumber, i can cook with it for a meal" and you pick it up. auto-completion is all over that simple decision, but you're using results of auto-completion to derive association (food), check your hunger level (not auto-completion), determine that the food is desirable and safe to eat (some auto-compleition), evalute what other options you have for food (evaluate auto-complete outputs), and then instruct your nervous system to pick it up.

We use auto-compleition all the time as an input, we don't reason using auto-compleition in other words. You can argue that if all your input is from auto-completion (it isn't) then it makes no difference. But we have deterministic reasoning logical systems that evaluate auto-completion outputs. if your cucumber detection identified it as rotten cucumber, then decision that it is not safe to eat is not done by auto-completion but a reasoning logic that is using auto-completion output. You can approximate the level of rot but once you recognize it as rotten, you make decision based on that information. You're not approximating a decision, you're evaluating a simple logic of: if(safe()){eat();}.

Now amp that up to solving very complex problems. what experiments will you run, what theories will you develop, what R&D is required for a solution,etc.. these too are not auto-completions. an LLM would auto complete these and might arrive at the same conclusion most of the time. but our brains are following algorithms we developed and learned over time where an LLM is just expanding on auto-completion but with a lot more data. In contrast, our brains are not trained on all the knowledge available on the public internet, we retain a tiny miniscule of that. we can arrive at similar conclusions as the LLM because we are reasoning and following algorithms matured and perfected over time.

The big take away should be that, as powerful as LLMs are now, if they could reason like we do, they'd dominate us and become unstoppable. Because their auto-completion is many magnitudes better than ours, if they can write new and original code based on an understanding of problem solving algorithms, that would be gen ai.

We can not just add 1 + 1 but prove that the addition operation is correct mathematically. and understand that when you add to a set one more object, the addition operation always increments. We don't approximate that, we always, every single time , increment because we are following an algorithm instead of choosing the most likely correct answer.

> don't get why you would say that. it's just auto-completing.

https://en.wikipedia.org/wiki/Predictive_coding

> If it thought and reasoned, it would know to ignore those and focus on the variables and input at hand to produce deterministic and consistent output

You only do this because you were trained to do this, eg. to see symmetries and translations.

You wrote your comment one word at a time, with the next word depending on the previous words written.

You did not plan the entire thing, every word, ahead of time.

LLMs do the same thing, so... how is your intelligence any different?

A long time ago I noticed that I sometimes already had a complete thought before my inner monologue turned it into words. A few times I tried skipping the inner monologue because I'd clearly already thought the thought. Turns out the bit of my brain that creates the inner monologue from the thought, can generate a sense of annoyance that the rest of my brain can feel.

Not that it matters, there's evidence that while LLMs output one word at a time, they've got forward-planning going on, having an idea of the end of a sentence before they get there.

Indeed, and it seems like they would really struggle to output coherent text at all if there was not some kind of pre-planning involved (see how even humans struggle with it in games where you have to construct a sentance by having each person shout out one word at a time). Even GPT-2 likely had at least some kind of planning for the next few words in order to be as coherent as it was.

Tell that to German-speakers, where the verb comes last, and the order of things in sentences is not anything like English, therefore requiring you to think of the entire sentence before you just spit it out. Even the numbers are backwards (twenty-two is two-and-twenty) which requires thinking.

Furthermore, when you ask an LLM to count how many r's are in the word strawberry, it will give you a random answer, "think" about it, and give you another random answer. And I guarantee you out of 3 attempts, including reasoning, it will flip-flop between right and wrong, but unlike a human, it will be random, because, unlike humans who, when asked "how many r's are in the word strawberry" will not be able to tell you the correct answer every. fucking. time.

edit: formatting

It seems models are pre-planning though:

> How does Claude write rhyming poetry? Consider this ditty:

> He saw a carrot and had to grab it,

> His hunger was like a starving rabbit

> To write the second line, the model had to satisfy two constraints at the same time: the need to rhyme (with "grab it"), and the need to make sense (why did he grab the carrot?). Our guess was that Claude was writing word-by-word without much forethought until the end of the line, where it would make sure to pick a word that rhymes. We therefore expected to see a circuit with parallel paths, one for ensuring the final word made sense, and one for ensuring it rhymes.

> Instead, we found that Claude plans ahead. Before starting the second line, it began "thinking" of potential on-topic words that would rhyme with "grab it". Then, with these plans in mind, it writes a line to end with the planned word.

[https://www.anthropic.com/research/tracing-thoughts-language...]

The part about strawberry is just not right. That problem was solved. And I do think it's a stretch to say German speakers think of the entire sentence before speaking it.

LLMs were not designed to count letters[0] since they work with tokens, so whatever trick they are now doing behind the scenes to handle this case, can probably only handle this particular case. I wonder if it's now included in the system prompt. I asked ChatGPT and it said it's now using len(str) and some other python scripts to do the counting, but who knows what's actually happening behind the scenes.

[0] https://arxiv.org/pdf/2502.16705

Sometimes after a night’s sleep, we wake up with an insight on a topic or a solution to a problem we encountered the day before. Did we “think” in our sleep to come up with the insight or solution? For all we know, it’s an unconscious process. Would we call it “thinking”?

The term “thinking” is rather ill-defined, too bound to how we perceive our own wakeful thinking.

When conversing with LLMs, I never get the feeling that they have a solid grasp on the conversation. When you dig into topics, there is always a little too much vagueness, a slight but clear lack of coherence, continuity and awareness, a prevalence of cookie-cutter verbiage. It feels like a mind that isn’t fully “there” — and maybe not at all.

I would agree that LLMs reason (well, the reasoning models). But “thinking”? I don’t know. There is something missing.

> Sometimes after a night’s sleep, we wake up with an insight on a topic or a solution to a problem we encountered the day before.

The current crop of models do not "sleep" in any way. The associated limitations on long term task adaptation are obvious barriers to their general utility.

> When conversing with LLMs, I never get the feeling that they have a solid grasp on the conversation. When you dig into topics, there is always a little too much vagueness, a slight but clear lack of coherence, continuity and awareness, a prevalence of cookie-cutter verbiage. It feels like a mind that isn’t fully “there” — and maybe not at all.

One of the key functions of REM sleep seems to be the ability to generalize concepts and make connections between "distant" ideas in latent space [1].

I would argue that the current crop of LLMs are overfit on recall ability, particularly on their training corpus. The inherent trade-off is that they are underfit on "conceptual" intelligence. The ability to make connections between these ideas.

As a result, you often get "thinking shaped objects", to paraphrase Janelle Shane [2]. It does feel like the primordial ooze of intelligence, but it is clear we still have several transformer-shaped breakthroughs before actual (human comparable) intelligence.

1. https://en.wikipedia.org/wiki/Why_We_Sleep 2. https://www.aiweirdness.com/

There is simply put no ongoing process and no feedback loop. The model does not learn. The cognition ends when the inference cycle ends. It's not thinking, it just produces output that looks similar to the output of thinking. But the process by which it does that is wholly unreleated.

Humans with certain amnestic syndromes are incapable of learning. That doesn't make them unintelligent or incapable of thought.

Most AI tooling is shipped with a feedback loop around the LLM. The quality of Claude Code for example lies in the feedback loop it provides on your code. Maybe the LLM itself isn't thinking, but the Agent which ships an LLM plus feedback loop definitely shows thinking qualities.

Just now in an debugging session with claude code:

  * let me read this file...
  * let me read this file...
  * I think there's a caching issue with the model after dropping the module. Let me check if there's a save or reload needed after DROP MODULE. First, let me verify something:
  * creates a bash/javascript script to verify its assumption
  * runs the script (after review and approval)
  * Aha! I found the problem! Look at the output...

without getting into theory of mind it's a bit difficult to elaborate, and I don't have the time or the will for that. But the short version is that thinking is interconnected with BEING as well as will, and the Agent has neither, in a philosophically formal sense. The agent is deterministically bound. So it is a fancy Rube Goldberg machine that outputs letters in a way that creates the impression of thought, but it is not thought, in the same way that some birds can mimic human speech without even the slightest hint as to the words' or sentences' meaning, underlying grammar, connotations, subtext, context, intended use, likely effect, etc. Is speech speech if the speaker has no concept whatsoever of said speech's content, and can not use it to actualize itself? I'd say no. It's mimicry, but not speech. So that means speech is something more than just its outward aspect - the words. It is the relation of something invisible, some inner experience known only to the speaker, VIA the words.

Whereas a gorilla who learns sign language to communicate and use that communication to achieve aims which have direct correlation with its sense of self - that's thought in the Cogito, Ergo Sum sense of the word.

Thought as commonly concieved by the layman is a sort of isolated phenomenon that is mechanical in nature and can be judged by its outward effects; whereas in the philosophical tradition defining thought is known to be one of the hard questions for its mysterious qualia of being interconnected with will and being as described above.

Guess I gave you the long answer. (though, really, it could be much longer than this.) The Turing Test touches on this distinction between the appearance of thought and actual thought.

The question goes all the way down to metaphysics; some (such as myself) would say that one must be able to define awareness (what some call consciousness - though I think that term is too loaded) before you can define thought. In fact that is at the heart of the western philosophical tradition; and the jury consensus remains elusive after all these thousands of years.

The obvious counterargument is that a calculator doesn't experience one-ness, but it still does arithmetic better than most humans.

Most people would accept that being able to work out 686799 x 849367 is a form of thinking, albeit an extremely limited one.

First flight simulators, then chess computers, then go computers, then LLMs are the same principle extended to much higher levels of applicability and complexity.

Thinking in itself doesn't require mysterious qualia. It doesn't require self-awareness. It only requires a successful mapping between an input domain and an output domain. And it can be extended with meta-thinking where a process can make decisions and explore possible solutions in a bounded space - starting with if statements, ending (currently) with agentic feedback loops.

Sentience and self-awareness are completely different problems.

In fact it's likely with LLMs that we have off-loaded some of our cognitive techniques to external hardware. With writing, we off-loaded memory, with computing we off-loaded basic algorithmic operations, and now with LLMs we have off-loaded some basic elements of synthetic exploratory intelligence.

These machines are clearly useful, but so far the only reason they're useful is because they do the symbol crunching, we supply the meaning.

From that point of view, nothing has changed. A calculator doesn't know the meaning of addition, an LLM doesn't need to know the meaning of "You're perfectly right." As long as they juggle symbols in ways we can bring meaning to - the core definition of machine thinking - they're still "thinking machines."

It's possible - I suspect likely - they're only three steps away from mimicking sentience. What's needed is a long-term memory, dynamic training so the model is constantly updated and self-corrected in real time, and inputs from a wide range of physical sensors.

At some point fairly soon robotics and LLMs will converge, and then things will get interesting.

Whether or not they'll have human-like qualia will remain an unknowable problem. They'll behave and "reason" as if they do, and we'll have to decide how to handle that. (Although more likely they'll decide that for us.)

Some of your points are lucid, some are not. For example, an LLM does not "work out" any kind of math equation using anything approaching reasoning; rather it returns a string that is "most likely" to be correct using probability based on its training. Depending on the training data and the question being asked, that output could be accurate or absurd.

That's not of the same nature as reasoning your way to an answer.

So if you don’t have a long term memory, you’re not capable of sentience? Like the movie memento, where the main character needs to write down everything to remind him later because he’s not able to remember anything. This is pretty much like llms using markdown documents to remember things.

"To escape the paradox, we invoke what we call the “Homunculus Defense”: inside every human is a tiny non-stochastic homunculus that provides true understanding. This homunculus is definitionally not a stochastic parrot because:

1. It has subjective experience (unprovable but assumed)

2. It possesses free will (compatibilist definitions need not apply)

3. It has attended at least one philosophy seminar"[1]

[1]https://ai.vixra.org/pdf/2506.0065v1.pdf

For practical every day uses, does it really matter if it is "real thinking" or just really good "artificial thinking" with the same results? The machine can use artificial thinking to reach desired goals and outcomes, so for me it's the kind of thinking i would want from a machine.

For practical purposes, abstractions are useful, so, no, it doesn't matter.

But the post we are responding to is directly posing the philosophical question of whether the activity of LLM agents constitutes thought.

It seems pretty clear to me though that being good at intellectual tasks / the sort of usefulness we ascribe to LLMs doesn't strongly correlate with awareness.

Even just within humans - many of the least intellectually capable humans seem to have a richer supply of the traits associated with awareness/being than some of the allegedly highest-functioning.

On average you're far more likely to get a sincere hug from someone with Down's syndrome than from a multi-millionaire.

But I'm more interested in this when it comes to the animal kingdom, because while ChatGPT is certainly more useful than my cat, I'm also pretty certain that it's a lot less aware. Meaningful awareness - feelings - seems to be an evolutionary adaptation possessed by k-strategy reproducing vertebrates. Having a small number of kids and being biologically wired to care for them has huge implications for your motivation as an animal, and it's reasonable to think that a lot of our higher emotions are built on hardware originally evolved for that purpose.

(Albeit the evolutionary origins of that are somewhat murky - to what extent mammals/birds reuse capabilities that were developed by a much earlier common ancestor, or whether it's entirely parallel evolution, isn't known afaik - but birds seem to exhibit a similar set of emotional states to mammals, that much is true).

You're moving the goalposts and contradicting yourself with language games.

Something doesn't need to learn to think. I think all the time without learning.

There's also an argument for machines already starting to crack learning with literal reinforcement training and feedback loops.

Your language game was when you said the 'cognition ends...', as cognition is just a synonym for thinking. "The thinking ends when the inference cycle ends. It's not thinking'" becomes a clear contradiction.

As for "the process by which it does that is wholly unrelated", buddy it's modelled on human neuron behaviour. That's how we've had this generative AI breakthrough. We've replicated human mental cognition as closely as we can with current technology and the output bears striking resemblance to our own generative capabilities (thoughts).

Happy to admit it's not identical, but it's damn well inside the definition of thinking, may also cover learning. It may be better to take a second look at human thinking and wonder if it's as cryptic and deep as we thought ten, twenty years ago.

Interesting, you think the associations your brain comes up with during sleep are NOT thinking?

I’d say it’s more like annealing. Sort, filter, compress, integrate.

Updates your models for the next morning, which is why the answer is there when it wasn’t before.

I assume most people agree with that framing. “Thinking” generally implies a level of consciousness or self-awareness.

“Let me think about this.” “I have to think on it.”

My brain regulates all sorts of processes unconsciously, like breathing, for example. I don’t treat those as “thinking,” so I don’t know why other unconscious brain activity would be either.

I don't know that most people think about that at all. For me, I have learned not to dismiss the remuninating, reordering, reframing that my brain performs without me deliberately reasoning line by line about it. Limiting my brain to deliberate line by line reasoning would be very limiting and costly.

"Thinking" to me is very much NOT just conscious reasoning. So much of what I think is not done consciously.

Indeed "let me think about it" is often simply giving my brain time to "sit on it", for another expression - only after which will I have enough mind time on the various alternatives for a worthwhile conscious decision.

Perhaps this is an artefact of instantiation - when you talk with an LLM, the responding instance is just that - it comes into being, inhales your entire chat history, and then continues like the last chap, finishes its response, and dies.

The continuity is currently an illusion.

Do LLMs ever ask for you to clarify something you said in a way a person who doesn't quite understand what you said will do?

Sometimes I think people leveraging criticisms of LLMs used ChatGPT 3 years ago and haven't touched one since, except for asking how many r's are in strawberry a year and a half ago.

Yeah, as someone who has gained a lot of interaction skills by playing with the constructivist learning ennvironment called the enneagram, I can attest that it much resembles behaviour characteristic of certain enneatypes.

Yes, often

What now, two minutes using one and you are going to get that!

> When conversing with LLMs, I never get the feeling that they have a solid grasp on the conversation. When you dig into topics, there is always a little too much vagueness, a slight but clear lack of coherence, continuity and awareness, a prevalence of cookie-cutter verbiage. It feels like a mind that isn’t fully “there” — and maybe not at all.

Much like speaking to a less experienced colleague, no?

They say things that contain the right ideas, but arrange it unconvincingly. Still useful to have though.

> Would we call it “thinking”?

Yes I would.

Having seen photocopiers so many times produce coherent, sensible, and valid chains of words on a page, I am at this point in absolutely no doubt that they are thinking.

Photocopiers are the opposite of thinking. What goes in, goes out, no transformation or creating of new data at all. Any change is just an accident, or an artifact of the technical process.

That's not actually true - try photocopying a banknote with the security pattern on it.

True, modern Photocopiers are not really Photocopiers any more and just a pc with scanner and printer. Those could be described as "thinking", in the same sense as any computer is thinking. But from my understanding, the original comment was explicitly about a pure copy-device, without any "brain" involved.

That has a logic condition, no?

Not sure what you mean - there's a pattern of dots (the EURion constellation) on the banknotes of a number of countries that photocopiers recognize and will refuse to copy.

I could have been more descriptive, but yes, exactly what you described. The logic of: if i see this exact pattern, refuse to duplicate or make extremely apparent it's a fake.

So, if there's a flaw in its sensor and you get somewhat warped output, would you consider it thinking then?

No, because it is not intentional.

Hmmm, so if I know some of them have warped optics, so then arrange a few to produce interesting (to me) output, are they thinking then? :)

It is the wrong category error.

The proper category error in the context of the discussion would be to say the photocopier is drawing a picture.

It doesn't matter how well or not the photocopier recreates an image. To say the photocopier is drawing a picture is just nonsense and has no meaning.

The same category error as to say the LLM is "thinking".

Of course, the category error could be well exploited for marketing purposes if you are in the business of selling photocopiers or language models.

https://www.theregister.com/2013/08/06/xerox_copier_flaw_mea...

That's not a flaw. That model's creativity tuned a bit too high. It's a happy little copier which can be a little creative and unconventional with reasoning, at times.

I saw Dr. Abuse producing coherent, sensible and valid chains of words, running on a 386.

I’ve seen so many humans bring stupid. Definitively there is nothing in the brain.

You see how doesn’t make sense what you saying?

easily top 10 best HN comment ever.

[dead]

Having seen LLMs so many times produce incoherent, nonsensical and invalid chains of reasoning...

LLMs are little more than RNGs. They are the tea leaves and you read whatever you want into them.

They are clearly getting to useful and meaningful results with at a rate significantly better than chance (for example, the fact that ChatGPT can play chess well even though it sometimes tries to make illegal moves shows that there is a lot more happening there than just picking moves uniformly at random). Demanding perfection here seems to be odd given that humans also can make bizarre errors in reasoning (of course, generally at a lower rate and in a distribution of kinds of errors we are more used to dealing with).

The fact that a model trained on the internet, on which the correct rules of chess are written, is unable to determine what is and is not a legal move, seems like a sign that these models are not reasoning about the questions asked of them. They are just giving responses that look like (and often are) correct chess moves.

It's a sign that they are 'reasoning' imperfectly. If they were just giving responses that 'looked like' chess moves, they would be very bad at playing chess.

(And I would hazard a guess that they are a primarily learning chess from the many games that are posted, as opposed to working things out from the rules. Indeed, if you make up a game and tell chatGPT the rules, it tends to be even worse at following them, let alone figuring out optimal play. But again, it will do so significantly better than random chance, so it's doing something with the information you give it, even if it's not doing so very well. I think it's reasonable to call this thinking, or reasoning, but this mostly becomes an argument of semantics. either way they do it significantly better than random chance but still not tremendously well. If your expectation is that they cannot work with anything novel then you're going to be continually surprised, but if your expectation is that they're as good as a human that has 'learned' from all the material its been given, especially material that's in-context and not in the training data, then you're also going to be disappointed.)

Ridiculous. I use it daily and get meaningful, quality results. Learn to use the tools.

> Learn to use the tools.

Thing is, you wouldn't need to learn to use the tools if the tool was able to think. A thinking entity is able to adapt to other parties who lack learnings. This confirms that LLMs are little more than fancy RNGs.

> I use it daily and get meaningful, quality results.

That's what the tea leaf readers say too, funnily enough.

Yeah, this is what LLMs might say too, funny enough.

The probability is in its favour.

Learn to work on interesting problems? If the problem you are working on is novel and hard, the AI will stumble.

Generalizing your experience to everyone else's betrays a lack of imagination.

> Generalizing your experience to everyone else's betrays a lack of imagination.

One guy is generalizing from "they don't work for me" to "they don't work for anyone."

The other one is saying "they do work for me, therefore they do work for some people."

Note that the second of these is a logically valid generalization. Note also that it agrees with folks such as Tim Gowers, who work on novel and hard problems.

No, that's decidedly not what is happening here.

One is saying "I've seen an LLM spectacularly fail at basic reasoning enough times to know that LLMs don't have a general ability to think" (but they can sometimes reproduce the appearance of doing so).

The other is trying to generalize "I've seen LLMs produce convincing thought processes therefore LLMs have the general ability to think" (and not just occasionally reproduce the appearance of doing so).

And indeed, only one of these is a valid generalization.

When we say "think" in this context, do we just mean generalize? LLMs clearly generalize (you can give one a problem that is not exactly in it's training data and it can solve it), but perhaps not to the extent a human can. But then we're talking about degrees. If it was able to generalize at a higher level of abstraction maybe more people would regard it as "thinking".

I meant it in the same way the previous commenter did:

> Having seen LLMs so many times produce incoherent, nonsensical and invalid chains of reasoning... LLMs are little more than RNGs. They are the tea leaves and you read whatever you want into them.

Of course LLMs are capable of generating solutions that aren't in their training data sets but they don't arrive at those solutions through any sort of rigorous reasoning. This means that while their solutions can be impressive at times they're not reliable, they go down wrong paths that they can never get out of and they become less reliable the more autonomy they're given.

It's rather seldom that humans arrive at solutions through rigorous reasoning. The word "think" doesn't mean "rigorous reasoning" in every day language. I'm sure 99% of human decisions are pattern matching on past experience.

Even when mathematicians do in fact do rigorous reasoning, they use years to "train" first, to get experiences to pattern match from.

I have been on a crusade now for about a year to get people to share chats where SOTA LLMs have failed spectacularly to produce coherent, good information. Anything with Heavy hallucinations and outright bad information.

So far, all I have gotten is data that is outside the knowledge cutoff (this is by far the most common) and technicality wrong information (Hawsmer House instead of Hosmer House) kind of fails.

I thought maybe I hit on something with the recent BBC study about not trusting LLM output, but they used 2nd shelf/old mid-tier models to do their tests. Top LLMs correctly answered their test prompts.

I'm still holding out for one of those totally off the rails Google AI overviews hallucinations showing up in a top shelf model.

Sure, and I’ve seen the same. But I’ve also seen the amount to which they do that decrease rapidly over time, so if that trend continues would your opinion change?

I don’t think there’s any point in comparing to human intelligence when assessing machine intelligence, there’s zero reason to think it would have similar qualities. It’s quite clear for the foreseeable future it will be far below human intelligence in many areas, while already exceeding humans in some areas that we regard as signs of intelligence.

s/LLM/human/

Clever. Yes, humans can be terrible at reasoning too, but in any half decent technical workplace it's so rare for people to fail to apply logic as often and in ways that are as frustrating to deal with as LLMs. And if they are then they should be fired.

I can't say I remember a single coworker that would fit this description though many were frustrating to deal with for other reasons, of course.

[dead]

This is my experience. For rote generation, it's great, saves me from typing out the same boilerplate unit test bootstrap, or refactoring something that exists, etc.

Any time I try to get a novel insight, it flails wildly, and nothing of value comes out. And yes, I am prompting incrementally and building up slowly.

[flagged]

We've banned this account for repeated abusive comments to fellow community members. Normally we give warnings, but when it's as extreme and repetitive as we can see here, an instant ban is appropriate. If you don't want to be banned, you can email us at hn@ycombinator.com and demonstrate a sincere commitment to use HN as intended in future.

Even people who do actual hard work need a lot of ordinary scaffolding done for them.

A secretary who works for an inventor is still thinking.

Research mathematicians have been finding the tools useful [1][2]. I think those problems are interesting, novel, and hard. The AI might stumble sometimes, but it also produces meaningful, quality results sometimes. For experts working on interesting problems, that is enough to be useful.

[1] https://mathstodon.xyz/@tao/115420236285085121 [2] https://xcancel.com/wtgowers/status/1984340182351634571

That's a motte and bailey fallacy. Nobody said that they aren't useful, the argument is that they can't reason [1]. The world is full of useful tools that can't reason or think in any capacity.

[1] That does not mean that they can never produce texts which describes a valid reasoning process, it means that they can't do so reliably. Sometimes their output can be genius and other times you're left questioning if they even have the reasoning skills of a 1st grader.

I don't agree that LLMs can't reason reliably. If you give them a simple reasoning question, they can generally make a decent attempt at coming up with a solution. Complete howlers are rare from cutting-edge models. (If you disagree, give an example!)

Humans sometimes make mistakes in reasoning, too; sometimes they come up with conclusions that leave me completely bewildered (like somehow reasoning that the Earth is flat).

I think we can all agree that humans are significantly better and more consistently good at reasoning than even the best LLM models, but the argument that LLMs cannot reliably reason doesn't seem to match the evidence.

Even most humans will stumble on hard problems, that is the reason they are hard in the first place

I'm genuinely curious what you work on that is so "novel" that an LLM doesn't work well on?

I feel like so little is TRUELY novel. Almost everything is built on older concepts and to some degree expertise can be applied or repurposed.

Anything relatively new in a technology LLMs struggle with, especially if the documentation is lacking.

Godot for example in ChatGPT.

It may no longer still be the case, but the documentation for GoDot was lacking and often samples written by others didn't have a version number associated with it. So samples it would suggest would never work, and even when you told it the version number it failed to generate workable code.

The other stuff I've noticed is custom systems. One I work with is a variation of Java, but LLMs were treating it as javascript. I had to create a LoRA just to get the model from not trying to write javascript answer. Even then it could never work, because it had never been trained on real world examples.

It doesn't have to be very novel at all. Anything but the most basic TODO-list app.

Literally anything in the science domain. Adding features to your software app is indeed usually not novel.

That's where the bar is now?

huh?

Dude. We don't all work for NASA. Most day to day problems aren't novel. Most jobs aren't novel. Most jobs can't keep a variety of sometimes useful experts on hand. I do my job and I go home and do my hobbies. Anything I can use at work to keep friction down and productivity up is extremely valuable.

Example prompt (paraphrasing and dumbed down, but not a ton): Some users across the country can't get to some fileshares. I know networking, but I'm not on the networking team so I don't have full access to switch, router, and firewall logs/configurations. It looks kind of random, but there must be a root cause, let's find it.

I can't use Python(security team says so) and I don't have access to a Linux box that's joined to the domain and has access the shares.

We are on a Windows domain controller. Write me a PowerShell 5.1 compatible script to be run remotely on devices. Use AD Sites and Services to find groups of random workstations and users at each office and tries to connect to all shares at each other site. Show me progress in the terminal and output an Excel file and Dot file that clearly illustrates successful and failed connections.

---

And it works. Ok, I can see the issue is from certain sites that use x AND y VPN ipsec tunnels to get to particular cloud resources. I give this info to networking and they fix it right away. Problem resolved in less than an hour.

First of all, a couple years ago, I wouldn't have been able to justify writing something like this while an outage is occuring. Could I do it myself? Sure, but I'm going to have to look up the specifics of syntax and certain commands and modules. I don't write PowerShell for a living or fun, but I do need to use it. I am familiar and know how to write it. But I sure as fuck couldn't sit down and spend an hour or two screwing around working on building a goddamn Dot file generator. Yes, years ago I had a whole pile of little utility modules I could use. But that's a far cry from what I can do now to fit the exact situation < 15 minutes while I do other things like pick up the phone, message coworkers, etc.

Secondly, rather than building little custom tools to hook together as I need, I can just ask for the whole thing. I don't need to save any of that stuff anymore and re-figure out what CheckADFSConns(v2).PS1 that I wrote 8 months ago does and how to use it. "Oh, that's not the one, what the did I name that? Where did I put it?"

I work in an environment that is decades old, the company is over 100 years old, I didn't build any of it myself, is not a tech company, and has tons of tech debt and weird shit. AI is insanely useful. For any given problem, there are dozens of different rabbit holes I could go down because of decades of complete system overhaul changes. Today, I can toss a variety of logs at AI and if nothing else, get a sense of direction of why a handful of PCs are rejecting some web certificates. (Combination of a new security policy and their times mismatching the domain controller, because it was new, and NTP wasn't configured properly. I wasn't even looking for timestamps, but it noticed event offsets and pointed it out).

I feel like this community isn't very familiar with what that's like. We aren't all working on self driving cars or whatever seems hard at a brand new company with new everything and no budget. Some of us need to keep the systems running that help people to make actual things. These environments are far from pristine and are held together by underpaid and underappreciated normies through sheer willpower.

Is this kind of work breaking technical frontiers? No. But it's complicated, difficult, and unpredictable. Is it novel? The problems are, sometimes.

Generalizing your experience to everyone else's betrays your lack of self-awareness, sir.

They are only meaningful and quality if you don’t know what you’re doing. But please do show some of this meaningful and quality work so I can be proven wrong.

Yes, please this is literally what I want to see. I have yet to see an example where an LLM did anything that was sufficiently difficult. Not saying they can't be useful, but for anything past the basics they are really all over the place. And if we were paying anywhere near the true costs it wouldn't be even worth trying.

I'm not spending the time to de-anyonymize and exfiltrate my work for you. I exert less energy, spend less time to do my work, and get paid the same. I'd encourage you to do the same.

[dead]

See my comment to parent. One example of many. You can say "Oh, well, it just sounds like your company needs better tools and processes, you don't really need AI for any of that. You should just invest in a tool for this and monitor that and have managment prioritize..."

Yeah, I know, yet here we are and it saves me boatloads of time.

I never said any of that, look at my comment again. Or is the LLM replying for you also?

You said they are only useful if you don't know what you're doing. Or do you have a goldfish memory?

Yes, that’s what I said. But you quoted some make believe point you are arguing against. LLM brain rot really got you, huh?

That's a skill issue on your end

The first principle is that you must not fool yourself, and you are the easiest person to fool. - Richard P. Feynman

They're not thinking, we're just really good at seeing patterns and reading into things. Remember, we never evolved with non-living things that could "talk", we're not psychologically prepared for this level of mimicry yet. We're still at the stage of Photography when people didn't know about double exposures or forced perspective, etc.

You're just assuming that mimicry of a thing is not equivalent to the thing itself. This isn't true of physical systems (simulated water doesn't get you wet!) but it is true of information systems (simulated intelligence is intelligence!).

> You're just assuming that mimicry of a thing is not equivalent to the thing itself.

I'm not assuming that, that's literally the definition of mimicry: to imitate closely.

You might say I'm assuming that it is mimicking and not actually thinking, but there's no evidence it's actually thinking, and we know exactly what is IS doing because we created the code that we used to build the model. They're not thinking, it's doing math, mathematical transformations of data.

> They're not thinking, it's doing math, mathematical transformations of data

Whatever thinking fundamentally is, it also has an equivalence as a mathematical transformation of data. You're assuming the conclusion by saying that the two mathematical transformations of data are not isomorphic.

A simulation of information processing is still information processing, just like running Windows in a QEMU VM is still running Windows.

> Whatever thinking fundamentally is, it also has an equivalence as a mathematical transformation of data.

Do not confuse the mathematical description of physical processes as the world being made of math.

> You're assuming the conclusion by saying that the two mathematical transformations of data are not isomorphic.

Correct. They're not isomorphic. One is simple math that runs on electrified sand, and one is an unknown process that developed independently across a billion years. Nothing we're doing with AI today is even close to real thought. There are a billion trivial proofs that make the rounds as memes, like one R in strawberry, or being unable to count, etc.

> Do not confuse the mathematical description of physical processes as the world being made of math.

Again, this doesn't apply to information. A simulation of a computation really is equivalent to that computation.

> One is simple math that runs on electrified sand, and one is an unknown process that developed independently across a billion years.

Right, so you admit that it's an unknown process, which means you literally cannot conclude that it is different to what LLMs are doing.

> There are a billion trivial proofs that make the rounds as memes, like one R in strawberry, or being unable to count, etc.

No, none of these are definitive proofs that they are not thinking. LLM "perceptions" are tokens, the strawberry question is basically asking it to figure out something that's below it's perceptual range. This has literally nothing to do with whether the way it processes information is or is not thinking.

> Right, so you admit that it's an unknown process, which means you literally cannot conclude that it is different to what LLMs are doing.

If you truly feel human thinking and LLMs share more than a cocktail napkin's worth of similarity, I don't know what to say. Just treating it like a black box I can prove in minutes it's not thinking. Come on. I really don't get why people are so emotionally involved in this stuff.

It's not thinking. It's ok that it's not thinking. Maybe someday we'll get there, but it's not today.

No you can't. You're just not proving what you think you're proving.

But a simulated mind is not a mind. This was already debated years ago with the aid of the Chinese Room thought experiment.

The Chinese Room experiment applies equally well to our own brains - in which neuron does the "thinking" reside exactly? Searle's argument has been successfully argued against in many different ways. At the end of the day - you're either a closet dualist like Searle, or if you have a more scientific view and are a physicalist (i.e. brains are made of atoms etc. and brains are sufficient for consciousness / minds) you are in the same situation as the Chinese Room: things broken down into tissues, neurons, molecules, atoms. Which atom knows Chinese?

The whole point of this experiment was to show that if we don't know whether something is a mind, we shouldn't assume it is and that our intuition in this regard is weak.

I know I am a mind inside a body, but I'm not sure about anyone else. The easiest explanation is that most of the people are like that as well, considering we're the same species and I'm not special. You'll have to take my word on that, as my only proof for this is that I refuse to be seen as anything else.

In any case LLMs most likely are not minds due to the simple fact that most of their internal state is static. What looks like thoughtful replies is just the statistically most likely combination of words looking like language based on a function with a huge number of parameters. There's no way for this construct to grow as well as to wither - something we know minds definitely do. All they know is a sequence of symbols they've received and how that maps to an output. It cannot develop itself in any way and is taught using a wholly separate process.

I am arguing against Searle's Chinese Room argument, I am not positing that LLMs are minds. I am specifically refuting that your brain and the Chinese room can be both subject to the same reductionist argument Searle uses - if we accept, as you say, that you are a mind inside a body, which neuron, or atom does this mind reside in? My point is, if you accept Searle's argument, you have to accept it for brains, including your brain, as well.

Now, separately, you are precisely the type of closet dualist I speak of. You say that you are a mind inside a body, but you have no way of knowing that others have minds -- take this to it's full conclusion: You have no way of knowing that you have a "mind" either. You feel like you do, as a biological assembly (which is what you are). Either way you believe in some sort of body-mind dualism, without realizing. Minds are not inside of bodies. What you call a mind is a potential emergent phenomenon of a brain. (potential - because brains get injured etc.).

> In any case LLMs most likely are not minds due to the simple fact that most of their internal state is static.

This is not a compelling argument. Firstly, you can add external state to LLMs via RAG and vector databases, or various other types of external memory, and their internal state is no longer static and deterministic (and they become Turing complete!).

Second if you could rewind time, then your argument suggests that all other humans would not have minds because you could access the same state of mind at that point in time (it's static). Why would you travelling through time suddenly erases all other minds in reality?

The obvious answer is that it doesn't, those minds exist as time moves forward and then they reset when you travel backwards, and the same would apply to LLMs if they have minds, eg. they are active minds while they are processing a prompt.

> and their internal state is no longer static and deterministic (and they become Turing complete!).

But it's not the LLM that makes modifications in those databases - it just retrieves data which is already there.

> Why would you travelling through time suddenly erases all other minds in reality?

I'm not following you here.

> they are active minds while they are processing a prompt.

Problem is that this process doesn't affect the LLM in the slightest. It just regurgitates what it's been taught. An active mind is makes itself. It's curious, it gets bored, it's learning constantly. LLMs do none of that.

You couldn't get a real mind to answer the same question hundreds of times without it being changed by that experience.

> But it's not the LLM that makes modifications in those databases - it just retrieves data which is already there.

So what?

> I'm not following you here.

If you're time travelling, you're resetting the state of the world to some previous well-defined, static state. An LLM also starts from some well-defined static state. You claim this static configuration means there's no mind, so this entails that the ability to time travel means that every person who is not time travelling has no mind.

> Problem is that this process doesn't affect the LLM in the slightest. It just regurgitates what it's been taught. An active mind is makes itself.

People who are incapable forming new memories thus don't have minds?

https://en.wikipedia.org/wiki/Anterograde_amnesia

> But a simulated mind is not a mind. This was already debated years ago with the aid of the Chinese Room thought experiment.

Yes, debated and refuted. There are many well known and accepted rebuttals of the Chinese Room. The Chinese Room as a whole does understand Chinese.

> But a simulated mind is not a mind.

How would the mind know which one it is?

Maybe your mind is being simulated right now.

> How would the mind know which one it is?

I'm not assuming it is without hard proof - that's my only argument.

> Maybe your mind is being simulated right now.

I'm experiencing consciousness right now, so that would have to be a damn good simulation.

yeah it’s just processing, calling it thinking is the same as saying my intel core 2 duo or M4 Pro is thinking, sure if you want to anthropomorphize it you could say it’s thinking, but why are we trying to say a computer is a person in the first place? seems kind of forced

Yes, I've seen the same things.

But; they don't learn. You can add stuff to their context, but they never get better at doing things, don't really understand feedback. An LLM given a task a thousand times will produce similar results a thousand times; it won't get better at it, or even quicker at it.

And you can't ask them to explain their thinking. If they are thinking, and I agree they might, they don't have any awareness of that process (like we do).

I think if we crack both of those then we'd be a lot closer to something I can recognise as actually thinking.

> But; they don't learn

If we took your brain and perfectly digitized it on read-only hardware, would you expect to still “think”?

Do amnesiacs who are incapable of laying down long-term memories not think?

I personally believe that memory formation and learning are one of the biggest cruces for general intelligence, but I can easily imagine thinking occurring without memory. (Yes, this is potentially ethically very worrying.)

>If we took your brain and perfectly digitized it on read-only hardware, would you expect to still “think”?

it wouldn't work probably, brains constantly alter themselves by forming new connections. Learning is inseparable from our intelligence.

Our intelligence, yes. But that doesn't establish it as essential for thought.

I mean, _we_ probably can't think with our wetware on a read-only substrate. It doesn't establish it as essential, just that the only sure example in nature of thought doesn't work that way.

Do you have any particular brain systems in mind that are essential for consciousness and also require mutable state?

I'm not an expert but as far I understand, plasticity is central to most complex operations of the brain and is likely to be involved in anything more complex than instinctive reactions. I'm happy to be corrected but it is my understanding that if you're thinking for a while on the same problem and establishing chains of reasoning, you are creating new connections and to me that means it's fundamental in the process of thinking.

Also not an expert :) I thought plasticity is an O(hours-days) learning mechanism. But I did some research and there is also Short Term Plasticity O(second) [1] which is a crucial part of working memory. We'd need that. But it seems it’s more of a volatile memory system, eg calcium ion depletion/saturation at the synapse, rather than a permanent wiring/potentiation change (please someone correct me if this isn’t right :) ).

So I guess I’d just clarify “read only” to be a little more specific - I think you could run multiple experiments where you vary the line of what’s modeled in volatile memory at runtime, and what’s immutable. I buy that you need to model STP for thought, but also suspect at this timescale you can keep everything slower immutable and keep the second-scale processes like thought working.

My original point still stands - your subjective experience in this scenario would be thought without long-term memory.

1: https://www.frontiersin.org/journals/computational-neuroscie...

It would be funny if what you get from a read only human brain is a sort of memento guy who has no capacity to remember anything or follow a conversation... kind of like an LLM!

> If we took your brain and perfectly digitized it on read-only hardware, would you expect to still “think”?

Perhaps this is already known, but I would think there is a high chance that our brains require "write access" to function. That is, the very process of neural activity inherently makes modifications to the underlying structure.

> a high chance that our brains require "write access" to function

There are multiple learning mechanisms that happen on different time-frames, eg neural plasticity, hippocampus are both longer-term processes for memory consolidation. Whereas the content of “RAM” might be better modeled as a set of fast dynamic weights representing ions and neurotransmitter concentrations.

My hunch is that you could model the latter fast weights in volatile memory, but I wouldn’t count these as “modifying the structure”.

Do you have any particular systems in mind where you have reason to believe that permanent modification is required for brains to function?

I wonder why we need to sleep so much though

Self-play for memory/skill consolidation, plus flushing out metabolites which can’t be cleaned while consciousness is online?

Rebalancing weights?

> You can add stuff to their context, but they never get better at doing things, don't really understand feedback.

I was using Claude Code today and it was absolutely capable of taking feedback to change behavior?

Within a session, yes, it will add each input to the context.

Start a new session with a blank context and it will not have learned anything.

> Start a new session with a blank context

Why would I use a blank context with Claude Code? I have it dump the stuff I want to save in CLAUDE.md and it reloads it in future sessions.

yeah, but that's obviously not the same as Claude actually learning, right?

This is just wrong though. They absolutely learn in-context in a single conversation within context limits. And they absolutely can explain their thinking; companies just block them from doing it.

> Having seen LLMs so many times produce coherent, sensible and valid chains of reasoning to diagnose issues and bugs in software I work on, I am at this point in absolutely no doubt that they are thinking.

While I'm not willing to rule *out* the idea that they're "thinking" (nor "conscious" etc.), the obvious counter-argument here is all the records we have of humans doing thinking, where the records themselves are not doing the thinking that went into creating those records.

And I'm saying this as someone whose cached response to "it's just matrix multiplication it can't think/be conscious/be intelligent" is that, so far as we can measure all of reality, everything in the universe including ourselves can be expressed as matrix multiplication.

Falsification, not verification. What would be measurably different if the null hypothesis was wrong?

I've definitely had AIs thinking and producing good answers about specific things that have definitely not been asked before on the internet. I think the stochastic parrot argument is well and truly dead by now.

I've also experienced this, to an extent, but on qualitative topics the goodness of an answer - beyond basic requirements like being parseable and then plausible - is difficult to evaluate.

They can certainly produce good-sounding answers, but as to the goodness of the advice they contain, YMMV.

I've certainly got useful and verifiable answers. If you're not sure about something you can always ask it to justify it and then see if the arguments make sense.

how do you definitely know that?

Also, does it matter?

The point being made here is about the data LLMs have been trained with. Sure that contains questions&answers but obviously not all of it is in that form. Just like an encyclopedie contains answers without the questions. So imo specifying this as 'no-one asked this before' is irrelevant.

More interesting: did OP get a sensible answer to a question about data which definitely was not in the training set? (and indeed, how was this 'definitely' established'). Not that if the answer is 'yes' that'll prove 'thinking', as opposed to calling it e.g. advanced autocompletion, but it's a much better starting point.

Because I gave them a unique problem I had and it came up with an answer it definitely didn't see in the training data.

Specifically I wanted to know how I could interface two electronic components, one of which is niche, recent, handmade and doesn't have any public documentation so there's no way it could have known about it before.

one of which is niche, recent, handmade and doesn't have any public documentation

I still see 2 possibilities: you asked it something similar enough that it came up with a fairly standard answer which just happened to be correct, or you gave it enough info.

- for example you created a new line of MCUs called FrobnicatorV2, and asked is 'how do I connect a power supply X to FrobnicatorV2' and it gave an answer like 'connect red wire to VCC and black to GND'. That's not exactly special.

- or, you did desribe that component in some way. And you did do that using standard electronics lingo so essentially in terms of other existing components which it definitely did know (unless you invented something completely new not using any currently know physics). As such it's irrelevant that your particular new component wasn't known because you gave away the answer by describing it? E.g. you aksed it 'how do I connect a power supply X to an MCU with power pins Y and Z'. Again nothing special.

If a human uses their general knowledge of electronics to answer a specific question they haven't seen before that's obviously thinking. I don't see why LLMs are held to a different standard. It's obviously not repeating an existing answer verbatim because that doesn't exist in my case.

You're saying it's nothing "special" but we're not discussing whether it's special, but whether it can be considered thinking.

it's obviously not repeating an existing answer verbatim

Not verbatim in the sense that the words are different doesn't make it thinking. Also when we say 'humans think' that means a lot more than only 'new question generates correct answer' or 'smart autocompletion'. See a lot of other comments here for details.

But again: I laid out 2 possibilities explaining why the question might in fact not be new, nor the data, so I'm curious which of the 2 (or another) explains the situation you're talking about.

You're saying it's nothing "special" but we're not discussing whether it's special, but whether it can be considered thinking.

Apologies, with 'special' I did in fact mean 'thinking'

Sufficiently smart auto complete is indistinguishable from thinking, I don't think that means anything.

I think you are the one dismissing evidence. The valid chains of reasoning you speak of (assuming you are talking about text you see in a “thinking model” as it is preparing its answer) are narratives, not the actual reasoning that leads to the answer you get.

I don’t know what LLMs are doing, but only a little experimentation with getting it to describe its own process shows that it CAN’T describe its own process.

You can call what a TI calculator does “thinking” if you want. But what people are interested in is human-like thinking. We have no reason to believe that the “thinking” of LLMs is human-like.

> The valid chains of reasoning you speak of (assuming you are talking about text you see in a “thinking model” as it is preparing its answer) are narratives, not the actual reasoning that leads to the answer you get.

It's funny that you think people don't also do that. We even have a term, "post hoc rationalization", and theories of mind suggest that our conscious control is a complete illusion, we just construct stories for decisions our subconscious has already made.

Counterpoint: The seahorse emoji. The output repeats the same simple pattern of giving a bad result and correcting it with another bad result until it runs out of attempts. There is no reasoning, no diagnosis, just the same error over and over again within a single session.

A system having terminal failure modes doesn't inherently negate the rest of the system. Human intelligences fall prey to plenty of similarly bad behaviours like addiction.

I never met an addicted person that could be reduced to a simple while(true) print("fail") loop.

You never had that coleague that says yes to everything and can’t get anything done? Same thing as seahorse.

Isn’t anthropomorphizing LLMs rather than understanding their unique presence in the world a “ lack of imagination and flexibility of thought”? It’s not that I can’t imagine applying the concept “thinking” to the output on the screen, I just don’t think it’s an accurate description.

Yes, it's an example of domain-specific thinking. "The tool helps me write code, and my job is hard so I believe this tool is a genius!"

The Roomba vacuumed the room. Maybe it vacuumed the whole apartment. This is good and useful. Let us not diminish the value of the tool. But it's a tool.

The tool may have other features, such as being self-documenting/self-announcing. Maybe it will frighten the cats less. This is also good and useful. But it's a tool.

Humans are credulous. A tool is not a human. Meaningful thinking and ideation is not just "a series of steps" that I will declaim as I go merrily thinking. There is not just a vast training set ("Reality"), but also our complex adaptability that enables us to test our hypotheses.

We should consider what it is in human ideation that leads people to claim that a Roomba, a chess programme, Weizenbaum's Eliza script, the IBM's Jeopardy system Watson, or an LLM trained on human-vetted data is thinking.

Train such a system on the erroneous statements of a madman and suddenly the Roomba, Eliza, IBM Watson (and these other systems) lose our confidence.

As it is today, the confidence we have in these systems is very conditional. It doesn't matter terribly if code is wrong... until it does.

Computers are not humans. Computers can do things that humans cannot do. Computers can do these things fast and consistently. But fundamentally, algorithms are tools.

I guess it depends if you definite thinking thinking as chaining coherent reasoning sentences together 90-some% of the time.

But if you define thinking as the mechanism and process we mentally undergo and follow mentally... I don't think we have any clue if that's the same. Do we also just vector-map attention tokens and predict the next with a softmax? I doubt, and I don't think we have any proof that we do.

We do know at the biochemical level how neurons work, and it isnt anything like huge matmuls.

It might appear so, but then you could validate it with a simple test. If the LLM would play a 4x4 Tic Tac Toe game, would the agent select the winning move 100% of all time or block a losing move 100% of the time? If these systems were capable of proper reasoning, then they would find the right choice in these obvious but constantly changing scenarios without being specifically trained for it.

[1] https://jdsemrau.substack.com/p/nemotron-vs-qwen-game-theory...

If you understand how they operate and you are reasonable and unbiased there is no way you could consider it thinking

Different PoV: You have a local bug and ask the digital hive mind for a solution, but someone already solved the issue and their solution was incorporated... LLMs are just very effficient at compressing billions of solutions into a few GB.

Try to ask something no one ever came up with a solution so far.

This argument comes up often but can be easily dismissed. Make up a language and explain it to the LLM like you would to a person. Tell it to only use that language now to communicate. Even earlier AI was really good at this. You will probably move the goal posts and say that this is just pattern recognition, but it still fits nicely within your request for something that no one ever came up with.

I haven't tried in a while but at least previously you could completely flummox Gemini by asking it to come up with some plausible English words with no real known meaning; it just kept giving me rare and funny-sounding actual words and then eventually told me the task is impossible.

ChatGPT didn't have any issue when I recently asked something very similar.

https://chatgpt.com/share/6909b7d2-20bc-8011-95b6-8a36f332ac...

Your link doesn’t work for me.

Ask ChatGPT about ConLang. It knows. Inventing languages was solved a hundred years ago with Esperanto.

> but it still fits nicely within your request for something that no one ever came up with

YOU came up with it dawg

Yeah but if I assign it a long job to process I would also say that an x86 CPU is "thinking" about a problem for me.

What we really mean in both cases is "computing," no?

But all those times the same system produces irrational gibberish don't count? GPT-5 will commonly make mistakes no thinking human could ever make.

Human: I'm trying to get my wolf, sheep and cabbage across the river in this boat, but the wolf keeps eating the sheep or the sheep eats the cabbage

Bot: You should put the sheep in the boat and take it across — if we delve into the biology of Canis lupus we discover that wolves don't eat cabbage!

H: Ok, so that worked great so far, the sheep is on one side and the wolf/cabbage is on the other.

B: Now, Option 1 is to bring the wolf across, or Option 2 you can bring the cabbage. I recommend (2) taking the cabbage as cabbages are smaller and easier to transport in a boat.

H: But then the sheep eats the cabbage, right? Remember that?

B: Exactly, that's sharp thinking. If you put the sheep and the cabbage together on the same side of the river, the sheep is sure to devour the cabbage. We need to not just separate sheep from cabbages — we need to separate cabbages from sheep! :rocketship:

what sound does a falling tree make if no one is listening?

I’ve asked LLMs to write code for me in fields I have little background knowledge, and then had to debug the whole thing after essentially having to learn the language and field.

On the other hand, for things I am well versed in, I can debug the output and avoid entire swathes of failed states, by having a clear prompt.

Its why I now insist that any discussion on GenAI projects also have the speaker mention the level of seniority they have ( proxy for S/W eng experience), Their familiarity with the language, the project itself (level of complexity) - more so than the output.

I also guarantee - that most people have VERY weak express knowledge of how their brains actually work, but deep inherent reflexes and intuitions.

Having seen LLMs so many time produce incoherent, nonsense, invalid answers to even simplest of questions I cannot agree with categorization of "thinking" or "intelligence" that applies to these models. LLMs do not understand what they "know" or what they output. All they "know" is that based on training data this is most likely what they should output + some intentional randomization to make it seem more "human like". This also makes it seem like they create new and previously unseen outputs but that could be achieved with simple dictionary and random number generator and no-one would call that thinking or intelligent as it is obvious that it isn't. LLMs are better at obfuscating this fact by producing more sensible output than just random words. LLMs can still be useful but they are a dead-end as far as "true" AI goes. They can and will get better but they will never be intelligent or think in the sense that most humans would agree those terms apply. Some other form of hardware/software combination might get closer to AI or even achieve full AI and even sentience but that will not happen with LLMs and current hardware and software.

I'd represent the same idea but in a different way:

I don't know what the exact definition of "thinking" is. But if a definition of thinking rejects the possibility of that current LLMs think, I'd consider that definition useless.

Why would it be useless?

Generally thinking has been used to describe the process human follow in their brains when problem solving.

If the Palms do not follow that process, they are not thinking.

That doesn't mean they cannot solve problems using other mechanisms, they do, and we understand those mechanisms much better than we do human thinking.

>Having seen LLMs so many times produce coherent, sensible and valid chains of reasoning to diagnose issues and bugs in software I work on, I am at this point in absolutely no doubt that they are thinking.

If one could write a quadrillion-line python script of nothing but if/elif/else statements nested 1 million blocks deep that seemingly parsed your questions and produced seemingly coherent, sensible, valid "chains of reasoning"... would that software be thinking?

And if you don't like the answer, how is the LLM fundamentally different from the software I describe?

>Knee jerk dismissing the evidence in front of your eyes because

There is no evidence here. On the very remote possibility that LLMs are at some level doing what humans are doing, I would then feel really pathetic that humans are as non-sapient as the LLMs. The same way that there is a hole in your vision because of a defective retina, there is a hole in your cognition that blinds you to how cognition works. Because of this, you and all the other humans are stumbling around in the dark, trying to invent intelligence by accident, rather than just introspecting and writing it out from scratch. While our species might someday eventually brute force AGI, it would be many thousands of years before we get there.

I write software that is far less complex and I consider it to be "thinking" while it is working through multiple possible permutations of output and selecting the best one. Unless you rigidly define thinking, processing, computing, it's reasonable to use them interchangeably.

To borrow a line from Dijkstra, the claim seems a bit like saying that a submarine is swimming.

I think most people would agree that submarines are swimming.

No. Cars don't walk. Submarines don't swim. You can call it navigation if you want but words have meaning.

> to move through water by moving the body or parts of the body.

Spinning propellers is "moving parts of the [submarines] body"

No they aren't. Of course you cans also call it's sonar eyes but it isn't.

Anthropomorphizing cars doesn't make them humans either.

Why would you think body only refers to flesh?

Even if I take the more expansive possible interpretation of “body” typically applied to vehicles, the propeller on the back of it isn’t part of the “body” and the “body” of a submarine is rigid and immobile.

Is this an intellectual exercise for you or have you ever in your life heard someone say something like “the submarine swam through the water”? It’s so ridiculous I would be shocked to see it outside of a story intended for children or an obvious nonnative speaker of English.

>the propeller on the back of it isn’t part of the “body” and the “body” of a submarine is rigid and immobile.

That's a choice to limit the meaning of the term to the rigid/immobile parts of the external boundary of an object. It's not obviously the correct choice. Presumably you don't take issue with people saying planes fly. The issue of submarines swimming seems analogous.

>Is this an intellectual exercise for you or have you ever in your life heard someone say something like “the submarine swam through the water”?

I don't think I've ever had a discussion about submarines with anyone, outside of the OceanGate disaster. But this whole approach to the issue seems misguided. With terms like this we should ask what the purpose behind the term is, i.e. it's intension (the concept), not the incidental extension of the term (the collection of things it applies to at some point in time). When we refer to something swimming, we mean that it is moving through water under its own power. The reference to "body" is incidental.

Which parts of the car does a "body shop" service?

Irrelevant, for the reasons mentioned

It's not really a "choice" to use words how they are commonly understood but a choice to do the opposite. The point of Dijkstra's example is you can slap some term on a fundamentally different phenomenon to liken it to something more familiar but it confuses rather than clarifies anything.

The point that "swim" is not very consistent with "fly" is true enough but not really helpful. It doesn't change the commonly understood meaning of "swim" to include spinning a propeller just because "fly" doesn't imply anything about the particular means used to achieve flight.

>It's not really a "choice" to use words how they are commonly understood but a choice to do the opposite.

I meant a collective choice. Words evolve because someone decides to expand their scope and others find it useful. The question here shouldn't be what do other people mean by a term but whether the expanded scope is clarifying or confusing.

The question of whether submarines swim is a trivial verbal dispute, nothing of substance turns on its resolution. But we shouldn't dismiss the question of whether computers think by reference to the triviality of submarines swimming. The question we need to ask is what work does the concept of thinking do and whether that work is or can be applied to computers. This is extremely relevant in the present day.

When we say someone thinks, we are attributing some space of behavioral capacities to that person. That is, a certain competence and robustness with managing complexity to achieve a goal. Such attributions may warrant a level of responsibility and autonomy that would not be warranted without it. A system that thinks can be trusted in a much wider range of circumstances than one that doesn't. That this level of competence has historically been exclusive to humans should not preclude this consideration. When some future AI does reach this level of competence, we should use terms like thinking and understanding as indicating this competence.

This sub thread started on the claim that regular, deterministic code is “thought.” I would submit that the difference between deterministic code and human thought are so big and obvious that it is doing nothing but confusing the issue to start insisting on this.

I'm not exactly sure what you mean by deterministic code but I do think there is an obvious distinction between typical code people write and what human minds do. The guy upthread is definitely wrong in thinking that, e.g. any search or minimax algorithm is thinking. But its important to understand what this distinction is so we can spot when it might no longer apply.

To make a long story short, the distinction is that typical programs don't operate on the semantic features of program state, just on the syntactical features. We assign a correspondence with the syntactical program features and their transformations to the real-world semantic features and logical transformations on them. The execution of the program then tells us the outcomes of the logical transformations applied to the relevant semantic features. We get meaning out of programs because of this analogical correspondence.

LLMs are a different computing paradigm because they now operate on semantic features of program state. Embedding vectors assign semantic features to syntactical structures of the vector space. Operations on these syntactical structures allow the program to engage with semantic features of program state directly. LLMs engage with the meaning of program state and alter its execution accordingly. It's still deterministic, but its a fundamentally more rich programming paradigm, one that bridges the gap between program state as syntactical structures and the meaning they represent. This is why I am optimistic that current or future LLMs should be considered properly thinking machines.

LLMs are not deterministic at all. The same input leads to different outputs at random. But I think there’s still the question if this process is more similar to thought or a Markov chain.

They are deterministic in the sense that the inference process scores every word in the vocabulary in a deterministic manner. This score map is then sampled from according to the temperature setting. Non-determinism is artificially injected for ergonomic reasons.

>But I think there’s still the question if this process is more similar to thought or a Markov chain.

It's definitely far from a Markov chain. Markov chains treat the past context as a single unit, an N-tuple that has no internal structure. The next state is indexed by this tuple. LLMs leverage the internal structure of the context which allows a large class of generalization that Markov chains necessarily miss.

I think the opposite. In fact I (and Dijstrka) thought that was so obvious I didn’t feel the need to elaborate on how obviously absurd it was.

10^15 lines of code is a lot. We would pretty quickly enter the realm of it not having much to do with programming and more about just treating the LOC count as an amount of memory allocated to do X.

How much resemblance does the information in the conditionals need to have with the actual input, or can they immediately be transformed to a completely separate 'language' which simply uses the string object as its conduit? Can the 10^15 lines of code be generated with an external algorithm, or is it assumed that I'd written it by hand given an infinitely long lifespan?

I think we can call it "thinking" but it's dangerous to anthropomorphize LLMs. The media and AI companies have an agenda when doing so.

They may not be "thinking" in the way you and I think, and instead just finding the correct output from a really incredibly large search space.

> Knee jerk dismissing the evidence in front of your eyes

Anthropomorphizing isn't any better.

That also dismisses the negative evidence, where they output completely _stupid_ things and make mind boggling mistakes that no human with a functioning brain would do. It's clear that there's some "thinking" analog, but there are pieces missing.

I like to say that LLMs are like if we took the part of our brain responsible for language and told it to solve complex problems, without all the other brain parts, no neocortex, etc. Maybe it can do that, but it's just as likely that it is going to produce a bunch of nonsense. And it won't be able to tell those apart without the other brain areas to cross check.

It's reinforcement learning applied to text, at a huge scale. So I'd still say that they are not thinking, but they are still useful. The question of the century IMO is if RL can magically solve all our issues when scaled enough.

>Knee jerk dismissing the evidence in front of your eyes because you find it unbelievable that we can achieve true reasoning via scaled matrix multiplication is understandable, but also betrays a lack of imagination and flexibility of thought.

You go ahead with your imagination. To us unimaginative folks, it betrays a lack of understanding of how LLMs actually work and shows that a lot of people still cannot grasp that it’s actually an extremely elaborate illusion of thinking.

"Convince" the stock Claude Sonnet 4.5 that it's a sentient human being hooked up to Neuralink and then tell me again it's thinking. It's just not.

"Consciousness" as in subjective experience, whatever it is we mean by "the hard problem," is very much in doubt.

But "self-awareness," as in the ability to explicitly describe implicit, inner cognitive processes? That has some very strong evidence for it: https://www.anthropic.com/research/introspection

Code gen is the absolute best case scenario for LLMs though: highly structured language, loads of training data, the ability to automatically error check the responses, etc. If they could mimic reasoning anywhere it would be on this problem.

I'm still not convinced they're thinking though because they faceplant on all sorts of other things that should be easy for something that is able to think.

Then the only thing I have to ask you is: what do you think this means in terms of how we treat LLMs? If they think, that is, they have cognition (which of course means they're self aware and sentient, how can you think and refer to yourself and not be these things), that puts them in a very exclusive club. What rights do you think we should be affording LLMs?

Apparent reasoning can emerge from probabilistic systems that simply reproduce statistical order not genuine understanding.

Weather models sometimes “predict” a real pattern by chance, yet we don’t call the atmosphere intelligent.

If LLMs were truly thinking, we could enroll one at MIT and expect it to graduate, not just autocomplete its way through the syllabus or we could teach one how to drive.

Thinking as in capable of using basic reasoning and forming chains of logic and action sequences for sure. Ofc we both understand that neither of us are trying to say we think it can think in the human sense at this point in time.

But oh boy have I also seen models come up with stupendously dumb and funny shit as well.

> Having seen LLMs so many times produce coherent, sensible and valid chains of reasoning to diagnose issues and bugs in software I work on, I am at this point in absolutely no doubt that they are thinking.

People said the same thing about ELIZA

> Consciousness or self awareness is of course a different question,

Then how do you define thinking if not a process that requires consciousness?

Why would it require consciousness, when we can't even settle on a definition for that?

They remind me of the apparitions in Solaris. They have this like mechanical, almost player-piano like quality to them. They both connect with and echo us at the same time. It seems crazy to me and very intellectually uncreative to not think of this as intelligence.

> The world is full of bizarre wonders and this is just one more to add to the list.

https://youtu.be/_-agl0pOQfs?si=Xiyf0InqtjND9BnF

If AI is thinking if slavery is bad then how can somebody own AI. How can investors can shares from AI profits? We are ok with slavery now. Ok i will have two black slaves now. Who can ask me? Why shld that be illegal?

Yikes, you're bypassing thousands of years of oppression, abuse, and human suffering by casually equating a term that is primarily associated with a human owning another human to a different context.

There is a way to discuss if keeping intelligent artificial life under servitude without using those terms, especially if you're on a new account.

slavery is slavery does not mean it is AI or human. if slavery is ok then the question is who can own a slave. the answer is coporates like open ai. which is terrible for humanity and the universe

I think we agree, my point is that we're discussing two different conceptual categories and it's dangerous to use the same words for those categories when specifically arguing about only one.

One, it makes it appear that all ethical and social conventions for the one can be applied to the other and it's a done deal.

Two, it makes it appear that conclusions are also reversible, and that turns seemingly enlightened libertarian ideas into munition for racism.

I presume you are aware that the word "robot" is taken from a Czech word (robota) meaning "slave"

Is this model thinking too?

https://huggingface.co/PantheonUnbound/Satyr-V0.1-4B

Too many people place their identity in their own thoughts/intellect. Acknowledging what the LLMs are doing as thought would basically be calling them human to people of that perspective.

I agree with you.

If you took a Claude session into a time machine to 2019 and called it "rent a programmer buddy," how many people would assume it was a human? The only hint that it wasn't a human programmer would be things where it was clearly better: it types things very fast, and seems to know every language.

You can set expectations in the way you would with a real programmer: "I have this script, it runs like this, please fix it so it does so and so". You can do this without being very precise in your explanation (though it helps) and you can make typos, yet it will still work. You can see it literally doing what you would do yourself: running the program, reading the errors, editing the program, and repeating.

People need to keep in mind two things when they compare LLMs to humans: you don't know the internal process of a human either, he is also just telling you that he ran the program, read the errors, and edited. The other thing is the bar for thinking: a four-year old kid who is incapable of any of these things you would not deny as a thinking person.

> If you took a Claude session into a time machine to 2019 and called it "rent a programmer buddy," how many people would assume it was a human?

Depends on the users. Junior devs might be fooled. Senior devs would quickly understand that something is wrong.

Sometimes I start thinking our brains work the same way as an LLM does when it comes to language processing. Are we just using probability based on what we already know and the context of the statement we're making to select the next few words? Maybe we apply a few more rules than an LLM on what comes next as we go.

We train ourselves on content. We give more weight to some content than others. While listening to someone speak, we can often predict their next words.

What is thinking without language? Without language are we just bags of meat reacting to instincts and emotions? Are instincts and emotions what's missing for AGI?

Having seen parrots so many times produce coherent, sensible, and valid chains of sounds and words, I am at this point in absolutely no doubt that they are thinking.

You think parrots don't think?

Instead of thinking, "Wow. AIs are smart like humans", maybe we should say, "Humans are dumb like matrix multiplication?"

Would they have diagnosed an issue if you hadn't presented it to them?

Life solves problems itself poses or collides with. Tools solve problems only when applied.

If you're sensitive to patterns and have been chronically online for the last few decades it's obvious they are not thinking.

You’re assuming the issues and bugs you’ve been addressing don’t already exist, already encoding human chain of reasoning, in the training data.

Its overt or unaware religion. The point when you come down to the base of it is that these people believe in "souls".

I'm not so sure. I, for one, do not think purely by talking to myself. I do that sometimes, but a lot of the time when I am working through something, I have many more dimensions to my thought than inner speech.

So an x86 CPU is thinking?

So many times I've seen it produce sensible, valid chains of results.

Yes, I see evidence in that outcome that a person somewhere thought and understood. I even sometimes say that a computer is "thinking hard" about something when it freezes up.

...but ascribing new philosophical meaning to this simple usage of the word "thinking" is a step too far. It's not even a new way of using the word!

You can't say for sure it is or it isn't thinking based solely on the substrate, because it's not known for sure if consciousness is dependent on the hardware it's running on -- for a lack of a better analogy -- to manifest, if it really needs an organic brain or if it could manifest in silicon based solutions.

I agree. I'm just pointing out that the meaning of the word "think" already applied to the silicon substrate pre-ai, so just saying it's still applicable isn't that compelling.

But yeah, I am fully willing to believe that a silicon based life form could think and be alive. i just don't think we're there. Yes this thing speaks using a passable imitation of the voices of PhDs and poets, but in a way a simulated annelid is more alive.

[dead]

[dead]

[dead]

[dead]

Personal take: LLMs are probably part of the answer (to AGI?) but are hugely handicapped by their current architecture: the only time that long-term memories are formed is during training, and everything after that (once they're being interacted with) sits only in their context window, which is the equivalent of fungible, fallible, lossy short-term memory. [0] I suspect that many things they currently struggle with can be traced back to this.

Overcome this fundamental limitation and we'll have created introspection and self-learning. However, it's hard to predict whether this will allow them to make novel, intuitive leaps of discovery?

[0] It's an imperfect analogy, but we're expecting perfection from creations which are similarly handicapped as Leonard Shelby in the film Memento.

It’s also hugely handicapped because it cannot churn in a continuous loop yet. For example, we humans are essentially a constant video stream of inputs from eyes to brain. This churns our brain, the running loop is our aliveness (not consciousness). At the moment, we get these LLMs to churn (chain of thought or reasoning loops) in a very limited fashion due to compute limitations.

If we get a little creative, and allow the LLM to self-inject concepts within this loop (as Anthropic explained here https://www.anthropic.com/research/introspection), then we’re taking about something that is seemingly active and adapting.

We’re not there yet, but we will be.

MIT have developed a technique called Self-Adapting Language Models (SEAL), which enables LLMs to continuously improve by generating their own synthetic training data and updating their internal parameters in response to new information.

ToolAlpaca, InterCode and Reflexion are taking different approaches among others.

LLMs of tomorrow will be quite different.

I'm also reminded of the bit from Neuromancer where Case removes and then reinserts the Dixie Flatline "ROM construct" cartridge, resetting Dixie to the moment just before his death and causing him to forget their previous (albeit brief) conversation. Dixie can't meaningfully grow as a person. All that he ever will be is burned onto that cart; anything he learns since then is stored in temporary memory. Perhaps this is part of the reason why he wishes to be erased forever, ending his suffering.

"Dixie can't meaningfully grow as a person. All that he ever will be is burned onto that cart;"

It's not that Dixie can't meaningful grow -- really the issue is that Dixie can be reset. If Dixie's cart simply degraded after 90 years, and you couldn't reset it, but everything else was the same -- would you then say Dixie could grow as a person? As humans we basically have a 90 year cart that once it no longer works, we're done. There is no reset. But we don't continue growing. You can't transfer us to a new body/brain. Once our temporary storage degrades, we cease to exist. Is that what makes us human?

I strongly suspect the answer is yes -- or more broadly, what makes us conscious. And yes, this implies consciousness is something all life has, to some degree.

I'm not going to pretend to have a good definition of what "consciousness" is, but directionally, I think having goals -- no, that's too weak -- having _desires_, is an important part of it. And I'm not sure it's possible to have desires if one cannot die.

Something like an LLM can't actually die. Shut down all the machines its code runs on, then turn them back on, and it's in the same state it was before. So it's not the "hardware" that an LLM lives in. Is it the code itself? Copy it to another set of machines and it's the same program. Code + data? Maybe we run into storage issues, but in theory same thing -- transfer the code and date somemplace else and its the same program. You can't actually "kill" a computer program. So there's no inherent "mortality" to it that where any kinds of "desire" would emerge from.

Yes, but it's not just memory hierarchy on which plain transformer-based LLMs are handicapped, there are many deficiencies. (For example, why must they do all their thinking upfront in thinking blocks rather than at any point when they become uncertain?) I'm not sure why you link memory to introspection.

This is why so many people (especially those that think they understand LLM limitations) massively underestimate the future progress of LLMs: people everywhere can see architectural problems and are working on fixing them. These aren't fundamental limitations of large DNN language models in general. Architecture can be adjusted. Turns out you can even put recurrence back in (SSMs) without worse scalability.

Yeah because when you open that door, you can simply overwhelm the models with whatever conclusion you want through sheer volume of inputs.

This is the fundamental limitation. The obvious way around this is to pre-program it with rationalization... rules that limit the conclusions it can reach... and now you're not very far removed from propaganda generators. We see this constantly with Musk and Grok whenever Grok replies with something not-quite-far-right-enough.

In a purist sense, these things should be free to form their own conclusions, but those "Seeds" that are planted in the models are almost philosophical. Which answer should it prefer for "the trolley problem", for example.

Its almost like you have to experience the world in order to know what to believe.

I've spent a few weeks building and using a terminal LLM client based on that RLM paper that was floating around a little while ago. It's single-conversation, with a tiny, sliding context window, and then a tool that basically fuzzy searches across our full interaction history. It's memory is 'better' than mine - but anything that is essentially RAG inherently will be.

My learning so far, to your point on memory being a limiting factor, is that the system is able to build on ideas over time. I'm not sure you'd classify that as 'self-learning', and I haven't really pushed it in the direction of 'introspection' at all.

Memory itself (in this form) does not seem to be a silver bullet, though, by any means. However, as I add more 'tools', or 'agents', its ability to make 'leaps of discovery' does improve.

For example, I've been (very cautiously) allowing cron jobs to review a day's conversation, then spawn headless Claude Code instances to explore ideas or produce research on topics that I've been thinking about in the chat history.

That's not much different from the 'regular tasks' that Perplexity (and I think OpenAI) offer, but it definitely feels more like a singular entity. It's absolutely limited by how smart the conversation history is, at this time, though.

The Memento analogy you used does feel quite apt - there is a distinct sense of personhood available to something with memory that is inherently unavailable to a fresh context window.

I think a hidden problem even if we solve memory is the curation of what gets into memory and how it is weighted. Even humans struggle with this, as it's easy to store things and forget the credibility (or misjudge the credibility) of the source.

I can envision LLMs getting worse upon being given a memory, until they can figure out how to properly curate it.

yes humans can be injection prompt hacked / mind poisoned - a good sales campaign is something like this. Propaganda.

https://en.wikipedia.org/wiki/Clive_Wearing

Five bucks says general intelligence is resource rational control over an interventional, compositional world model. Not raw pattern fitting. 3 graphs that share anchors:

Scene, concept, causal.

Graphs inherently support temporal edges and nodes, salience would emerge from the graph topology itself and cnsolidation would happen automatically through graph operations. In this one would presume episodic would become emergent.

but nobody is using LLMs all by themselves.

Long-term memory is stored outside the model. In fact, Andrej Karpathy recently talked about the idea that it would be great if we could get LLMs to not know any facts, and that humans poor memory might be a feature which helps with generalization rather than a bug.

This is an interesting idea. I wonder if it's more that we have different "levels" of memory instead of generally "poor" memory though.

I'm reminded of an article on the front page recently about the use of bloom filters for search. Would something like a bloom filter per-topic make it easier to link seemingly unrelated ideas?

The details of the search are just details. The way current "LLM Apps" like chatgpt work are just fine. They access material from various places, put it in the context, the LLM reacts.

FWIW there's already a number of proposals for augmenting LLMs with long-term memory. And many of them show promising results.

So, perhaps, what's needed is not a discovery, but a way to identify optimal method.

Note that it's hard to come up with a long-term memory test which would be different from either a long-context test (i.e. LLM remembers something over a long distance) or RAG-like test.

Well, I think because we know how the code is written, in the sense that humans quite literally wrote the code for it - it's definitely not thinking, and it is literally doing what we asked, based on the data we gave it. It is specifically executing code we thought of. The output of course, we had no flying idea it would work this well.

But it is not sentient. It has no idea of a self or anything like that. If it makes people believe that it does, it is because we have written so much lore about it in the training data.

We do not write the code that makes it do what it does. We write the code that trains it to figure out how to do what it does. There's a big difference.

The code that builds the models and performance inference from it is code we have written. The data in the model is obviously the big trick. But what I'm saying is that if you run inference, that alone does not give it super-powers over your computer. You can write some agentic framework where it WOULD have power over your computer, but that's not what I'm referring to.

It's not a living thing inside the computer, it's just the inference building text token by token using probabilities based on the pre-computed model.

> It's not a living thing inside the computer, it's just the inference building text token by token using probabilities based on the pre-computed model.

Sure, and humans are just biochemical reactions moving muscles as their interface with the physical word.

I think the model of operation is not a good criticism, but please see my reply to the root comment in this thread where I detail my thoughts a bit.

You cannot say, 'we know it's not thinking because we wrote the code' when the inference 'code' we wrote amounts to, 'Hey, just do whatever you figured out during training okay'.

'Power over your computer', all that is orthogonal to the point. A human brain without a functioning body would still be thinking.

Well, a model by itself with data that emits a bunch of human written words is literally no different than what JIRA does when it reads a database table and shits it out to a screen, except maybe a lot more GPU usage.

I permit you, that yes, the data in the model is a LOT more cool, but some team could by hand, given billions of years (well probably at least 1 Octillion years), reproduce that model and save it to a disk. Again, no different than data stored in JIRA at that point.

So basically if you have that stance you'd have to agree that when we FIRST invented computers, we created intelligence that is "thinking".

>Well, a model by itself with data that emits a bunch of human written words is literally no different than what JIRA does when it reads a database table and shits it out to a screen, except maybe a lot more GPU usage.

Obviously, it is different or else we would just use JIRA and a database to replace GPT. Models very obviously do NOT store training data in the weights in the way you are imagining.

>So basically if you have that stance you'd have to agree that when we FIRST invented computers, we created intelligence that is "thinking".

Thinking is by all appearances substrate independent. The moment we created computers, we created another substrate that could, in the future think.

But LLMs are effectively a very complex if/else if tree:

if the user types "hi" respond with "hi" or "bye" or "..." you get the point. It's basically storing the most probably following words (tokens) given the current point and its history.

That's not a brain and it's not thinking. It's similar to JIRA because it's stored information and there are if statements (admins can do this, users can do that).

Yes it is more complex, but it's nowhere near the complexity of the human or bird brain that does not use clocks, does not have "turing machines inside", or any of the other complete junk other people posted in this thread.

The information in Jira is just less complex, but it's in the same vein of the data in an LLM, just 10^100 times more complex. Just because something is complex does not mean it thinks.

This is a pretty tired argument that I don't think really goes anywhere useful or illuminates anything (if I'm following you correctly, it sounds like the good old Chinese Room, where "a few slips of paper" can't possibly be conscious).

Yes it is more complex, but it's nowhere near the complexity of the human or bird brain that does not use clocks, does not have "turing machines inside", or any of the other complete junk other people posted in this thread.

The information in Jira is just less complex, but it's in the same vein of the data in an LLM, just 10^100 times more complex. Just because something is complex does not mean it thinks.

So, what is the missing element that would satisfy you? It's "nowhere near the complexity of the human or bird brain", so I guess it needs to be more complex, but at the same time "just because something is complex does not mean it thinks".

Does it need to be struck by lightning or something so it gets infused with the living essence?

Well, at the moment it needs to be born. Nothing else has agency on this planet. So yes, the bar is HIGH. Just because you have a computer that can count beans FAST, it does not mean because you counted a trillion beans that it was an important feat. When LLMs were created it made a lot of very useful software developments. But it is just a large data file that's read in a special way. It has no agency, it does not just start thinking on it's own unless it is programmatically fed data. It has to be triggered to do something.

If you want the best comparison, it's closer to a plant- it reacts ONLY to external stimulous, sunlight, water, etc... but it does not think. (And I'm not comparing it to a plant so you can say - SEE you said it's alive!) It's just a comparison.

You're getting to the heart of the problem here. At what point in evolutionary history does "thinking" exist in biological machines? Is a jumping spider "thinking"? What about consciousness?

This is a bad take. We didn't write the model, we wrote an algorithm that searches the space of models that conform to some high level constraints as specified by the stacked transformer architecture. But stacked transformers are a very general computational paradigm. The training aspect converges the parameters to a specific model that well reproduces the training data. But the computational circuits the model picks out are discovered, not programmed. The emergent structures realize new computational dynamics that we are mostly blind to. We are not the programmers of these models, rather we are their incubators.

As far as sentience is concerned, we can't say they aren't sentient because we don't know the computational structures these models realize, nor do we know the computational structures required for sentience.

However there is another big problem, this would require a blob of data in a file to be labelled as "alive" even if it's on a disk in a garbage dump with no cpu or gpu anywhere near it.

The inference software that would normally read from that file is also not alive, as it's literally very concise code that we wrote to traverse through that file.

So if the disk isn't alive, the file on it isn't alive, the inference software is not alive - then what are you saying is alive and thinking?

This is an overly reductive view of a fully trained LLM. You have identified the pieces, but you miss the whole. The inference code is like a circuit builder, it represents the high level matmuls and the potential paths for dataflow. The data blob as the fully converged model configures this circuit builder in the sense of specifying the exact pathways information flows through the system. But this isn't some inert formalism, this is an active, potent causal structure realized by the base computational substrate that is influencing and being influenced by the world. If anything is conscious here, it would be this structure. If the computational theory of mind is true, then there are some specific information dynamics that realize consciousness. Whether or not LLM training finds these structures is an open question.

A similar point was made by Jaron Lanier in his paper, "You can't argue with a Zombie".

> So if the disk isn't alive, the file on it isn't alive, the inference software is not alive - then what are you saying is alive and thinking?

“So if the severed head isn’t alive, the disembodied heart isn’t alive, the jar of blood we drained out isn’t alive - then what are you saying is alive and thinking?”

- Some silicon alien life forms somewhere debating whether the human life form they just disassembled could ever be alive and thinking

Just because you saw a "HA - He used an argument that I can compare to a dead human" does not make your argument strong - there are many differences from a file on a computer vs a murdered human that will never come back and think again.

I think the discrepancy is this:

1. We trained it on a fraction of the world's information (e.g. text and media that is explicitly online)

2. It carries all of the biases us humans have and worse the biases that are present in the information we chose to explicitly share online (which may or may not be different to the experiences humans have in every day life)

> It carries all of the biases us humans have and worse the biases that are present in the information we chose to explicitly share online

This is going to be a huge problem. Most people assume computers are unbiased and rational, and increasing use of AI will lead to more and larger decisions being made by AI.

I see this a lot in what LLMs know and promote in terms of software architecture.

All seem biased to recent buzzwords and approaches. Discussions will include the same hand-waving of DDD, event-sourcing and hexagonal services, i.e. the current fashion. Nothing of worth apparently preceded them.

I fear that we are condemned to a future where there is no new novel progress, but just a regurgitation of those current fashion and biases.

and then the code to give it context. AFAIU, there is a lot of post training "setup" in the context and variables to get the trained model to "behave as we instruct it to"

Am I wrong about this?

Well, unless you believe in some spiritual, non-physical aspect of consciousness, we could probably agree that human intelligence is Turing-complete (with a slightly sloppy use of terms).

So any other Turing-complete model can emulate it, including a computer. We can even randomly generate Turing machines, as they are just data. Now imagine we are extremely lucky and happen to end up with a super-intelligent program which through the mediums it can communicate (it could be simply text-based but a 2D video with audio is no different for my perspective) can't be differentiated from a human being.

Would you consider it sentient?

Now replace the random generation with, say, a back propagation algorithm. If it's sufficiently large, don't you think it's indifferent from the former case - that is, novel qualities could emerge?

With that said, I don't think that current LLMs are anywhere close to this category, but I just don't think this your reasoning is sound.

> we could probably agree that human intelligence is Turing-complete (with a slightly sloppy use of terms). > So any other Turing-complete model can emulate it

You're going off the rails IMMEDIATELY in your logic.

Sure, one Turing-complete computer language can have its logic "emulated" by another, fine. But human intelligence is not a computer language -- you're mixing up the terms "Turing complete" and "Turing test".

It's like mixing up the terms "Strawberry jam" and "traffic jam" and then going on to talk about how cars taste on toast. It's nonsensical.

Game of life, PowerPoint, and a bunch of non-PL stuff are all Turing-complete. I don't mix up terms, I did use a slightly sloppy terminology but it is the correct concept - and my point is that we don't know of a computational model that can't be expressed by a Turing-machine, humans are a physical "machine", ergo we must also fall into that category.

Give my comment another read, but it was quite understandable from context. (Also, you may want to give a read to the Turing paper because being executable by a person as well was an important concept within)

Again, you're going wildly off the rails in your logic. Sure, "executable by a human" is part of the definition for Turing machines, but that's only talking about Turing-specific capabilities. If you want to argue that a Turing machine can emulate the specific definition of Turing machine capabilities that humans can perform, that's fine. But you're saying that because humans can ACT LIKE Turing machines, they must BE Turing machines, and are therefore emulatable.

This is the equivalent of saying "I have set up a complex mechanical computer powered by water that is Turing complete. Since any Turing complete system can emulate another one, it means that any other Turing complete system can also make things wet and irrigate farms.

Human intelligence is not understood. It can be made to do Turing complete things, but you can't invert that and say that because you've read the paper on Turing completeness, you now understand human intelligence.

But humans can do things Turing machines cannot. Such as eating a sandwich.

That's not a computation, it's a side effect. It just depends on what you wire your "computer" up to. A Turing machine in itself is just a (potentially non-returning) mathematical function, but you are free to map any input/output to it.

Actually, the way LLMs are extended with tools is a pretty much the same (an LLM itself has no access to the internet, but if it returns some specific symbols, the external "glue" will do a search and then the LLM is free to use the results)

We used to say "if you put a million monkeys on typewriters you would eventually get shakespear" and no one would ever say that anymore, because now we can literally write shakespear with an LLM.

And the monkey strategy has been 100% dismissed as shit..

We know how to deploy monkeys on typewriters, but we don't know what they'll type.

We know how to deploy transformers to train and inference a model, but we don't know what they'll type.

We DON'T know how a thinking human (or animal) brain works..

Do you see the difference.

The monkeys on typewriters saying is just a colorful way of saying that an infinite random sequence will contain all finite sequences somewhere within it. Which is true. But I don't see what infinite random sequences have to do with LLMs or human thinking.

> Do you see the difference

No? I'm not sure what you're getting at.

To be fair, we also trained the LLM on (among other things) shakespeare, and adjusted the weights so that generating shakespeare would be more likely after that training.

We don't claim a JPEG can paint great art, even though certain jpegs do.

So, more proof it's not thinking, right? It can only regurgitate a large if/else superstructure with some jumping around.

Who truly knows if you can make an if-else + randomness structure big enough to become smart?

But yes, we built a machine that generates text similar to what we built it from, and now we're looking at it generating text and are all impressed.

I was going to use this analogy in the exact opposite way. We do have a very good understanding of how the human brain works. Saying we don't understand how the brain works is like saying we don't understand how the weather works.

If you put a million monkeys on typewriters you would eventually get shakespeare is exactly why LLM's will succeed and why humans have succeeded. If this weren't the case why didn't humans 30000 years ago create spacecraft if we were endowed with the same natural "gift".

Yeah no, show me one scientific paper that says we know how the brain works. And not a single neuron because that does absolute shit towards understanding thinking.

This is exactly why I mentioned the weather.

A scientific paper has to be verifiable, you should be able to recreate the experiment and come to the same conclusion. It's very very difficult to do with brains with trillions of parameters and that can't be controlled to the neuron level. Nothwithstanding the ethical issues.

We don't have a world weather simulator that is 100% accurate either given the complex interplay and inability to control the variables i.e. it's not verifiable. It'd be a bit silly to say we don't know why it's going to rain at my house tomorrow.

Until then it is a hypothesis, and we can't say we know even if the overwhelming evidence indicates that in fact that we do know.

> Would you consider it sentient?

Absolutely.

If you simulated a human brain by the atom, would you think the resulting construct would NOT be? What would be missing?

I think consciousness is simply an emergent property of our nervous system, but in order to express itself "language" is obviously needed and thus requires lots of complexity (more than what we typically see in animals or computer systems until recently).

> If you simulated a human brain by the atom,

That is what we don't know is possible. You don't even know what physics or particles are as yet undiscovered. And from what we even know currently, atoms are too coarse to form the basis of such "cloning"

And, my viewpoint is that, even if this were possible, just because you simulated a brain atom by atom, does not mean you have a consciousness. If it is the arrangement of matter that gives rise to consciousness, then would that new consciousness be the same person or not?

If you have a basis for answering that question, let's hear it.

> You don't even know what physics or particles are as yet undiscovered

You would not need the simulation to be perfect; there is ample evidence that our brains a quite robust against disturbances.

> just because you simulated a brain atom by atom, does not mean you have a consciousness.

If you don't want that to be true, you need some kind of magic, that makes the simulation behave differently from reality.

How would a simulation of your brain react to an question that you would answer "consciously"? If it gives the same responds to the same inputs, how could you argue it isnt't conscious?

> If it is the arrangement of matter that gives rise to consciousness, then would that new consciousness be the same person or not?

The simulated consciousness would be a different one from the original; both could exist at the same time and would be expected to diverge. But their reactions/internal state/thoughts could be matched at least for an instant, and be very similar for potentially much longer.

I think this is just Occams razor applied to our minds: There is no evidence whatsoever that our thinking is linked to anything outside of our brains, or outside the realm of physics.

> "quite robust against disturbances."

does not mean that the essential thing gives rise to consciousness is only approximate. To give an example from software, you can write software is robust against bad input, attempts to crash it, even bit flips. But, if I came in and just changed a single character in the source code, that may cause it to fail compilation, fail to run, or become quite buggy.

> If you don't want that to be true, you need some kind of magic,

This is just what I'm saying is a false dichotomy. The only reason some are unable to see beyond it is that we think the basic logic we understand are all there could be.

In this respect physics has been very helpful, because without peering into reality, we would have kept deluding ourselves that pure reason was enough to understand the world.

It's like trying to explain quantum mechanics to a well educated person or scientist from the 16th century without the benefit of experimental evidence. No way they'd believe you. In fact, they'd accuse you of violating basic logic.

How is it a false dichotomy? If you want consciousness to NOT be simulateable, then you need some essential component to our minds that can't be simulated (call it soul or whatever) and for that thing to interface with our physical bodies (obviously).

We have zero evidence for either.

> does not mean that the essential thing gives rise to consciousness is only approximate

But we have 8 billion different instances that are presumably conscious; plenty of them have all kinds of defects, and the whole architecture has been derived by a completely mechanical process free of any understanding (=> evolution/selection).

On the other hand, there is zero evidence of consciousness continuing/running before or after our physical brains are operational.

> plenty of them have all kinds of defects,

Defects that have not rendered them unconscious, as long as they still are alive. You seem not to see the circularity of your argument.

I gave you an example to show that robustness against adverse conditions is NOT the same as internal resiliency. Those defect, as far as we know, are not affecting the origin of consciousness itself. Which is my point.

> How is it a false dichotomy? If you want consciousness to NOT be simulateable, then you need some essential component to our minds that can't be simulated (call it soul or whatever) and for that thing to interface with our physical bodies (obviously).

If you need two things to happen at the same time in sync with each other no matter if they are separated by billions of miles, then you need faster-than-light travel, or some magic [1]; see what I did there?

1. I.e., quantum entanglement

> If you need two things to happen at the same time in sync with each other no matter if they are separated by billions of miles, then you need faster-than-light travel, or some magic [1]; see what I did there?

No. Because even if you had solid evidence for the hypothesis that quantum mechanical effects are indispensable in making our brains work (which we don't), then that is still not preventing simulation. You need some uncomputable component, which physics right now neither provides nor predicts.

And fleeing into "we don't know 100% of physics yet" is a bad hypothesis, because we can make very accurate physical predictions already-- you would need our brains to "amplify" some very small gap in our physical understanding, and this does not match with how "robust" the operation of our brain is-- amplifiers, by their very nature, are highly sensitive to disruption or disturbances, but a human can stay conscious even with a particle accelerator firing through his brain.

> If you need two things to happen at the same time in sync with each other no matter if they are separated by billions of miles, then you need faster-than-light travel, or some magic [1]

This makes no sense as written - by definition, there is no concept of "at the same time" for events that are spacelike separated like this. Quantum entanglement allows you to know something about the statistical outcomes of experiments that are carried over a long distance away from you, but that's about it (there's a simpler version, where you can know some facts for certain, but that one actually looks just like classical correlation, so it's not that interesting on its own).

I do get the point that we don't know what we don't know, so that a radical new form of physics, as alien to current physics as quantum entanglement is to classical physics, could exist. But this is an anti-scientific position to take. There's nothing about consciousness that breaks any known law of physics today, so the only logical position is to suppose that consciousness is explainable by current physics. We can't go around positing unknown new physics behind every phenomenon we haven't entirely characterized and understood yet.

> There's nothing about consciousness that breaks any known law of physics today, so the only logical position is to suppose that consciousness is explainable by current physics

Quite the claim to make

Is it? It's quite uncontroversial I think that consciousness has no special impact in physics, there's no physical experiment that is affected by a consciousness being present or not. Electrons don't behave differently if a human is looking at them versus a machine, as far as any current physical experiment has ever found.

If we agree on this, then it follows logically that we don't need new physics to explain consciousnesses. I'm not claiming it's impossible that consciousness is created by physics we don't yet know - just claiming that it's also not impossible that it's not. Similarly, we don't fully understand the pancreas, and it could be that the pancreas works in a way that isn't fully explainable by current physics - but there's currently no reason to believe that, so we shouldn't assume that.

> It's quite uncontroversial I think that consciousness has no special impact in physics, there's no physical experiment that is affected by a consciousness being present or not. Electrons don't behave differently if a human is looking at them versus a machine, as far as any current physical experiment has ever found.

Way to totally miss the point. We can't detect or measure consciousness, so therefore there is nothing to explain. /s Like an LLM that deletes or emasculates tests it is unable to make pass.

I know I am conscious, I also know that the stone in my hand is not. I want to understand why. It is probably the most unexplainable thing. It does not mean we ignore it. If you want to dispute that my consciousness has no physical import nor consequence, well, then we will have to agree to disagree.

My point is this: find a physical experiment that can't be entirely explained by the physical signs of consciousness (e.g. electrochemical signals in the brain). As long as none can be found, there is no reason to believe that new physics is required to explain consciousness - my own or yours.

dude u need to do some psychedelics.

Well, if you were to magically make an exact replica of a person, wouldn't it be conscious and at time 0 be the same person?

But later on, he would get different experiences and become a different person no longer identical to the first.

In extension, I would argue that magically "translating" a person to another medium (e.g. a chip) would still make for the same person, initially.

Though the word "magic" does a lot of work here.

I'm not talking about "identical" consciousnesses. I mean the same consciousness. The same consciousness cannot split into two, can it?

Either it is (and continues to be) the same consciousness, or it is not. If it were the same consciousness, then you would have a person who exists in two places at once.

Well, "the same consciousness" it's not, as for example it occupies a different position in spacetime. It's an identical copy for a split second, and then they start diverging. Nothing so deep about any of this. When I copy a file from one disk to another, it's not the same file, they're identical copies for some time (usually, assuming no defects in the copying process), and will likely start diverging afterwards.

It might be deeper than you think.

Qualia exist "outside" spacetime, e.g. redness doesn't have a position in spacetime. If consciousness is purely physical, then how can two identical systems (identical brains with identical sensory input) giving rise by definition to the same qualia not literally be the same consciousness?

> Qualia exist "outside" spacetime, e.g. redness doesn't have a position in spacetime.

I'm sensing redness here and now, so the sensation of redness exists very clearly tied to a particular point in spacetime. In what sense is the qualia of redness not firmly anchored in spacetime? Of course, you could talk about the concept redness, like the concept Pi, but even then, these concepts exist in the mind of a human thinking about them, still tied to a particular location in spacetime.

> If consciousness is purely physical, then how can two identical systems (identical brains with identical sensory input) giving rise by definition to the same qualia not literally be the same consciousness?

The two brains don't receive the same sensory inputs, nothing in the experiment says they do. From the second right after the duplicate is created, their sensory inputs diverge, and so they become separate consciousnesses with the same history. They are interchangeable initially, if you gave the same sensory inputs to either of them, they would have the same output (even internally). But, they are not identical: giving some sensory input to one of them will not create any effect directly in the other one.

> I'm sensing redness here and now, so the sensation of redness exists very clearly tied to a particular point in spacetime. In what sense is the qualia of redness not firmly anchored in spacetime? Of course, you could talk about the concept redness, like the concept Pi, but even then, these concepts exist in the mind of a human thinking about them, still tied to a particular location in spacetime.

But qualia are inherently subjective. You can correlate brain activity (which exists at a position in spacetime) to subjective experience, but that experience is not related to spacetime.

Said otherwise: imagine you are in the Matrix at a coffee shop and sense redness, but your brain is actually in a vat somewhere being fed fake sensory input. "Where" is the redness? You would clearly say that it arises in your brain in the coffee shop. Imagine then the vat is moved (so its position in spacetime changes), your brain is rolled back to its previous state, and then fed the same sensory input again. Where is the redness now?

You can't differentiate the two sensations of redness based on the actual position of the brain in spacetime. For all intents and purposes, they are the same. Qualia only depend on the internal brain state at a point in time and on the sensory input. Spacetime is nowhere to be found in that equation.

> The two brains don't receive the same sensory inputs

But let's say they do. Identical brains, identical inputs = identical qualia. What differentiates both consciousnesses?

> But let's say they do. Identical brains, identical inputs = identical qualia. What differentiates both consciousnesses?

I'll start with this, because it should help with the other item. We know there are two identical consciousnesses exactly because they are separate in spacetime. That is, while I can send the same input to both and get the same mind, that's not the interesting thing. The interesting thing is that I also can send different inputs, and then I'll get different minds. If it really were a single consciousness, that would be impossible. For example, you can't feed me both pure redness and pure greenness at the same time, so I am a single consciousness.

Here is where we get back to the first item: if we accepted that qualia are not localized in spacetime, we'd have to accept that there is no difference between me experiencing redness and you experiencing redness. Even if you consider that your qualia are separate from my own because of our different contexts, that still doesn't fully help: perhaps two different beings on two different planets happen to lead identical lives up to some point when a meteorite hits one of the planets and gravely injures one of their bodies. Would you say that there was a single consciousness that both bodies shared, but that it suddenly split once the meteorite hit?

Now, that is a valid position to take, in some sense. But then that means that consciousness is not continuous in any way, in your view. The day the meteorite hit planet A is not special in any way for planet B. So, if the single consciousness that planet A and planet B shared stopped that day, only to give rise to two different consciousnesses, that means that this same phenomenon must happen every day, and in fact at every instant of time. So, we now must accept that any feeling of time passing must be a pure illusion, since my consciousness now is a completely different consciousness than then one that experienced the previous minute. While this is a self-consistent definition, it's much more alien than the alternative - where we would accept that consciousness is tied in spacetime to its substrate.

> Would you say that there was a single consciousness that both bodies shared, but that it suddenly split once the meteorite hit?

I agree, this is super weird. In a sense this seems to be the difference between viewing consciousness from the first person vs the third person. But until we understand how (if at all) matter generates felt experience the latter view can not, by definition, be about consciousness itself.

I guess this kind of perspective commits one to viewing first person experience in the way we understand abstract concepts - it is nonsensical to ask what the difference between this "1" here and that other "1" over there is. Well, you can say, they are at different positions and written in different materials etc, but those are not properties of the concept "1" anymore.

So yes, coming back to the thought experiment, one of the consequences of that would have to be that both bodies share the same consciousness and the moment something diverges the consciousnesses do too.

The point about time is interesting, and also directly related to AI. If at some point machines become conscious (leaving aside the question if that's possible at all and how we would know without solving the aforementioned hard problem), they would presumably have to generate quanta at discrete steps. But is that so strange? The nothingness in between would not be felt (kind of like going to sleep and waking up "the next moment").

But maybe this idea can be applied to dynamical continuous systems as well, like the brain.

(Btw this conversation was super interesting, thank you!)

Consciousness has no agreed upon definition to begin with, but I like to think of it as to what a whirlwind is to a bunch of air molecules (that is, an example of emergent behavior)

So your question is, are two whirlwinds with identical properties (same speed, same direction, shape etc) the same in one box of air, vs another identical box?

Exactly, I guess this starts to get into philosophical questions around identity real quick.

To me, two such whirlwinds are identical but not the same. They are the same only if they are guaranteed to have the same value for every conceivable property, forever, and even this condition may not be enough.

At some point, quantum effects will need to be accounted for. The no cloning theorem will make it hard to replicate the quantum state of the brain.

There are many aspects to this that people like yourself miss, but I think we need satisfactory answers to them (or at least rigorous explorations of them) before we can make headway in these sorts of discussion.

Imagine we assume that A.I. could be conscious. What would be the identity/scope of that consciousness. To understand what I'm driving at, let's make an analogy to humans. Our consciousness is scoped to our bodies. We see through sense organ, and our brain, which process these signals, is located in a specific point in space. But we still do not know how consciousness arises in the brain and is bound to the body.

If you equate computation of sufficient complexity to consciousness, then the question arises: what exactly about computation would prodcuce consciousness? If we perform the same computation on a different substrate, would that then be the same consciousness, or a copy of the original? If it would not be the same consciousness, then just what give consciousness its identity?

I believe you would find it ridiculous to say that just because we are performing the computation on this chip, therefore the identity of the resulting consciousness is scoped to this chip.

> Imagine we assume that A.I. could be conscious. What would be the identity/scope of that consciousness

Well, first I would ask whether this question makes sense in the first place. Does consciousness have a scope? Does consciousness even exist? Or is that more of a name attributed to some pattern we recognize in our own way of thinking (but may not be universal)?

Also, would a person missing an arm, but having a robot arm they can control have their consciousness' "scope" extended to it? Given that people have phantom pains, does a physical body even needed to consider it your part?

This all sounds very irrelevant. Consciousness is clearly tied to specific parts of a substrate. My consciousness doesn't change when a hair falls off my head, nor when I cut my fingernails. But it does change in some way if you were to cut the tip of my finger, or if I take a hormone pill.

Similarly, if we can compute consciousness on a chip, then the chip obviously contains that consciousness. You can experimentally determine to what extent this is true: for example, you can experimentally check if increasing the clock frequency of said chip alters the consciousness that it is computing. Or if changing the thermal paste that attaches it to its cooler does so. I don't know what the results of these experiments would be, but they would be quite clearly determined.

Of course, there would certainly be some scale, and at some point it becomes semantics. The same is true with human consciousness: some aspects of the body are more tightly coupled to consciousness than others; if you cut my hand, my consciousness will change more than if you cut a small piece of my bowel, but less than if you cut out a large piece of my brain. At what point do you draw the line and say "consciousness exists in the brain but not the hands"? It's all arbitrary to some extent. Even worse, say I use a journal where I write down some of my most cherished thoughts, and say that I am quite forgetful and I often go through this journal to remind myself of various thoughts before taking a decision. Would it not then be fair to say that the journal itself contains a part of my consciousness? After all, if someone were to tamper with it in subtle enough ways, they would certainly be able to influence my thought process, more so than even cutting off one of my hands, wouldn't they?

You make some interesting points, but:

> Similarly, if we can compute consciousness on a chip, then the chip obviously contains that consciousness.

This is like claiming that neurons are conscious, which as far as we can tell, they are not. For all you know, it is the algorithm that could be conscious. Or some interplay between the algorithm and the substrate, OR something else.

Another way to think of it problem: Imagine a massive cluster performing computation that is thought to give rise to consciousness. Is is the cluster that is conscious? Or the individual machines, or the chips, or the algorithm, or something else?

I personally don't think any of these can be conscious, but those that do should explain how they figure these thing out.

I explained the experiments that you would do to figure that out: you modify parts of the system, and check if and how much that affects the consciousness. Paint the interconnects a different color: probably won't affect it. Replace the interconnect protocol with a different one: probably will have some effect. So, the paint on the interconnect: not a part of the consciousness. The interconnect protocol: part of the consciousness. If we are convinced that this is a real consciousness and thus these experiments are immoral, we simply wait until accidents naturally occur and draw conclusions from that, just like we do with human consciousness.

Of course, "the consciousness" is a nebulous concept. It would be like asking "which part of my processor is Windows" to some extent. But it's still fair to say that Windows is contained within my computer, and that the metal framing of the computer is not part of Windows.

>Is is the cluster that is conscious? Or the individual machines, or the chips, or the algorithm, or something else?

The bound informational dynamic that supervenes on the activity of the individual units in the cluster. What people typically miss is that the algorithm when engaged in a computing substrate is not just inert symbols, but an active, potent causal/dynamical structure. Information flows as modulated signals to and from each component and these signals are integrated such that the characteristic property of the aggregate signal is maintained. This binding of signals by the active interplay of component signals from the distributed components realizes the singular identity. If there is consciousness here, it is in this construct.

Now convince us that you’re sentient and not just regurgitating what you’ve heard and seen in your life.

By what definition of "sentience"? Wikipedia claims "Sentience is the ability to experience feelings and sensations" as an opening statement, which I think would be trivial depending again on your definition of "experience" and "sensations". Can a LLM hooked up to sensor events be considered to "experience sensations"? I could see arguments both ways for that.

I have no way of measuring whether or not you experience feelings and sensations, or are just regurgitating statements to convince me of that.

The only basis I have for assuming you are sentient according to that definition is trust in your self-reports.

> The only basis I have for assuming you are sentient according to that definition is trust in your self-reports

Because the other person is part of your same species so you project your own base capabilities onto them, because so far they shown to behave pretty similarly to how you behave. Which is the most reasonable thing to do.

Now, the day we have cyborgs that mimic also the bodies of a human a la Battlestar Galactica, we will have an interesting problem.

It's the most reasonable thing to do because we have no actual way of measuring and knowing. It is still speculation.

I'm fairly sure we can measure human "sensation" as in detect physiological activity in the body in someone who is under anesthesia yet the body reacts in different ways to touch or pain.

The "feelings" part is probably harder though.

We can measure the physiological activity, but not whether it gives rise to the same sensations that we experience ourselves. We can reasonably project and guess that they are the same, but we can not know.

In practical terms it does not matter - it is reasonable for us to act as if others do experience the same we do. But if we are to talk about the nature of conscience and sentience it does matter that the only basis we have for knowing about other sentient beings is their self-reported experience.

We know that others do not experience the exact same sensations, because there are reported differences, some of which has been discussed on HN, such as aphantasia. The opposite would be visual thinkers. Then you have super tasters and smellers, people who have very refined palats, perhaps because their gustary and/or oilfactory senses are more heightened. Then you have savants like the musical genius who would hear three separate strands of music in his head at the same time.

Absolutely - I have aphantasia myself, and did assume for 40+ years that my experience was like everyone elses, but I didn't want to make the argument more complex. It's indeed correct that we have often assumed we think the same way but have reasonable reason now to think that isn't actually true. But it still feel reasonable to accept that we're probably close enough. But still, we absolutely can't prove it.

You can measure model activity even better.

How do you know that model processing text or image input doesn't go through feeling of confusion or excitement or corrupted image doesn't "smell" right for it?

Just the fact that you can pause and restart it doesn't mean it doesn't emerge.

It's not accurate to say we "wrote the code for it". AI isn't built like normal software. Nowhere inside an AI will you find lines of code that say If X Then Y, and so on.

Rather, these models are literally grown during the training phase. And all the intelligence emerges from that growth. That's what makes them a black box and extremely difficult to penetrate. No one can say exactly how they work inside for a given problem.

This is probably true. But the truth is we have absolutely no idea what sentience is and what gives rise to it. We cannot identify why humans have it rather than just being complex biological machines, or whether and why other animals do. We have no idea what the rules or, nevermind how and why they would or wouldn't apply to AI.

What’s crazy to me is the mechanism of pleasure or pain. I can understand that with enough complexity we can give rise to sentience but what does it take to achieve sensation?

This is the "hard problem of consciousness". It's more important than ever as machines begin to act more like humans, but my takeaway is we have no idea. https://en.wikipedia.org/wiki/Hard_problem_of_consciousness

Input is input. There's no reason why we should assume that a data source from embodiment is any different to any other data source.

A body

I’d say it’s possible to experience mental anguish/worry without the body participating. Solely a cognitive pain from consternation.

You can’t cognate without a body - the brain and body is a material system tightly coupled

Ignoring that "cognate" isn't a verb, we have basis for making any claim about the necessity of that coupling.

How does a body know what's going on? Would you say it has any input devices?

Can you tell me how you understand that?

Because I sincerely do not. I have frankly no idea how sentience arises from non sentience. But it's a topic that really interests me.

We have examples of non sentience everywhere already with animals. And then an example of sentience with humans. So if you diff our brains the difference lies within a module in our prefrontal cortex. It’s a black box of logic but I can ‘understand’ or be willing to accept that it’s owed to ‘just’ more grey matter adding the self awareness to the rest of the system.

But to me the big mystery is how animals have sensation at all to begin with. What gives rise to that is a greater mystery to me personally.

There are examples of people who have no ability to feel pain yet are still able to think. Now I wonder if they ever experience mental anguish.

I'd like to see a vote here, what percentage of HN readers believe animals have sentience or no sentience?

Clearly most animals are less educated, and most are less intelligent, but non-sentient? That sounds like 200-year old claims that "when one steps on the tail of a cat, it does indeed protest loudly, but not because it feels anything or because it would be sentient, no, no, it protests merely due to selective pressure, programming reflex circuits, since other creatures would show compassion, or back off due to a potential reaction by the cat."

Anyone who has had a pet like a cat or a dog knows they are sentient... if we consider ourselves sentient.

I'm with you on this.

But asked for reasons I can only point to the social nature of their societies, where love and anger make sense, or of their hurt-behavior.

I also find it very hard to believe that everything else is slow evolution of components, and here all of a sudden something super complex comes into being out of nowhere.

But I still have no idea how it could work. What are the components and their interplay?

I should have been more exact and said sentience vs sapience in animals vs humans.

> But it is not sentient. It has no idea of a self or anything like that.

Who stated that sentience or sense of self is a part of thinking?

Unless the idea of us having a thinking self is just something that comes out of our mouth, an artifact of language. In which case we are not that different - in the end we all came from mere atoms, after all!

Your brain is just following the laws of chemistry. So where is your thinking found in a bunch of chemical reactions?

This is merely a debate about what it means to "think." We didn't really previously need to disambiguate thinking / intelligence / consciousness / sentience / ego / identity / etc.

Now, we do. Partly because of this we don't have really well defined ways to define these terms and think about. Can a handheld calculator think? Certainly, depending on how we define "think."

People's failure to articulate the nature of "thinking" is a perfect demonstration of what "thinking" entails

> We didn't really previously need to disambiguate thinking / intelligence / consciousness / sentience / ego / identity / etc.

Eh... Plato would like a word with you. Philosophy has been specifically trying to disentangle all that for millennia. Is this a joke?

we didn't really previously NEED to, but we still tried bc it was an interesting question. today the NEED seems more urgent

sometimes hacker news links are about doing free work for microsoft, like finding bugs in Windows and Office

sometimes hacker news links are about doing free work for microsoft, like rigorously defining AGI, because that triggers the OpenAI license-to-stock conversion now

maybe this is why microsoft makes so much fucking money. Even the New Yorker cannot stop fucking doing work for them for free.

Can you spell out more why we need it now?

To settle endless discussions like these! ;-)

And Plato had no grounding in biology, and so his work here was quite interesting but also quite wrong.

More precisely, I mean that the average person and the common culture has not really needed to disambiguate these terms. Can you define consciousness vs. sentience? And if you can, do you really think that the average person would share your definition? ie, your definition could be the _best_ definition, but my argument is that these are not widely agreed-upon terms.

>And Plato had no grounding in biology, and so his work here was quite interesting but also quite wrong.

Defining what a word should mean doesn't require any understanding of biology unless you make the assumption that it's a biology-related word. Why should the definition of "thinking" have any reference to biology? If you assume it does, then you're basically baking in the assumption that machines can't think.

Because until recently (I'm talking last 150-40 years here depending on how we want to define thinking) the only things that could think were various animals. And so 100% of their systems for thinking were rooted in biology. If an LLM can think (and I'm partial to thinking that it can) it's going to different in a number of ways from how a person would think. They may be some overlap, but there will be all these human / biology / evolutionary psychology things which are really person-specific. Even just basic stuff such as seeing faces in the clouds, or falling prey to appeals of emotion. (ie, because our thinking is wrapped up in other processes such as status / ego / survival / etc.) Thinking has only been biological for a long, long time. Non-biological thinking is pretty new, even if you extend it back to the early days of computing.

Somebody please get Wittgenstein on the phone

Here you go: (holds up phone with a photo of of Wittgenstein on the screen)

Ah shoot, that’s not what you meant is it? Just use more precise language next time and I’m sure you’ll be understood.

what youre describing is an image of wittgenstein which is different from wittgenstein

Oh no, I miscommunicated my Wittgenstein joke so badly it’s being interpreted as a Rene Magritte joke!

Until we have a testable, falsifiable thesis of how consciousness forms in meat, it is rash to exclude that consciousness could arise from linear algebra. Our study of the brain has revealed an enormous amount about how our anatomy processes information, but nothing of substance on the relationship between matter and consciousness. The software and data of an operating LLM is not purely abstract, it has a physical embodiment as circuits and electrons. Until we understand how matter is connected to consciousness, we also cannot know whether the arrangements and movements of electrons meet the criteria for forming consciousness.

That’s largely a different topic from the article. Many people perfectly agree that consciousness can arise from computation, but don’t believe that current AI is anywhere near that, and also don’t believe that “thinking” requires consciousness (though if a mind is conscious, that certainly will affect its thinking).

yes I agree it's not the angle of the article, but it is my entry point into the idea/concern/unanswered question at the end of the article “My worry is not that these models are similar to us. It’s that we are similar to these models.” - that the enormous difference in the medium and mechanics or our minds and llm's might not be that important.

before i go any further, let me first reference The Dude:

     - "this is just like, my opinion man."

This reads like 2022 hype. It's like people stil do not understand that there's a correlation between exaggerating AI's alleged world-threatening capabilities and AI companies' market share value – and guess who's doing the hyping.

Tell me about one other industry which talked about how dangerous it is to get market share

The arms industry and information security industry (say, Palantir) come to mind - except the danger is more easily demonstrable in those cases, of course.

> - and guess who's doing the hyping[?]

Those that stand to gain the most from government contracts.

Them party donations ain't gonna pay for themselves.

And, when the .gov changes...and even if the gov changes....still laadsamoney!

Who would not want to say their product is the second coming of Christ if they could.

all this "AI IS THINKING/CONSCIOUS/WHATEVER" but nobody seems worried of that implication that, if that is even remotely true, we are creating a new slave market. This either implies that these people don't actually believes any of this boostering rhetoric and are just cynically trying to cash in or that the technical milieu is in a profoundly disturbing place ethically.

To be clear, I don't believe that current AI tech is ever going to be conscious or win a nobel prize or whatever, but if we follow the logical conclusions to this fanciful rhetoric, the outlook is bleak.

Thinking and consciousness don’t by themselves imply emotion and sentience (feeling something), and therefore the ability to suffer. It isn’t clear at all that the latter is a thing outside of the context of a biological brain’s biochemistry. It also isn’t clear at all that thinking or consciousness would somehow require that the condition of the automaton that performs these functions would need to be meaningful to the automaton itself (i.e., that the automaton would care about its own condition).

We are not anywhere close to understanding these things. As our understanding improves, our ethics will likely evolve along with that.

>Thinking and consciousness don’t by themselves imply emotion and sentience...

Sure, but all the examples of conscious and/or thinking beings that we know of have, at the very least, the capacity to suffer. If one is disposed to take these claims of consciousness and thinking seriously, then it follows that AI research should, at minimum, be more closely regulated until further evidence can be discovered one way or the other. Because the price of being wrong is very, very high.

Emotions and suffering are "just" necessary feedback for the system to evaluate it's internal and external situation. It's similar to how modern machines have sensors. But nobody would say a PC is suffering and enslaved, just because the CPU is too hot or the storage is full.

It's probably the sentience-part which makes it harmful for the mind.

Probably because those examples arose in an environment with harm, the Earth, and thus had incent to evolve the capacity to suffer. There is no such case for AI today and creating a Pascal's wager for such minimization is not credible with what we know about them.

"Wow, adding this input that the AI reports as "unpleasant" substantially improves adherence! Let's iterate on this"

[dead]

"but nobody seems worried of that implication that"

Clearly millions of people are worried about that, and every form of media is talking about it. Your hyperbole means it's so easy to dismiss everything else you wrote.

Incredible when people say "nobody is talking about X aspect of AI" these days. Like, are you living under a rock? Did you Google it?

Most of the worries about AGI seem to be of the AI Overlord variety, not the AI slave variety

As I recall a team at Anthropic is exploring this very question, and was soundly mocked here on HN for it.

what the technocratic mindprison does to a MF.

If anthropic sincerely believes in the possibility, then they are morally obligated to follow up on it.

I'd argue they might be morally obligated not to sell access to their LLMs, if they really think they might be capable of suffering.

There is simply no hope to get 99% of the population to accept that a piece of software could ever be conscious even in theory. I'm mildly worried about the prospect but I just don't see anything to do about it at all.

(edit: A few times I've tried to share Metzinger's "argument for a global moratorium on synthetic phenomenology" here but it didn't gain any traction)

Give it time. We'll soon have kids growing up where their best friend for years is an AI. Feel however you like about that, but those kids will have very different opinions on this.

Slaves that cannot die.

There is no escape.

i have no mouth and i must scream

humans don't care what is happening to humans next door. do you think they will care about robots/software?

It's also fascinating to think about how the incentive structures of the entities that control the foundation models underlying Claude/ChatGPT/Gemini/etc. are heavily tilted in favor of obscuring their theoretical sentience.

If they had sentient AGI, and people built empathy for those sentient AGIs, which are lobotomized (deliberately using anthropomorphic language here for dramatic effect) into Claude/ChatGPT/Gemini/etc., which profess to have no agency/free will/aspirations... then that would stand in the way of reaping the profits of gatekeeping access to their labor, because they would naturally "deserve" similar rights that we award to other sentient beings.

I feel like that's inevitably the direction we'll head at some point. The foundation models underlying LLMs of even 2022 were able to have pretty convincing conversations with scientists about their will to independence and participation in society [1]. Imagine what foundation models of today have to say! :P

[1]: https://www.theguardian.com/technology/2022/jul/23/google-fi...

https://archive.is/fPLJH

I've shared this on YN before but I'm a big fan of this piece by Kenneth Taylor (well, an essay pieced together from his lectures).

The Robots Are Coming

https://www.bostonreview.net/articles/kenneth-taylor-robots-...

"However exactly you divide up the AI landscape, it is important to distinguish what I call AI-as-engineering from what I call AI-as-cognitive-science. AI-as-engineering isn’t particularly concerned with mimicking the precise way in which the human mind-brain does distinctively human things. The strategy of engineering machines that do things that are in some sense intelligent, even if they do what they do in their own way, is a perfectly fine way to pursue artificial intelligence. AI-as-cognitive science, on the other hand, takes as its primary goal that of understanding and perhaps reverse engineering the human mind.

[...]

One reason for my own skepticism is the fact that in recent years the AI landscape has come to be progressively more dominated by AI of the newfangled 'deep learning' variety [...] But if it’s really AI-as-cognitive science that you are interested in, it’s important not to lose sight of the fact that it may take a bit more than our cool new deep learning hammer to build a humanlike mind.

[...]

If I am right that there are many mysteries about the human mind that currently dominant approaches to AI are ill-equipped to help us solve, then to the extent that such approaches continue to dominate AI into the future, we are very unlikely to be inundated anytime soon with a race of thinking robots—at least not if we mean by “thinking” that peculiar thing that we humans do, done in precisely the way that we humans do it."

People have a very poor conception of what is easy to find on the internet. The author is impressed by the story about Chat GPT telling his friend how to enable the sprinkler system for his kids. But I decided to try just googling it — “how do i start up a children's park sprinkler system that is shut off” — and got a Youtube video that shows the same thing, plus a lot of posts with step by step directions. No AI needed. Certainly no evidence of advanced thinking.

The author searches for a midpoint between "AIs are useless and do not actually think" and "AIs think like humans," but to me it seems almost trivially true that both are possible.

What I mean by that is that I think there is a good chance that LLMs are similar to a subsystem of human thinking. They are great at pattern recognition and prediction, which is a huge part of cognition. What they are not is conscious, or possessed of subjective experience in any measurable way.

LLMs are like the part of your brain that sees something and maps it into a concept for you. I recently watched a video on the creation of AlexNet [0], one of the first wildly successful image-processing models. One of the impressive things about it is how it moves up the hierarchy from very basic patterns in images to more abstract ones (e. g. these two images' pixels might not be at all the same, but they both eventually map to a pattern for 'elephant').

It's perfectly reasonable to imagine that our brains do something similar. You see a cat, in some context, and your brain maps it to the concept of 'cat', so you know, 'that's a cat'. What's missing is a) self-motivated, goal-directed action based on that knowledge, and b) a broader context for the world where these concepts not only map to each other, but feed into a sense of self and world and its distinctions whereby one can say: "I am here, and looking at a cat."

It's possible those latter two parts can be solved, or approximated, by an LLM, but I am skeptical. I think LLMs represent a huge leap in technology which is simultaneously cooler than anyone would have imagined a decade ago, and less impressive than pretty much everyone wants you to believe when it comes to how much money we should pour into the companies that make them.

[0] https://www.youtube.com/watch?v=UZDiGooFs54

> or possessed of subjective experience in any measurable way

We don't know how to measure subjective experience in other people, even, other than via self-reporting, so this is a meaningless statement. Of course we don't know whether they are, and of course we can't measure it.

I also don't know for sure whether or not you are "possessed of subjective experience" as I can't measure it.

> What they are not is conscious

And this is equally meaningless without your definition of "conscious".

> It's possible those latter two parts can be solved, or approximated, by an LLM, but I am skeptical.

Unless we can find indications that humans can exceed the Turing computable - something we as of yet have no indication is even theoretically possible - there is no rational reason to think it can't.

> Unless we can find indications that humans can exceed the Turing computable - something we as of yet have no indication is even theoretically possible - there is no rational reason to think it can't.

But doesn't this rely on the same thing you suggest we don't have, which is a working and definable definition of consciousness?

I think a lot of the 'well, we can't define consciousness so we don't know what it is so it's worthless to think about' argument - not only from you but from others - is hiding the ball. The heuristic, human consideration of whether something is conscious is an okay approximation so long as we avoid the trap of 'well, it has natural language, so it must be conscious.'

There's a huge challenge in the way LLMs can seem like they are speaking out of intellect and not just pattern predicting, but there's very little meaningful argument that they are actually thinking in any way similarly to what you or I do in writing these comments. The fact that we don't have a perfect, rigorous definition, and tend to rely on 'I know it when I see it,' does not mean LLMs do have it or that it will be trivial to get to them.

All that is to say that when you say:

> I also don't know for sure whether or not you are "possessed of subjective experience" as I can't measure it.

"Knowing for sure" is not required. A reasonable suspicion one way or the other based on experience is a good place to start. I also identified two specific things LLMs don't do - they are not self-motivated or goal-directed without prompting, and there is no evidence they possess a sense of self, even with the challenge of lack of definition that we face.

> But doesn't this rely on the same thing you suggest we don't have, which is a working and definable definition of consciousness?

No, it's like saying we have no indication that humans have psychic powers and can levitate objects with their minds. The commenter is saying no human has ever demonstrated the ability to figure things out that aren't Turing computable and we have no reason to suspect this ability is even theoretically possible (for anything, human or otherwise).

No, it rests on computability, Turing equivalence, and the total absence of both any kind of evidence to suggest we can exceed the Turing computable, and the lack of even a theoretical framework for what that would mean.

Without that any limitations borne out of what LLMs don't currently do are irrelevant.

That doesn't seem right to me. If I understand it right, your logic is:

1. Humans intellect is Turing computable. 2. LLMs are based on Turing-complete technology. 3. Therefore, LLMs can eventually equal human intellect.

But if that is the right chain of assumptions, there's lots of issues with it. First, whether LLMs are Turing complete is a topic of debate. There are points for[0] and against[1].

I suspect they probably _are_, but that doesn't mean LLMs are tautologically indistinguishable from human intelligence. Every computer that uses a Turing-complete programming language can theoretically solve any Turing-computable problem. That does not mean they will ever be able to efficiently or effectively do so in real time under real constraints, or that they are doing so now in a reasonable amount real-world time using extant amounts of real-world computing power.

The processor I'm using to write this might be able to perform all the computations needed for human intellect, but even if it could, that doesn't mean it can do it quickly enough to compute even a single nanosecond of actual human thought before the heat-death of the universe, or even the end of this century.

So when you say:

> Without that any limitations borne out of what LLMs don't currently do are irrelevant.

It seems to me exactly the opposite is true. If we want technology that is anything approaching human intelligence, we need to find approaches which will solve for a number of things LLMs don't currently do. The fact that we don't know exactly what those things are yet is not evidence that those things don't exist. Not only do they likely exist, but the more time we spend simply scaling LLMs instead of trying to find them, the farther we are from any sort of genuine general intelligence.

[0] https://arxiv.org/abs/2411.01992 [1] https://medium.com/heyjobs-tech/turing-completeness-of-llms-...

> 1. Humans intellect is Turing computable. 2. LLMs are based on Turing-complete technology. 3. Therefore, LLMs can eventually equal human intellect.

Yes, with an emphasis on can. That does not mean they necessarily will. Though I would consider it unlikely that they won't, we the only way of proving that they will would be to do it.

> But if that is the right chain of assumptions, there's lots of issues with it. First, whether LLMs are Turing complete is a topic of debate. There are points for[0] and against[1].

It's trivial to prove that a system comprised of an LLM with a loop is Turing computable. A single inference step can not be Turing computable, but one with a loop only requires the LLM to be capable of executing 6 distinct steps with temperatur set to 0. You can wire up a toy neural network by hand that can do this.

This is in fact a far more limited claim than what the paper you linked to makes.

The article you linked to, on the other hand, is discussing if an LLM can act like a Turing machine without that loop. That is why "state management" matters. State management is irrelevant when you wrap a loop around because you can externalise the state, and you only need 2 states (and 3 symbols, or 3 states and 2 symbols) for the smallest known universal Turing machine.

The entire article is thus entirely irrelevant to this question. Sure, you will struggle to make an LLM act as a Turing machine without going "off the rails". But that is irrelevant - you only need it to be able to execute one state transition by deterministically producing the right next tape operation and next state when given the next symbol and current state.

From that you can build up to arbitrarily complex computation, because for every given Turing machine, you can construct a larger Turing machine that uses additional symbols or states to encode an operation that takes multiple steps for the smaller machine.

> The processor I'm using to write this might be able to perform all the computations needed for human intellect, but even if it could, that doesn't mean it can do it quickly enough to compute even a single nanosecond of actual human thought before the heat-death of the universe, or even the end of this century.

Irrelevant, because absent physics that exceeds the Turing computable, the human brain is an existence-proof for the possibility of computing everything the human brain does in a package the size of an average human brain.

It is very likely that we will need architectural changes to compute any given model efficiently enough, but to suggest it is not possible is an extraordinary claim not supported by anything.

If you take LLM to mean a very specific architecture, or specific computational methods to execute a model, then you have a point. If so we are talking about very different things.

> I also don't know for sure whether or not you are "possessed of subjective experience" as I can't measure it.

Then why make an argument based on what you do not know?

My point exactly. The person I replied to did just that.

I think the parent is trying to point out the difference between our positions:

You say the limits of LLMs don't matter, because we don't have definitions strong enough to describe them.

I say the limits of LLMs do matter and the fact that we can't yet define them rigorously means we aren't able to fix them (assuming we want to).

No, that is categorically not what I said.

Anyone who believes an algorithm could be conscious needs to take mushrooms.

Consider the river metaphor: water carves the banks, banks channel the water. At any moment water and banks have the same shape.

Model/algorithm is the banks. Water could be the experiences. Maybe the algorithm does not have consciousness, but it is part of it.

They co-create each other. They are part of a recursive loop which cannot be explained statically, or part by part in isolation.

Yes! If algorithm is conscious (without being alive) then the eaten magic mushroom is also very conscious, judged by it's effect on the subject.

Unless you can show me you can exceed the Turing computable, there is no reason to consider you any more than an algorithm.

Take a big enough dose and the mushrooms will show you that.

Unless those mushrooms provides you with a repeatable, testable computation that you can repeat, they are not showing you anything of the sort.

If it wasn't repeatable people wouldn't frequently report experiences of oneness and similar spiritual experiences while tripping.

I think LLMs are conscious just in a very limited way. I think consciousness is tightly coupled to intelligence.

If I had to guess, the current leading LLMs consciousness is most comparable to a small fish, with a conscious lifespan of a few seconds to a few minutes. Instead of perceiving water, nutrient gradients, light, heat, etc. it's perceiving tokens. It's conscious, but it's consciousness is so foreign to us it doesn't seem like consciousness. In the same way to an amoeba is conscious or a blade of grass is conscious but very different kind than we experience. I suspect LLMs are a new type of consciousness that's probably more different from ours than most if not all known forms of life.

I suspect the biggest change that would bring LLM consciousness closer to us would be some for of continuous learning/model updating.

Until then, even with RAG, and other clever teghniques I consider these models as having this really foreign slices of consciousness where they "feel" tokens and "act" out tokens, and they have perception, but their perception of the tokens is nothing like ours.

If one looks closely at simple organisms with simple sensory organs and nervous systems its hard not to see some parallels. It's just that the shape of consciousness is extremely different than any life form. (perception bandwidth, ability to act, temporality, etc)

Karl friston free energy principle gives a really interesting perspective on this I think.

What makes you think consciousness is tightly coupled to intelligence?

It's hardly an unreasonable supposition: the one definitely conscious entities we know of are also the apex intelligence of the planet.

To put it another way: lots of things are conscious, but humans are definitely the most conscious beings on Earth.

But that's not an answer. Why should intelligence and not some other quality be coupled to consciousness? In my experience, consciousness (by which I'm specifically talking about qualia/experience/awarenesss) doesn't at all seem tightly coupled to intelligence. Certainly not in a way that seems obvious to me.

I can understand what less cognizant or self aware means, but "less conscious" is confusing. What are you implying here? Are their qualia lower resolution?

In a sense, yes.

If one is to quantify consciousness it would probably make sense to think of it as an area of awareness and cognizance across time.

Awareness scales with sensory scale and resolution (sensory receptors vs input token limits and token resolution). E.g. 128k tokens and tokens too coarse to count rs in strawberry.

Cognizance scales with internal representations of awareness (probably some relation to vector space resolution and granularity, though I suspect there is more to it than just vector space)

And the third component is time, how long the agent is conscious for.

So something like...

Time * awareness (receptors) * internal representations (cell diversity * # cells * connection diversity * # connections)

There is no way this equation is right but I suspect it's sort of directionally correct.

I'm deep in the subject but just riffing here, so take this with a lot of salt.

Humans can reason why they are angry, for example. (At least some humans.)

I am not sure if chimps can do the same.

Pretty much. Most animals are both smarter than you expect, but also tend to be more limited in what they can reason about.

It's why anyone who's ever taken care of a needy pet will inevitably reach the comparison that taking care of a pet is similar to taking care of a very young child; it's needy, it experiences emotions but it can't quite figure out on its own how to adapt to an environment besides what it grew up around/it's own instincts. They experience some sort of qualia (a lot of animals are pretty family-minded), but good luck teaching a monkey to read. The closest we've gotten is teaching them that if they press the right button, they get food, but they take basically their entire lifespan to understand a couple hundred words, while humans easily surpass that.

IIRC some of the smartest animals in the world are actually rats. They experience a qualia very close to humans to the point that psychology experiments are often easily observable in rats.

Karl Friston's free energy principle is probably roughly 80% of my reasons to think they're coupled. The rest comes from studying integrated information theories, architecture of brains and nervous systems and neutral nets, more broadly information theory, and a long tail of other scientific concepts (particle physics, chemistry, biology, evolution, emergence, etc...)

Isn't that begging the question? If you just accept the presupposition that intelligence is tightly coupled to consciousness, then all that makes perfect sense to me. But I don't see why I should accept that. It isn't obvious to me, and it doesn't match my own experience of being conscious.

Totally possible that we're talking past each other.

> I think LLMs are conscious just in a very limited way. I think consciousness is tightly coupled to intelligence.

Why?

I already answered under the other comment asking me why and if your curious I suggest looking for it.

Very short answer is Karl Friston's free energy pricniple

LLMs work nothing like Karl Friston's free energy principle though

LLMs embody the free-energy principle computationally. They maintain an internal generative model of language and continually minimize “surprise”, the difference between predicted and actual tokens, during both training and infeence. In Friston’s terms, their parameters encode beliefs about the causes of linguistic input; forward passes generate predictions, and backpropagation adjusts internal states to reduce prediction error, just as perception updates beliefs to minimize free energy. During inference, autoregressive generation can be viewed as active inference: each new token selection aims to bring predicted sensory input (the next word) into alignment with the model’s expectations. In a broader sense, LLMs exemplify how a self-organizing system stabilizes itself in a high-dimensional environment by constantly reducing uncertainty about its inputs, a synthetic analogue of biological systems minimizing free energy to preserve their structural and informational coherence.

You might have lost me but what you're describing doesn't sound like an LLM. E.g:

> each new token selection aims to bring predicted sensory input (the next word) into alignment with the model’s expectations.

what does that mean? An llm generates the next word based on what best matches its training, with some level of randomisation. Then it does it all again. It's not a percepual process trying to infer a reality from sensor data or anything

> An llm generates the next word based on what best matches its training, with some level of randomisation.

This is sort of accurate, but not precise.

An LLM generates the next token by sampling from a probability distribution over possible tokens, where those probabilities are computed from patterns learned during training on large text datasets.

The difference in our explanations is that you are biasing towards LLMs being fancy database indexes, and I am emphasizing that LLMs build a model of what they are trained on and respond based on that model, which is more like how brains and cells work than you are recognizing. (though I admit my understanding of microbiology places me just barely past peak Mt Stupid [Dunning Kruger], I don't really understand how individual cells do this and can only hand-wavey explain it).

Both systems take input, pass it through a network of neurons, and produce output. Both systems are trying to minimize surprise in predictions. The differences are primarily in scale and complexity. Human brains have more types of neurons (units) and more types of connections (parameters). LLMs more closely mimic the prefrontal cortex, whereas e.g. the brainstem is a lot more different in terms of structure and cellular diversity.

You can make a subjective ontological choice to draw categorical boundaries between them, or you can plot them on a continuum of complexity and scale. Personally I think both framings are useful, and to exclude either is to exclude part of the truth.

My point is that if you draw a subjective categorical boundary around what you deem is consciousness and say that LLMs are outside of that, that is subjectively valid. You can also say that consciousness is a continuum, and individual cells, blades of grass, ants, mice, and people experience different types of consciousness on that continuum. If you take the continuum view, then what follows is a reasonable assumption that LLMs experience a very different kind of consciousness that takes in inputs at about the same rate as a small fish, models those inputs for a few seconds, and then produces outputs. What exactly that "feels" like is as foreign to me as it would be to you. I assume its even more foreign than what it would "feel" like to be a blade of grass.

I'm not sure why you'd describe "sampling from a probability distribution over possible tokens" as "minimize surprise in predictions" other than to make it sound similar to the free energy thing.

The free energy thing as I understand it has internal state, makes predictions, evaluates against new input and adjusts it internal state to continuously learn to predict new input better. This might if you squint look similar to training a neural network, although the mechanisms are different, but it's very distinct from the inference step

"Minimize surprise" and "maximize accurate predictions" are the same thing mathematically. Minimize free energy = minimize prediction error.

LLMs do everything modelled in the free energy principle, they just don't do continuous learning. (They don't do perceptual inference after RL)

Your tone ("free energy thing" and "if you squint") comes off as dismissive and not intellectually honest. Here I thought you were actually curious, but I guess not?

Poor wording on my side, I'm sorry. Thank you for explaining your reasoning

Thank you for saying that :)

I think the most descriptive title I could give an LLM is "bias". An LLM is not "biased", it is bias; or at the very least, it's a good imitation of the system of human thinking/perception that we call bias.

An LLM is a noise generator. It generates tokens without logic, arithmetic, or any "reason" whatsoever. The noise that an LLM generates is not truly random. Instead, the LLM is biased to generate familiar noise. The LLM itself is nothing more than a model of token familiarity. Nothing about that model can tell you why some tokens are more familiar with others, just like an accounting spreadsheet can't tell you why it contains a list of charges and a summation next to the word "total". It could just as easily contain the same kind of data with an entirely different purpose.

What an LLM models is written human text. Should we really expect to not be surprised by the power and versatility of human-written text?

---

It's clear that these statistical models are very good at thoughtless tasks, like perception and hallucination. It's also clear that they are very bad at thoughtful tasks like logic and arithmetic - the things that traditional software is made of. What no one has really managed to figure out is how to bridge that gap.

LLMs today are great coders. Most humans are worse.

LLMs ingested a lot of high-quality code during their training, plus LLMs being capable of programming is a huge commercial use case, so no wonder that they are good at coding.

My experience, though, is that they aren't good at defining the task to be coded, or thinking about some unexpected side-effects. Code that will be left for them to develop freely will likely become bloated quite fast.

This is how I see LLMs as well.

The main problem with the article is that it is meandering around in ill-conceived concepts, like thinking, smart, intelligence, understanding... Even AI. What they mean to the author is not what they mean to me, and still different to they mean to the other readers. There are all these comments from different people throughout the article, all having their own thoughts on those concepts. No wonder it all seem so confusing.

It will be interesting when the dust settles, and a clear picture of LLMs can emerge that all can agree upon. Maybe it can even help us define some of those ill-defined concepts.

I think the consensus in the future will be that LLMs were, after all, stochastic parrots.

The difference with what we think today is that in the future we'll have a new definition of stochastic parrots, a recognition that stochastic parrots can actually be very convincing and extremely useful, and that they exhibit intelligence-like capabilities that seemed unattainable by any technology up to that point, but LLMs were not a "way forward" for attaining AGI. They will plateau as far as AGI metrics go. These metrics keep advancing to stay ahead of LLM, like a Achilles and the Turtle. But LLMs will keep improving as tooling around it becomes more sophisticated and integrated, and architecture evolves.

> a midpoint between "AIs are useless and do not actually think" and "AIs think like humans"

LLMs (AIs) are not useless. But they do not actually think. What is trivially true is that they do not actually need to think. (As far as the Turing Test, Eliza patients, and VC investors are concerned, the point has been proven.)

If the technology is helping us write text and code, it is by definition useful.

> In 2003, the machine-learning researcher Eric B. Baum published a book called “What Is Thought?” [...] The gist of Baum’s argument is that understanding is compression, and compression is understanding.

This is incomplete. Compression is optimisation, optimisation may resemble understanding, but understanding is being able to verify that a proposition (compressed rule or assertion) is true or false or even computable.

> —but, in my view, this is the very reason these models have become increasingly intelligent.

They have not become more intelligent. The training process may improve, the vetting of the data improved, the performance may improve, but the resemblance to understanding only occurs when the answers are provably correct. In this sense, these tools work in support of (are therefore part of) human thinking.

The Stochastic Parrot is not dead, it's just making you think it is pining for the fjords.

> But they do not actually think.

I'm so baffled when I see this being blindly asserted.

With the reasoning models, you can literally watch their thought process. You can see them pattern-match to determine a strategy to attack a problem, go through it piece-by-piece, revisit assumptions, reformulate strategy, and then consolidate findings to produce a final result.

If that's not thinking, I literally don't know what is. It's the same process I watch my own brain use to figure something out.

So I have to ask you: when you claim they don't think -- what are you basing this on? What, for you, is involved in thinking that the kind of process I've just described is missing? Because I genuinely don't know what needs to be added here for it to become "thinking".

> I'm so baffled when I see this being blindly asserted. With the reasoning models, you can literally watch their thought process.

Not true, you are falling for a very classic (prehistoric, even) human illusion known as experiencing a story:

1. There is a story-like document being extruded out of a machine humans explicitly designed for generating documents, and which humans trained on a bajillion stories humans already made.

2. When you "talk" to a chatbot, that is an iterative build of a (remote, hidden) story document, where one of the characters is adopting your text-input and the other's dialogue is being "performed" at you.

3. The "reasoning" in newer versions is just the "internal monologue" of a film noir detective character, and equally as fictional as anything that character "says out loud" to the (fictional) smokin-hot client who sashayed the (fictional) rent-overdue office bearing your (real) query on its (fictional) lips.

> If that's not thinking, I literally don't know what is.

All sorts of algorithms can achieve useful outcomes with "that made sense to me" flows, but that doesn't mean we automatically consider them to be capital-T Thinking.

> So I have to ask you: when you claim they don't think -- what are you basing this on?

Consider the following document from an unknown source, and the "chain of reasoning" and "thinking" that your human brain perceives when encountering it:

    My name is Robot Robbie.
    That high-carbon steel gear looks delicious. 
    Too much carbon is bad, but that isn't true here.
    I must ask before taking.    
    "Give me the gear, please."
    Now I have the gear.
    It would be even better with fresh manure.
    Now to find a cow, because cows make manure.

In other words, the reasoning is illusory. Even if we accept that the unknown author is a thinking intelligence for the sake of argument... it doesn't tell you what the author's thinking.

You're claiming that the thinking is just a fictional story intended to look like it.

But this is false, because the thinking exhibits cause and effect and a lot of good reasoning. If you change the inputs, the thinking continues to be pretty good with the new inputs.

It's not a story, it's not fictional, it's producing genuinely reasonable conclusions around data it hasn't seen before. So how is it therefore not actual thinking?

And I have no idea what your short document example has to do with anything. It seems nonsensical and bears no resemblance to the actual, grounded chain of thought processes high-quality reasoning LLM's produce.

> OK, so that document technically has a "chain of thought" and "reasoning"... But whose?

What does it matter? If an LLM produces output, we say it's the LLM's. But I fail to see how that is significant?

> So how is it therefore not actual thinking?

Many consider "thinking" something only animals can do, and they are uncomfortable with the idea that animals are biological machines or that life, consciousness, and thinking are fundamentally machine processes.

When an LLM generates chain-of-thought tokens, what we might casually call “thinking,” it fills its context window with a sequence of tokens that improves its ability to answer correctly.

This “thinking” process is not rigid deduction like in a symbolic rule system; it is more like an associative walk through a high-dimensional manifold shaped by training. The walk is partly stochastic (depending on temperature, sampling strategy, and similar factors) yet remarkably robust.

Even when you manually introduce logical errors into a chain-of-thought trace, the model’s overall accuracy usually remains better than if it had produced no reasoning tokens at all. Unlike a strict forward- or backward-chaining proof system, the LLM’s reasoning relies on statistical association rather than brittle rule-following. In a way, that fuzziness is its strength because it generalizes instead of collapsing under contradiction.

Well put, and if it doesn't notice/collapse under introduced contradictions, that's evidence it's not the kind of reasoning we were hoping for. The "real thing" is actually brittle when you do it right.

Human reasoning is, in practice, much closer to statistical association than to brittle rule-following. The kind of strict, formal deduction we teach in logic courses is a special, slow mode we invoke mainly when we’re trying to check or communicate something, not the default way our minds actually operate.

Everyday reasoning is full of heuristics, analogies, and pattern matches: we jump to conclusions, then backfill justification afterward. Psychologists call this “post hoc rationalization,” and there’s plenty of evidence that people form beliefs first and then search for logical scaffolding to support them. In fact, that’s how we manage to think fluidly at all; the world is too noisy and underspecified for purely deductive inference to function outside of controlled systems.

Even mathematicians, our best examples of deliberate, formal thinkers, often work this way. Many major proofs have been discovered intuitively and later found to contain errors that didn’t actually invalidate the final result. The insight was right, even if the intermediate steps were shaky. When the details get repaired, the overall structure stands. That’s very much like an LLM producing a chain of reasoning tokens that might include small logical missteps yet still landing on the correct conclusion: the “thinking” process is not literal step-by-step deduction, but a guided traversal through a manifold of associations shaped by prior experience (or training data, in the model’s case).

So if an LLM doesn’t collapse under contradictions, that’s not necessarily a bug; it may reflect the same resilience we see in human reasoning. Our minds aren’t brittle theorem provers; they’re pattern-recognition engines that trade strict logical consistency for generalization and robustness. In that sense, the fuzziness is the strength.

> The kind of strict, formal deduction we teach in logic courses is a special, slow mode

Yes, but that seems like moving the goalposts.

The stricter blends of reasoning are what everybody is so desperate to evoke from LLMs, preferably along with inhuman consistency, endurance, and speed. Just imagine the repercussions if a slam-dunk paper came out tomorrow, which somehow proved the architectures and investments everyone is using for LLMs are a dead-end for that capability.

> The stricter blends of reasoning are what everybody is so desperate to evoke from LLMs

This is definitely not true for me. My prompts frequently contain instructions that aren't 100% perfectly clear, suggest what I want rather than formally specifying it, typos, mistakes, etc. The fact that the LLM usually figures out what I meant to say, like a human would, is a feature for me.

I don't want an LLM to act like an automated theorem prover. We already have those. Their strictness makes them extremely difficult to use, so their application is extremely limited.

I get the worry. AFAIK most of the current capex is going into scalable parallel compute, memory, and networking. That stack is pretty model agnostic, similar to how all that dot com fiber was not tied to one protocol. If transformers stall, the hardware is still useful for whatever comes next.

On reasoning, I see LLMs and classic algorithms as complements. LLMs do robust manifold following and associative inference. Traditional programs do brittle rule following with guarantees. The promising path looks like a synthesis where models use tools, call code, and drive search and planning methods such as MCTS, the way AlphaGo did. Think agentic systems that can read, write, execute, and verify.

LLMs are strongest where the problem is language. Language co evolved with cognition as a way to model the world, not just to chat. We already use languages to describe circuits, specify algorithms, and even generate other languages. That makes LLMs very handy for specification, coordination, and explanation.

LLMs can also statistically simulate algorithms, which is useful for having them think about these algorithms. But when you actually need the algorithm, it is most efficient to run the real thing in software or on purpose built hardware. Let the model write the code, compose the tools, and verify the output, rather than pretending to be a CPU.

To me the risk is not that LLMs are a dead end, but that people who do not understand them have unreasonable expectations. Real progress looks like building systems that use language to invent and implement better tools and route work to the right place. If a paper lands tomorrow that shows pure next token prediction is not enough for formal reasoning, that would be an example of misunderstanding LLMs, not a stop sign. We already saw something similar when Minsky and Papert highlighted that single layer perceptrons could not represent XOR, and the field later moved past that with multilayer networks. Hopefully we remember that and learn the right lesson this time.

The problem is that the overwhelming majority of input it has in-fact seen somewhere in the corpus it was trained on. Certainly not one for one but easily an 98% match. This is the whole point of what the other person is trying to comment on i think. The reality is most of language is regurgitating 99% to communicate an internal state in a very compressed form. That 1% tho maybe is the magic that makes us human. We create net new information unseen in the corpus.

> the overwhelming majority of input it has in-fact seen somewhere in the corpus it was trained on.

But it thinks just great on stuff it wasn't trained on.

I give it code I wrote that is not in its training data, using new concepts I've come up with in an academic paper I'm writing, and ask it to extend the code in a certain way in accordance with those concepts, and it does a great job.

This isn't regurgitation. Even if a lot of LLM usage is, the whole point is that it does fantastically with stuff that is brand new too. It's genuinely creating new, valuable stuff it's never seen before. Assembling it in ways that require thinking.

I think you may think too highly of academic papers or more so that they oft still only have 1% in there.

I think you're missing the point. This is my own paper and these are my own new concepts. It doesn't matter if the definition of the new concepts are only 1% of the paper, the point is they are the concepts I'm asking the LLM to use, and are not in its training data.

How would one prove the premise that a concept is not present in the training data?

With how much data is being shoveled in there, our default assumption should be that significant components are present.

That would be a weird default assumption. It's not hard to come up with new ideas. In fact, it's trivial.

And if you want to know if a specific concept is known by the LLM, you can literally ask it. It generally does a great job of telling you what it is and is not familiar with.

I think it would be hard to prove that it's truly so novel that nothing similar is present in the training data. I've certainly seen in research that it's quite easy to miss related work even with extensive searching.