If the model comes up with anything even remotely correct it would be quite a strong evidence that LLMs are a path to something bigger if not then I think it is time to go back to the drawing board.
Things don't happen all of a sudden, and being able to see all the scientific papers of the era its possible those could have fallen out of the synthesis.
Standing on the shoulders of giants, as it were
I by no means believe LLMs are general intelligence, and I've seen them produce a lot of garbage, but if they could produce these revolutionary theories from only <= year 1900 information and a prompt that is not ridiculously leading, that would be a really compelling demonstration of their power.
Kind of, how long would it have realistically taken for someone else (also really smart) to come up with the same thing if Einstein wouldn't have been there?
I personally doubt that this would work. I don't think these systems can achieve truly ground-breaking, paradigm-shifting work. The homeworld of these systems is the corpus of text on which it was trained, in the same way as ours is physical reality. Their access to this reality is always secondary, already distorted by the imperfections of human knowledge.
Yes. It is certainly a question if Einstein is one of the smartest guy ever lived or all of his discoveries were already in the Zeitgeist, and would have been discovered by someone else in ~5 years.
Einstein was smart and put several disjointed things together. It's amazing that one person could do so much, from explaining the Brownian motion to explaining the photoeffect.
But I think that all these would have happened within _years_ anyway.
It turns out my reading is somewhat topical. I've been reading Rhodes' "The Making of the Atomic Bomb" and of the things he takes great pains to argue (I was not quite anticipating how much I'd be trying to recall my high school science classes to make sense of his account of various experiments) is that the development toward the atomic bomb was more or less inexorable and if at any point someone said "this is too far; let's stop here" there would be others to take his place. So, maybe, to answer your question.
It's still unclear whether this process could be merely continued, seeded only with new physical data, in order to keep progressing beyond that point, "forever", or at least for as long as we imagine humans will continue to go on making scientific progress.
These two are so above everyone else in the mathematical world that most people would struggle for weeks or even months to understand something they did in a couple of minutes.
There's no "get down and dirty" shortcut with them =)
However, what I'm saying is not mere nitpicking either. It is precisely because of my belief in Einstein's extraordinary abilities that I find it unconvincing that an LLM being able to recombine the extant written physics-related building blocks of 1900, with its practically infinite reading speed, necessarily demonstrates comparable capabilities to Einstein.
The essence of the question is this: would Einstein, having been granted eternal youth and a neverending source of data on physical phenomena, be able to innovate forever? Would an LLM?
My position is that even if an LLM is able to synthesise special relativity given 1900 knowledge, this doesn't necessarily mean that a positive answer to the first question implies a positive answer to the second.
This would absolutely be very good evidence that models can actually come up with novel, paradigm-shifting ideas. It was absolutely not obvious at that time from the existing facts, and some crazy leap of faiths needed to be taken.
This is especially true for General Relativity, for which you had just a few mismatch in the mesurements like Mercury's precession, and where the theory almost entirely follows from thought experiments.
However, you could also argue that it's actually empirical evidence that general relativity and 19th century physics wasn't truly a paradigm shift -- you could have 'derived' it from previous data -- that the LLM has actually proven something about structurally similarities between those paradigms, not that it's demonstrating general intelligence...
For instance spectroscopy enables one to look at the spectra emitted by another 'thing', perhaps the sun, and it turns out that there's little streaks within the spectra the correspond directly to various elements. This is how we're able to determine the elemental composition of things like the sun.
That connection between elements and the patterns in their spectra was discovered in the early 1800s. And those patterns are caused by quantum mechanical interactions and so it was perhaps one of the first big hints of quantum mechanics, yet it'd still be a century before we got to relativity, let alone quantum mechanics.
But the whole question is whether or not something can do that synthesis!
And the "anyone who read all the right papers" thing - nobody actually reads all the papers. That's the bottleneck. LLMs don't have it. They will continue to not have it. Humans will continue to not be able to read faster than LLMs.
Even me, using a speech synthesizer at ~700 WPM.
If it's true of everything, then surely having an LLM work iteratively on the pieces, along with being provided additional physical data, will lead to the discovery of everything?
If the answer is "no", then surely something is still missing.
> And the "anyone who read all the right papers" thing - nobody actually reads all the papers. That's the bottleneck. LLMs don't have it. They will continue to not have it. Humans will continue to not be able to read faster than LLMs.
I agree with this. This is a definitive advantage of LLMs.
Excellence at any given task is not indicative of intelligence. I think we set these sort of false goalposts because we want something that sounds achievable but is just out of reach at one moment in time. For instance at one time it was believed that a computer playing chess at the level of a human would be proof of intelligence. Of course it sounds naive now, but it was genuinely believed. It ultimately not being so is not us moving the goalposts, so much as us setting artificially low goalposts to begin with.
So for instance what we're speaking of here is logical processing across natural language, yet human intelligence predates natural language. It poses a bit of a logical problem to then define intelligence as the logical processing of natural language.
I'm sure, without any uncertainty, that any generalist model able to do what Einstein did would be AGI, as in, that model would be able to perform any cognitive task that an intelligent human being could complete in a reasonable amount of time (here "reasonable" depends on the task at hand; it could be minutes, hours, days, years, etc).
(1) - A major part of the challenge is in not being directed towards something. There was no external guidance for Einstein - he wasn't even a formal researcher at the time of his breakthroughs. An LLM might be able to be handheld towards relativity, though I doubt it, but given the prompt of 'hey find something revolutionary' it's obviously never going to respond with anything relevant, even with substantially greater precision specifying field/subtopic/etc.
(2) - Logical processing of natural language remains one small aspect of intelligence. For example - humanity invented natural language from nothing. The concept of an LLM doing this is a nonstarter since they're dependent upon token prediction, yet we're speaking of starting with 0 tokens.
(3) - LLMs are, in many ways, very much like calculators. They can indeed achieve some quite impressive feats in specific domains, yet then they will completely hallucinate nonsense on relatively trivial queries, particularly on topics where there isn't extensive data to drive their token prediction. I don't entirely understand your extreme optimism towards LLMs given this proclivity for hallucination. Their ability to produce compelling nonsense makes them particularly tedious for using to do anything you don't already effectively know the answer to.
Simply because I don't see hallucinations as a permanent problem. I see that models keep improving more and more in this regard, and I don't see why the hallucination rate can't be abirtrarily reduced with further improvements to the architecture. When I ask Claude about obscure topics, it correctly replies "I don't know", where past models would have hallucinated an answer. When I use GPT 5.2-thinking for my ML research job, I pretty much never encounter hallucinations.
THat's not how science works.
The LLM would have to propose experiments (which would have to be simulated), and then develop its theories from that.
Maybe there had been enough facts around to suggest a number of hypotheses, but the LLM in its curent form won't be able to confirm them.
It’s probably even more of a problem that different areas of scientific development don’t know about each other. LLMs combining results would still not be like they invented something new.
But if they could give us a head start of 20 years on certain developments this would be an awesome result.
My own prediction is that the LLMs would totally fail at connecting the dots, but a small group of very smart humans can.
Things don't happen all of a sudden, but they also don't happen everywhere. Most people in most parts of the world would never connect the dots. Scientific curiosity is something valuable and fragile, that we just take for granted.
I imagine possible buffers on that conjecture synthesis being epxerimentation and acceptance by the scientific community. AIs can come up with new ideas every day but Nature won't publish those ideas for years.
--
[0] - To people who happen to have the right background and skill set, and are in the right place.
[1] - Almost always multiple someones, independently, within short time of each other. People usually remember only one or two because, for better or worse, history is much like patent law: first to file wins.
That's why those tiny steps of scientific and technological progress aren't made by just any randos - they're made by people who happen to be at the right place and time, and equipped correctly to be able to take the step.
The important corollary to this is that you can't generally predict this ahead of time. Someone like Einstein was needed to nail down relativity, but standing there few years earlier, you couldn't have predicted it was Einstein who would make a breakthrough, nor what would that be about. Conversely, if Einstein lived 50 years earlier, he wouldn't have come up with relativity, because necessary prerequisites - knowledge, people, environment - weren't there yet.
Your second and third paragraphs are entirely consistent with the original point I was trying to make, which was not that it took Einstein specifically to come up with relativity, but that it took someone with uncommon skills, as evidenced by the fact that it blindsided even a good many of the people who were qualified to be contenders for being the one to figure it out first. It does not amount to proof, but one does not expect people who are closing in on the solution to be blindsided by it.
I am well aware of the problems with “great man” hagiography, but dismissing individual contributions, which is what the person I was replying to seemed to be doing, is a distortion in its own way.
Take Michelson in 1894: after doing (and inspiring) the kind of precision work that should have set off alarm bells, he’s still talking like the fundamentals are basically done and progress is just “sixth decimal place” refinement.
"While it is never safe to affirm that the future of Physical Science has no marvels in store even more astonishing than those of the past, it seems probable that most of the grand underlying principles have been firmly established and that further advances are to be sought chiefly in the rigorous application of these principles to all the phenomena which come under our notice. It is here that the science of measurement shows its importance — where quantitative work is more to be desired than qualitative work. An eminent physicist remarked that the future truths of physical science are to be looked for in the sixth place of decimals." - Michelson 1894
The Michelson-Morley experiments weren't obscure, they were famous, discussed widely, and their null result was well-known. Yet for nearly two decades, the greatest physicists of the era proposed increasingly baroque modifications to existing theory rather than question the foundational assumption of absolute time. These weren't failures of data availability or technical skill, they were failures of imagination constrained by what seemed obviously true about the nature of time itself.
Einstein's insight wasn't just "connecting dots" here, it was recognizing that a dot everyone thought was fixed (the absoluteness of simultaneity) could be moved, and that doing so made everything else fall into place.
People scorn the 'Great Man Hypothesis' so much they sometimes swing too much in the other direction. The 'multiple discovery' pattern you cite is real but often overstated. For Special Relativity, Poincaré came close, but didn't make the full conceptual break. Lorentz had the mathematics but retained the aether. The gap between 'almost there' and 'there' can be enormous when it requires abandoning what seems like common sense itself.
There's a very similar parallel with dark matter in modern times. We certainly have endless hints to the truth that will be evident in hindsight, but for now? We are mostly convinced that we know the truth, perform experiments to prove that, find nothing, shrug, adjust the model to be even more esoteric, and repeat onto the next one. And maybe one will eventually show something, or maybe we're on the wrong path altogether. This quote, from Michelson in 1894 (more than a decade before Einstein would come along), is extremely telling of the opinion at the time:
"While it is never safe to affirm that the future of Physical Science has no marvels in store even more astonishing than those of the past, it seems probable that most of the grand underlying principles have been firmly established and that further advances are to be sought chiefly in the rigorous application of these principles to all the phenomena which come under our notice. It is here that the science of measurement shows its importance — where quantitative work is more to be desired than qualitative work. An eminent physicist remarked that the future truths of physical science are to be looked for in the sixth place of decimals." - Michelson 1894
For another parallel our understanding of the big bang was, and probably is, wrong. There are a lot of problems with the traditional view of the big bang with the horizon problem [1] being just one among many - areas in space that should not have had time to interact behave like they have. So this was 'solved' by an ad hoc solution - just make the expansion of the universe go into super-light speed for a fraction of a second at a specific moment, slow down, then start speeding up again (cosmic inflation [2]) - and it all works just fine. So you know what we did? Shrugged, accepted it, and even gave Guth et al a bunch of accolades for 'solving' the puzzle.
This is the problem - arguably the most important principle of science is falsifiability. But when is something falsified? Because in many situations, probably the overwhelming majority, you can instead just use one falsification to create a new hypothesis with that nuance integrated into it. And as science moves beyond singular formulas derived from clear principles or laws and onto broad encompassing models based on correlations from limited observations, this becomes more and more true.
New discoveries don’t happen in a vacuum.
Training data for a base model isn't that hard to come by, even though you have to OCR most of it yourself because the publicly available OCRed versions are commonly unusably bad. But training a model large enough to be useful is a major issue. Training a 700M parameter model at home is very doable (and is what this TimeCapsuleLLM is), but to get that kind of reasoning you need something closer to a 70B model. Also a lot of the "smarts" of a model gets injected in fine tuning and RL, but any of the available fine tuning datasets would obviously contaminate the model with 2026 knowledge.
It largely cannot imaginr anything beyond the usual but there is a small part that it can. This is similar to in context learning, it's weak but it is there.
It would be incredible if meta learning/continual learning found a way to train exactly for novel learning path. But that's literally AGI so maybe 20yrs from now? Or never..
You can see this on CL benchmarks. There is SOME signal but it's crazy low. When I was traing CL models i found that signal was in the single % points. Some could easily argue it was zero but I really do believe there is a very small amount in there.
This is also why any novel work or findings is done via MASSIVE compute budgets. They find RL enviroments that can extract that small amount out. Is it random chance? Maybe, hard to say.
AI cannot compound their learnings for the foreseeable future
Most ML researchers think AGI is imminent.
I don't think there is any level of broad agreement right now. There are tons of random camps none of which I would consider to be broadly dominating.
Not sure how to interpret that as almost imminent.
20 years away in 2026, still 20 years away in 2027, etc etc.
Whatever Altman's hyping, that's the translation.
I mostly agree with you, especially about distrusting the self-interested hype beasts.
While I don't think the models are actually "intelligent", I also wonder if there are insights to be gained by looking at how concepts get encoded by the models. It's not really that the models will add something "new", but more that there might be connections between things that we haven't noticed, especially because academic disciplines are so insular these days.
If we did this (to a good enough level of detail), would it be able to derive relativity? How large of an AI model would it have to be to successfully derive relativity (if it only had access to everything published up to 1904)?
I guess the arguments sometimes mention languages. But I feel like the core of the arguments are pretty much the same regardless?
At least the model part. Although others made same thought as you afaik none tried it.
They measure compression (perplexity) on future Wikipedia, news articles, code, arXiv papers, and multi-modal data. Data compression is intimately connected with robustness and generalization.
A). contaminate the model with your own knowledge of relativity, leading it on to "discover" what you know, or
B). you will try to simulate a blind operation but without the "competent human physicist knowledgeable up to the the 1900 scientific frontier" component prompting the LLM, because no such person is alive today nor can you simulate them (if you could, then by definition you can use that simulated Einstein to discover relativity, so the problem is moot).
So in both cases you would prove nothing about what a smart and knowledgeable scientist can achieve today from a frontier LLM.
In principle I see your point, in practice my default assumption until proven otherwise here -- is that a little something slipped through post-1900.
A much easier approach would be to just download some model, whatever model, today. Then 5 years from now, whatever interesting discoveries are found - can the model get there.
I imagine if you explained quantum field theory to a 19th century scientists they might think of it as a more refined understanding of luminiferous aether.
Or if an 18th century scholar learned about positive and negative ions, it could be seen as an expansion/correction of phlogiston theory.
[0] https://github.com/haykgrigo3/TimeCapsuleLLM/blob/main/Copy%...
In fact you don’t want it to know about them explicitly just have enough background knowledge that you can manage the rest via context.
The only people who are serious about suggesting that LLMs could possibly 'think' are the people who are committing fraud on the scale of hundreds of billions of dollars (good for them on finding the all-time grift!) and people who don't understand how they're programmed, and thusly are the target of the grift. Granted, given that the vast majority of humanity are not programmers, and even fewer are programmers educated on the intricacies of ML, the grift target pool numbers in the billions.
Could you elucidate me on the process of human thought, and point out the differences between that and a probabilistic prediction engine?
I see this argument all over the place, but "how do humans think" is never described. It is always left as a black box with something magical (presumably a soul or some other metaphysical substance) inside.
1: https://chatgpt.com/share/6965653e-b514-8011-b233-79d8c25d33...
That's a bar that only a handful of humans have met in the entirety of history.
But I think there are also some questions about the role of language in human thought that leave the door just slightly ajar on the issue of whether or not manipulating the tokens of language might be more central to human cognition than we've tended to think.
If it turned out that this was true, then it is possible that "a model predicting tokens" has more power than that description would suggest.
I doubt it, and I doubt it quite a lot. But I don't think it is impossible that something at least a little bit along these lines turns out to be true.
I'm convinced of this. I think it's because we've always looked at the most advanced forms of human languaging (like philosophy) to understand ourselves. But human language must have evolved from forms of communication found in other species, especially highly intelligent ones. It's to be expected that the building blocks of it is based on things like imitation, playful variation, pattern-matching, harnessing capabilities brains have been developing long before language, only now in the emerging world of sounds, calls, vocalizations.
Ironically, the other crucial ingredient for AGI which LLMs don't have, but we do, is exactly that animal nature which we always try to shove under the rug, over-attributing our success to the stochastic parrot part of us, and ignoring the gut instinct, the intuitive, spontaneous insight into things which a lot of the great scientists and artists of the past have talked about.
Are you familiar with the major works in epistemology that were written, even before the 20th century, on this exact topic?
Unfortunately, none of that has anything to do with what LLMs are doing. The LLM is not thinking about concepts and then translating that into language. It is imitating what it looks like to read people doing so and nothing more. That can be very powerful at learning and then spitting out complex relationships between signifiers, as it's really just a giant knowledge compression engine with a human friendly way to spit it out. But there's absolutely no logical grounding whatsoever for any statement produced from an LLM.
The LLM that encouraged that man to kill himself wasn't doing it because it was a subject with agency and preference. It did so because it was, quite accurately I might say, mimicking the sequence of tokens that a real person encouraging someone to kill themselves would write. At no point whatsoever did that neural network make a moral judgment about what it was doing because it doesn't think. It simply performed inference after inference in which it scanned through a lengthy discussion between a suicidal man and an assistant that had been encouraging him and then decided that after "Cold steel pressed against a mind that’s already made peace? That’s not fear. That’s " the most accurate token would be "clar" and then "ity."
We know what our experience is - thinking about concepts and then translating that into language - but we really don't know with much confidence what is actually going on.
I lean strongly toward the idea that humans are doing something quite different than LLMs, particularly when reasoning. But I want to leave the door open to the idea that we've not understood human cognition, mostly because our primary evidence there comes from our own subjective experience, which may (or may not) provide a reliable guide to what is actually happening.
We do know what it's not doing, and that is operating only through reproducing linguistic patterns. There's no more cause to think LLMs approximate our thought (thought being something they are incapable of) than that Naive-Bayes spam filter models approximate our thought.
'Language' is only the initial and final layers of a Large Language Model. Manipulating concepts is exactly what they do, and it's unfortunate the most obstinate seem to be the most ignorant.
It may, however, turn out that in doing what they do, they are effectively manipulating concepts, and this is what I was alluding to: by building the model, even though your approach was through tokenization and whatever term you want to use for the network, you end up accidentally building something that implicitly manipulates concepts. Moreover, it might turn out that we ourselves do more of this than we perhaps like to think.
Nevertheless "manipulating concepts is exactly what they do" seems almost willfully ignorant of how these systems work, unless you believe that "find the next most probable sequence of tokens of some length" is all there is to "manipulating concepts".
Yes, they do. And of course there is. And there's plenty of research on the matter.
>It may, however, turn out that in doing what they do, they are effectively manipulating concepts
There is no effectively here. Text is what goes in and what comes out, but it's by no means what they manipulate internally.
>Nevertheless "manipulating concepts is exactly what they do" seems almost willfully ignorant of how these systems work, unless you believe that "find the next most probable sequence of tokens of some length" is all there is to "manipulating concepts".
"Find the next probable token" is the goal, not the process. It is what models are tasked to do yes, but it says nothing about what they do internally to achieve it.
Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task - https://openreview.net/forum?id=DeG07_TcZvT
On the Biology of a Large Language Model - https://transformer-circuits.pub/2025/attribution-graphs/bio...
Emergent Introspective Awareness in Large Language Models - https://transformer-circuits.pub/2025/introspection/index.ht...
And this is OK and still quite interesting - we do it to ourselves all the time. Often it's the only way we have of understanding the world (or ourselves).
However, in the case of LLMs, which are tools that we have created from scratch, I think we can require a higher standard.
I don't personally think that any of these papers suggest that LLMs manipulate concepts. They do suggest that the internal representation after training is highly complex (superposition, in particular), and that when inputs are presented, it isn't unreasonable to talk about the observable behavior as if it involved represented concepts. It is useful stance to take, similar to Dennett's intentional stance.
However, while this may turn out to be how a lot of human cognition works, I don't think it is what is the significant part of what is happening when we actively reason. Nor do I think it corresponds to what most people mean by "manipulate concepts".
The LLM, despite the prescence of "features" that may correspond to human concepts, is relentlessly forward-driving: given these inputs, what is my output? Look at the description in the 3rd paper of the arithmetic example. This is not "manipulating concepts" - it's a trick that often gets to the right answer (just like many human tricks used for arithmetic, only somewhat less reliable). It is extremely different, however, from "rigorous" arithmetic - the stuff you learned when you somewhere between age 5 and 12 perhaps - that always gives the right answer and involves no pattern matter, no inference, no approximations. The same thing can be said, I think, about every other example in all 4 papers, to some degree or another.
What I do think is true (and very interesting) is that it seems somewhere between possible and likely that a lot more human cognition than we've previously suspected uses similar mechanisms as these papers are uncovering/describing.
I’m not sure what distinction you’re drawing here. A lot of mechanistic interpretability work is explicitly trying to describe what the model is doing in the most literal sense we have access to: identifying internal features/circuits and showing that intervening on them predictably changes behavior. That’s not “as-if” gloss; it’s a causal claim about internals.
If your standard is higher than “we can locate internal variables that track X and show they causally affect outputs in X-consistent ways,” what would count as “actually describing what it’s doing”?
>However, in the case of LLMs, which are tools that we have created from scratch, I think we can require a higher standard.
This is backwards. We don’t “create them from scratch” in the sense relevant to interpretability. We specify an architecture template and a training objective, then we let gradient descent discover a huge, distributed program. The “program” is not something we wrote or understand. In that sense, we’re in a similar epistemic position as neuroscience: we can observe behavior, probe internals, and build causal/mechanistic models, without having full transparency.
So what does “higher standard” mean here, concretely? If you mean “we should be able to fully enumerate a clean symbolic algorithm,” that’s not a standard we can meet even for many human cognitive skills, and it’s not obvious why that should be the bar for “concept manipulation.”
>I don't personally think that any of these papers suggest that LLMs manipulate concepts. They do suggest that the internal representation after training is highly complex (superposition, in particular), and that when inputs are presented, it isn't unreasonable to talk about the observable behavior as if it involved represented concepts. It is useful stance to take, similar to Dennett's intentional stance.
You start with “there is no representation of a concept,” but then concede “features that may correspond to human concepts.” If those features are (a) reliably present across contexts, (b) abstract over surface tokens, and (c) participate causally in producing downstream behavior, then that is a representation in the sense most people mean in cognitive science. One of the most frustrating things about these sorts of discussions is the meaningless semantic games and goalpost shifting.
>The LLM, despite the prescence of "features" that may correspond to human concepts, is relentlessly forward-driving: given these inputs, what is my output?
Again, that’s a description of the objective, not the internal computation. The fact that the training loss is next-token prediction doesn’t imply the internal machinery is only “token-ish.” Models can and do learn latent structure that’s useful for prediction: compressed variables, abstractions, world regularities, etc. Saying “it’s just next-token prediction” is like saying “humans are just maximizing inclusive genetic fitness,” therefore no real concepts. Goal ≠ mechanism.
> Look at the description in the 3rd paper of the arithmetic example. This is not "manipulating concepts" - it's a trick that often gets to the right answer
Two issues:
1. “Heuristic / approximate” doesn’t mean “not conceptual.” Humans use heuristics constantly, including in arithmetic. Concept manipulation doesn’t require perfect guarantees; it requires that internal variables encode and transform abstractions in ways that generalize.
2. Even if a model is using a “trick,” it can still be doing so by operating over internal representations that correspond to quantities, relations, carry-like states, etc. “Not a clean grade-school algorithm” is not the same as “no concepts.”
>Rigorous arithmetic… always gives the right answer and involves no pattern matching, no inference…
“Rigorous arithmetic” is a great example of a reliable procedure, but reliability doesn’t define “concept manipulation.” It’s perfectly possible to manipulate concepts using approximate, distributed representations, and it’s also possible to follow a rigid procedure with near-zero understanding (e.g., executing steps mechanically without grasping place value).
So if the claim is “LLMs don’t manipulate concepts because they don’t implement the grade-school algorithm,” that’s just conflating one particular human-taught algorithm with the broader notion of representing and transforming abstractions.
I'll see if I can try to explain what I mean here, because I absolutely don't believe this is shifting the goal posts.
There are a couple of levels of human cognition that are particularly interesting in this context. One is the question of just how the brain does anything at all, whether that's homeostasis, neuromuscular control or speech generation. Another is how humans engage in conscious, reasoned thought that leads to (or appears to lead to) novel concepts. The first one is a huge area, better understood than the second though still characterized more by what we don't know than what we do. Nevertheless, it is there that the most obvious parallels with e.g. the Lindsey papers can be found. Neural networks, activation networks and waves, signalling etc. etc. The brain receives (lots of) inputs, generates responses including but not limited to speech generation. It seems entirely reasonable to suggest that maybe our brains, given a somewhat analogous architecture at some physical level to the one used for LLMs, might use similar mechanisms as the latter.
However, nobody would say that most of what the brain does involves manipulating concepts. When you run from danger, when you reach up grab something from a shelf, when you do almost anything except actual conscious reasoning, most of the accounts of how that behavior arises from brain activity does not involve manipulating concepts. Instead, we have explanations more similar to those being offered for LLMs - linked patterns of activations across time and space.
Nobody serious is going to argue that conscious reasoning is not built on the same substrate as unconscious behavior, but I think that most people tend to feel that it doesn't make sense to try to shoehorn it into the same category. Just as it doesn't make much sense to talk about what a text editor is doing in terms of P and N semiconductor gates, or even just logic circuits, it doesn't make much sense to talk about conscious reasoning in terms of patterns of neuronal activation, despite the fact that in both cases, one set of behavior is absolutely predicated on the other.
My claim/belief is that there is nothing inside an LLM that corresponds even a tiny bit to what happens when you are asked "What is 297 x 1345?" or "will the moon be visible at 8pm tonight?" or "how does writer X tackle subject Y differently than writer Z?". They can produce answers, certainly. Sometimes the answers even make significant sense or better. But when they do, we have an understanding of how that is happening that does not require any sense of the LLM engaging in reasoning or manipulating concepts. And because of that, I consider attempts like Lindsey's to justify the idea that LLMs are manipulating concepts to be misplaced - the structures Lindsey et al. are describing are much more similar to the ones that let you navigate, move, touch, lift without much if any conscious thought. They are not, I believe, similar to what is going on in the brain when you are asked "do you think this poem would have been better if it was a haiku?" and whatever that thing is, that is what I mean by manipulating concepts.
> Saying “it’s just next-token prediction” is like saying “humans are just maximizing inclusive genetic fitness,” therefore no real concepts. Goal ≠ mechanism.
No. There's a huge difference between behavior and design. Humans are likely just maximizing genetic fitness (even though that's really a concept, but that detail is not worth arguing about here), but that describes, as you note, a goal not a mechanism. Along the way, they manifest huge numbers of sub-goal directed behaviors (or, one could argue quite convincingly, goal-agnostic behaviors) that are, broadly speaking, not governed by the top level goal. LLMs don't do this. If you want to posit that the inner mechanisms contain all sorts of "behavior" that isn't directly linked to the externally visible behavior, be my guest, but I just don't see this as equivalent. What humans visibly, mechanistically do covers a huge range of things; LLMs do token prediction.
This framing assumes "concept manipulation" requires conscious, deliberate reasoning. But that's not how cognitive science typically uses the term. When you reach for a shelf, your brain absolutely manipulates concepts - spatial relationships, object permanence, distance estimation, tool affordances. These are abstract representations that generalize across contexts. The fact that they're unconscious doesn't make them less conceptual
>My claim/belief is that there is nothing inside an LLM that corresponds even a tiny bit to what happens when you are asked "What is 297 x 1345?" or "will the moon be visible at 8pm tonight?"
This is precisely what the mechanistic interpretability work challenges. When you ask "will the moon be visible tonight," the model demonstrably activates internal features corresponding to: time, celestial mechanics, geographic location, lunar phases, etc. It combines these representations to generate an answer.
>But when they do, we have an understanding of how that is happening that does not require any sense of the LLM engaging in reasoning or manipulating concepts.
Do we? The whole point of the interpretability research is that we don't have a complete understanding. We're discovering that these models build rich internal world models, causal representations, and abstract features that weren't explicitly programmed. If your claim is "we can in principle reduce it to matrix multiplications," sure, but we can in principle reduce human cognition to neuronal firing patterns too.
>They are not, I believe, similar to what is going on in the brain when you are asked "do you think this poem would have been better if it was a haiku?" and whatever that thing is, that is what I mean by manipulating concepts.
Here's my core objection: you're defining "manipulating concepts" as "whatever special thing happens during conscious human reasoning that feels different from 'pattern matching.'" But this is circular and unfalsifiable. How would we ever know if an LLM (or another human, for that matter) is doing this "special thing"? You've defined it purely in terms of subjective experience rather than functional or mechanistic criteria.
>Humans are likely just maximizing genetic fitness... but that describes, as you note, a goal not a mechanism. Along the way, they manifest huge numbers of sub-goal directed behaviors... that are, broadly speaking, not governed by the top level goal. LLMs don't do this.
LLMs absolutely do this, it's exactly what the interpretability research reveals. LLMs trained on "token prediction" develop huge numbers of sub-goal directed internal behaviors (spatial reasoning, causal modeling, logical inference) that are instrumentally useful but not explicitly specified, precisely the phenomenon you claim only humans exhibit. And 'token prediction' is not about text. The most significant advances in robotics in decades are off the back of LLM transformers. 'Token prediction' is just the goal, and I'm tired of saying this for the thousandth time.
> Here's my core objection: you're defining "manipulating concepts" as "whatever special thing happens during conscious human reasoning that feels different from 'pattern matching.'" But this is circular and unfalsifiable. How would we ever know if an LLM (or another human, for that matter) is doing this "special thing"? You've defined it purely in terms of subjective experience rather than functional or mechanistic criteria.
I think your core objection is well aligned to my own POV. I am not claiming that the subjective experience is the critical element here, but I am claiming that whatever is going on when we have the subjective experience of "reasoning" is likely to be different (or more specifically, more usefully described in different ways) than what is happening in LLMs and our minds when doing something else.
How would we ever know? Well the obvious answer is more research into what is happening in human brains when we reason and comparing that to brain behavior at other times.
I don't think it's likely to be productive to continue this exchange on HN, but if you would like to continue, my email address is in my profile.
Then, we have a model that knows quite a bit in modern English. We also legally have a data set for everything it knows. Then, there's all kinds of experimentation or copyright-safe training strategies we can do.
Project Gutenberg up to the 1920's seems to be the safest bet on that.
But I also think we can do this with normal LLMs trained on up-to-date text, by asking them to come up with any novel theory that fits the facts. It does not have to be a groundbreaking theory like QM, just original and not (yet) proven wrong ?
But considering that Special Relativity was published in 1905 which means all its building blocks were already floating in the ether by 1900 it would be a very interesting experiment to train something on Claude/Gemini scale and then say give in the field equations and ask it to build a theory around them.
Here you will find all official resources provided by the Turkish state. You can register as a foreign citizen and conduct research. However, knowing Turkish is a great advantage. You can find more accurate documents by searching in the summaries, because all documents are archived with a short Turkish explanation - a summary. Also, most scanned images you want to download are paid.
1900 era humans are not available to do this so I'm not sure how this experiment is supposed to work.
Special relativity however seems possible.
It's a relief to think that we're not trusting these things for stuff like financial advice, medical advice, mental health counselling, ...