Position: LLMs Can't Jump
openreview.net
openreview.net
So, I do sort of buy into this idea that Einstein was simulating the world and running experiments on those simulations in ways that were beyond what you could encode in natural language. Will AI be capable of doing this, if it is bounded by training data that is composed almost entirely on language? One might argue that if AI is training on a lossy encoding/representation of the human experience, how will it be able to simulate anything beyond that experience? Unless it does so in a way that we manage to do when we image objects beyond 3D. But now I'm just rambling.
If you read every single book there is about The Grand Canyon, and watched every single video and/or documentary about The Grand Canyon, do you believe that you have fully experienced The Grand Canyon? Or do you just have to be there to fully experience it.
I dunno. Substitute in whatever you want for "The Grand Canyon". Maybe climbing Mount Everest or walking on the Moon. The point is that maybe the human experience is more vast than what is written about it.
Worth... quite a lot imo.
First, in the slightly objective sense of "What is this place like in real life, under X conditions." But, more subjectively, watching a video of a glacier and imagining the wind/rain/cold doesn't even approach 5% of the intensity and awe of climbing up a mountain yourself to see a seemingly endless expanse of ice, struggling to stand steady because of the wind, shivering because of the cold and rain. And finally, in the financial sense, a lot of people routinely spend many thousands of dollars and days-weeks of their time to experience natural wonders in real life.
It's not that the value of the grand canyon is low in an absolute sense, it's that in a relative sense the gap between "empty numb void" and "hiking at home (plus grand canyon videos)" is far far greater than the gap between "hiking at home (plus grand canyon videos)" and "actual grand canyon".
Have you ever been to a place which is really high up? You're above the trees and can not only see as far as the curvature of the earth allows you to, but are high enough that the curvature of the earth allows you to see for miles. The air is thinner so it's easier to take a breath yet harder to catch it. Your sweat evaporates faster but a breeze doesn't cool you as much. The sun is brighter, rises earlier and sets later. The wildlife makes unfamiliar sounds and sound itself is changed. Even the same food tastes different.
It's not something you can get by sitting at home and changing the background on your computer.
Back to the topic, I'm sure it's very impressive, but I think you're underestimating just how impressive our baseline is.
Sure, no doubt. But I don't really think "video of my local state park + imagining the parts that aren't conveyed over video" is particularly close to the real thing either.
Also, you touched on how even modern tech still falls short of the true experience. Just look at the history of motion picture since the early 1900s. We have added sound, color, bigger screens, faster refresh rates, more pixels, 3D (sort of), surround sound, spatial audio, IMAX, etc. Almost seems like video leaves a lot to be desired.
It's similar in dismissing artificial audio based on listening to some pop song on a basic mobile speaker versus listening to well-designed binaural audio on good headphones. The latter can sound 'freakily' 'real'.
Rather one should ask "Would you be able to answer any question and fulfill any request about the Grand Canyon given to you by others in a way that will be indistinguishable from someone who has been to the Grand Canyon"
I'd be fantastically surprised if you could provide a precise answer to any such question without extensive security footage that doesn't exist.
An estimate, perhaps, but you probably couldn't even answer a simpler "how many people are currently in the canyon" without resorting to estimation given the lack of information currently available via the mentioned sources.
You might equally contrast someone who is physically in the Grand Canyon with an LLM given access to a drone with sensory attachments. I'd bet that if you told groups of humans and drone-controlling AIs to find something interesting, the humans would be more variable and cover a wider range of interesting discoveries but often just give up on the task, while the AI would be more likely to find something but would cluster around certain discoveries and make fewer overall.
But in terms of objective knowledge about the Grand Canyon, reading every article and scholarly work would give you a much better grounding. Unless you're a geologist/ecologist doing actual field research there, actually seeing the Grand Canyon would add little, if anything, to your factual knowledge and understanding of the site.
If I'm going to hire a guide, I'm going to hire the guy who spent a week hiking it instead of the guy who spent a week reading about it or watching videos about it.
With apologies if this is common knowledge at this point, 3Blue1Brown has been doing an excellent series on compression, and its relationship to intelligence (or more controversially, that they are one and the same): https://www.youtube.com/watch?v=l6DKRf-fAAM
But that also throws in sharp relief, that there is vastly more to the human experience than intelligence alone: qualia, desire, gut instinct, intuition, emergent creativity. (Whether the "God of the Gaps" for the delta between capabilities of human vs AIs is fixed, or diminishing, or even shrinking to zero, remains an open experiment we're all living through.)
The compression series sounds interesting!
I think the deltas are growing at different rates.
The 3b1b videos mention how this concept may be used to find similarities between different languages. Researchers have been able to get results that closely resemble how languages are usually grouped into families.
Math is pure reasoning and stock LLM is already better at it than 99.9% of humans.
And as for human language, agentic LLM can produce a perfectly human text. The fact that AI texts have tells is just because default settings are the same for millions of people using given LLM and almost everybody just pretty much on-shots the text instead of doing it agentically with anti-slop detection and rephrasing in the loop.
given that, you can very mechanically convert it to a statistical compressor by just attaching an entropy coder. This gets very good compression.
Yes, and sometimes this is very intentional. Take for example a short poem which if you sit and really think about it for a long time, you could go off on a mental tangent of imagining what sort of kingdom or empire created a statue that is now "two vast and trunkless legs of stone", for instance. Being terse and allowing for human interpretation is kind of the entire point of something being written like this.
I met a traveller from an antique land
Who said: Two vast and trunkless legs of stone
Stand in the desert. Near them, on the sand,
Half sunk, a shattered visage lies, whose frown,
And wrinkled lip, and sneer of cold command,
Tell that its sculptor well those passions read
Which yet survive, stamped on these lifeless things,
The hand that mocked them and the heart that fed:
And on the pedestal these words appear:
"My name is Ozymandias, king of kings:
Look on my works, ye Mighty, and despair!"
Nothing beside remains. Round the decay
Of that colossal wreck, boundless and bare
The lone and level sands stretch far away.
It's also to a certain extent why LLMs work.
When IBM Watson was playing Jeopardy, one of the game prompts was:
> It was the anatomical oddity of U.S. gymnast George Eyser, who won a gold medal on the parallel bars in 1904
The man was missing a leg and used a prosthetic. Watson's output was, "What is leg?"
At first it was regarded as correct. If a human said that you could conclude that they knew the answer. But then the judges decided not to give Watson the point because its output didn't provide enough specificity to prove that it understood the context.
If you ask an LLM what kinds of things taste sweet it can give you examples like cotton candy or strawberries, but it has never actually tasted anything. All it knows is that the training data contains the association between those tokens. But the human reading the output knows what strawberries are, which is what allows the output to be meaningful.
> The fundamental problem of communication is that of reproducing at one point either exactly or approximately a message selected at another point. Frequently the messages have meaning; that is they refer to or are correlated according to some system with certain physical or conceptual entities. These semantic aspects of communication are irrelevant to the engineering problem.
—Albert Einstein
Quoted in Using Spaced Repetition Systems to See Through a Piece of Mathematics,
https://news.ycombinator.com/item?id=18895613
which describes the author's experience that if you approach a field obsessively enough, eventually you begin to understand it at a level deeper than language.
If an LLM is big enough, I imagine something similar is happening.
Then again maybe the philosophy department has a different opinion :)
https://www.noahpinion.blog/p/what-will-more-intelligence-ac...
> Another way of saying this is that there may be laws of the universe that humans can’t understand but AI can. I call these “cloud laws” — causal regularities that can be exploited by technology, but which are too diffuse and complex for an individual human being to either intuit or communicate. Human language seems to obey cloud laws, so why not other phenomena too? Perhaps social sciences like economics, sociology, and political science obey similarly complex regularities, and AI can help us find them. Perhaps there are physical processes — plasma, or topological materials, or aerial turbulence, etc. — that obey cloud laws instead of chaos?
"A post-graduate student equipped with honours and diplomas went to Agassiz to receive the final and finishing touches. The great man offered him a small fish and told him to describe it. Post-Graduate Student: “That’s only a sun-fish” Agassiz: “I know that. Write a description of it.” After a few minutes the student returned with the description of the Ichthus Heliodiplodokus, or whatever term is used to conceal the common sunfish from vulgar knowledge, family of Heliichterinkus, etc., as found in textbooks of the subject. Agassiz again told the student to describe the fish. The student produced a four-page essay. Agassiz then told him to look at the fish. At the end of the three weeks the fish was in an advanced state of decomposition, but the student knew something about it."
sourced from: https://nabeelqu.co/understanding
Building AGI Using Language Models – https://bmk.sh/2020/08/17/Building-AGI-Using-Language-Models...
But ultimately this is just a matter of training data. I do not say that it is easy to obtain the required data, but it is not a fundamental problem LLMs can't overcome.
> From the two postulates, Einstein derived the Lorentz trans- formation ...
If Einstein derived them, who is "Lorentz"?
The groundwork for Special Relativity was the study of electrodynamics and symmetries of Maxwell equations. The Einsteins paper was literally called "On the Electrodynamics of Moving Bodies" and never cites Michelson and Morley.
> If Einstein derived them, who is "Lorentz"?
You can (re-) derive a lot of existing stuff.
Einstein was aware of Lorentz and the transform. He was aware of Poincaré as well. He knew the state of the art for his time.
Imho there's also a clean argument against the existence of LLM-understanding: chatbots have been unable to summarize to experts (see my reply to you in the other thread) their own findings.
Even after prolonged interrogation. They were unable to _compress_ their own findings. Thus they might not actually understand what they have actually done. (They might barely pass an oral thesis defense)
(One may object-- that proofs aren't data that can be "compressed". But then doesn't the process of abduction generalise the very idea of data? to.. ?)
Note that there are some differences between the different breakthroughs. The ones that humans were close to solving already were better explained imho
Einstein is not the first great scientist who are in denial of other important prior contributions, and he also not the last one. Newton also probably knew too well about Al-Haytham (Alhazen), arguably the father of modern science, and his breakthrough experiments but never directly cited Alhazen's works in his seminal books on Optics.
[1] Millikan, Einstein, and the Birth of Relativity (4 letters):
https://www.aps.org/archives/publications/apsnews/200403/let...
"Abraham Pais, who knew Einstein well and wrote his scientific biography, was certain that Einstein did know about Michelson's experiment before 1905. He points out that Einstein was over seventy and in poor health when he spoke to Shankland; at the first interview he probably did not remember that Michelson's experiment is discussed in Lorentz's 1895 monograph, the famous "Versuch", which Einstein had definitely read before 1905."
Furthermore I can't help but feel some logic would lead us to conclude it is not natural to cite, why? Because in our time if you do not cite your sources in a lot of occupations dealing with knowledge you will be punished for it. It would be absurd to construct punishments for failing to do what people will naturally do, and to have those punishments so often called on.
Even if we give a very lenient interpretation of not citing sources we would probably have around 10% of papers with errors, and that is with the fear of punishment.
Citing when one should and accurately is not normal, it requires discipline to do it, or there would not be such a high failure rate at doing it.
So it seems you have established that no, Newton was not expected to cite everyone in his time, very well, this now puts the question further up the chain over my entry into the conversation - "Were the people Newton neglected to cite actually integral to his work?"
My definition of integral would be, two possible definitions:
1. do I write something here that someone reading it would say, "What?! What do you mean, explain yourself, how did you arrive at this surprising bit of information?!!" then it is integral.
2. Am I doing this based on some work that someone did without which I would definitely not be doing this? If I am doing it with knowledge of the prior work but whether or not the prior work existed I would still be doing it, the prior work is not actually integral.
Newton was accusing Leibniz plagiarising his work on calculus and engaged in a bitter, lifelong priority dispute over the invention of calculus.
“[Newton] was not the first of the age of reason. He was the last of the magicians.”The size of an atom was computed for the first time 40 years earlier, in 1865, by Johann Josef Loschmidt, who was an Austrian scientist.
Loschmidt had determined the value of what today is called "Avogadro's number", despite the fact that Avogadro had no idea about how one could find the value at that number. The contribution of Avogadro had been the law that ideal gases have the same number of molecules per volume when the pressure and the temperature are the same, but he did not know how that number might be determined.
Based on the results of Loschmidt, which were also used by Maxwell in his works on statistical physics, a few years later, in 1874, George Johnstone Stoney computed the value of the elementary electric charge, which he later named "electron".
Loschmidt was the one who first measured and weighed the atoms, establishing their existence beyond any reasonable doubt. Before him, many believed that the hypothesis of Dalton about the existence of atoms is just a convenient theory for explaining the laws of chemistry, but some other better explanation might be discovered in the future.
Einstein published an explanation of the Brownian motion, which verified that everything fits in the known theories and it matches the expected values. It did not contain any novel number or law, but it was a useful confirmation of the current theories.
In my opinion, the greatest achievement of Einstein has been the paper published by him in which he revealed the existence of the stimulated emission of electromagnetic waves by hot bodies, besides their well known absorption and spontaneous emission.
This paper had tremendous practical applications, leading to the discovery of masers and lasers, without which the modern technology would have never existed.
While Einstein's theory of gravitation was more intellectually challenging, its practical applications are minor, almost negligible. Moreover, even after more than a century we do not know exactly how adequate it is for describing gravitation.
What is certain is that Einstein's theory of gravitation cannot be anything else but an approximation, as it is based on averaging the distribution of matter in space, without taking into account the concentration of matter into elementary particles.
Another contribution of Einstein that is more important than his most frequently cited works is his support for Satyendra Nath Bose, which lead to the recognition of the Bose-Einstein statistics, which together with the Fermi-Dirac statistics plays an essential role in quantum physics.
What I have written is not a dismissal of Einstein. On the contrary, I have said that Einstein is the source of a few extremely important discoveries, but those are not the theories cited by most people, who have never read the original work of Einstein but speak about him from hearsay.
As I have written above, when speaking about Einstein, almost nobody mentions the discovery of stimulated emission, which has been far more important than anything else done by Einstein. Einstein did not discover special relativity, the photoelectric effect or Brownian motion. He just provided alternative explanations for them, which did not include any new formula or quantity.
On the other hand, stimulated emission was a surprise, even if it is one of those things that appears to be obvious in retrospect, but for decades nobody who studied the blackbody radiation had noticed that something is missing in the equations.
This discovery eventually lead to the invention of masers and lasers and today there exists almost nothing that incorporates modern technology and which could exist without lasers.
Already for decades, no modern integrated circuits, no CPU and no memory can be fabricated without using lasers. Even many purely mechanical things, but which require great precision, cannot be fabricated without using measurement instruments that depend on lasers.
So the contributions of Einstein are very important, but they are not those that most people associate with him. Besides stimulated emission his most original work was his theory of gravitation, a.k.a. general relativity, but if he had not developed it that might have changed nothing, because Hilbert had developed an equivalent theory almost simultaneously, so had Einstein not published first, Hilbert would have become its discoverer.
After stimulated emission, the most important in practice research of Einstein is the Bose-Einstein statistics. But again, here like for special relativity, the statistics had been discovered by someone else, i.e. by Bose. Einstein has nonetheless the great merit of understanding its value. He demonstrated that the statistics was much more generally applicable than initially believed by Bose and his authority convinced everybody about its importance.
Thus there is no doubt that Einstein is one of the most important physicists of the 20th century. Nevertheless, there are dozens who made similarly important discoveries, e.g. Sommerfeld, Dirac, de Broglie, Schroedinger, Fermi, Rutherford and many others.
fair, there's lots of really awesome scientists, many of which not talked about even 1/100th as often as einstein
>when he doesn’t even belong in that conversation.
suddenly, the pendulum has swung way too far in the other direction.
But I'd like more explanation as to why Maxwell or Newton are any less replaceable: all 4 of "Maxwell's equations" have other names attached to them (the exception is Ampere-Maxwell law [1] which Maxwell contributed an important term to), and Newton had a number of contemporaries who were making similar discoveries but are often forgotten.
[1]: https://en.wikipedia.org/wiki/Amp%C3%A8re's_circuital_law
What a depressingly peculiar time to be alive.
Reddit is over that way, my friend. You might find the crowd there more amenable to this nonsense.
Einstein’s fame isn’t the problem. It’s the narrative that he is somehow the greatest scientist ever, when it’s just transparently false when you look at his actual contributions and the historical momentum of the fields he contributed to. There’s nothing wrong with his fame, the problem is that it overshadows actual juggernauts, like Maxwell and Newton. Maxwell not being famous at all among the ordinary public is a great tragedy, when he genuinely is in the conversation as having have been the most important physicist to have ever lived.
Einstein is indeed one of the most important physicists of the 20th century, but it is really meaningless to try to establish a hierarchy between physicists, because their works are like open-source programs, each work is based on the works of the predecessors and it would have been impossible to do without those previous works.
In any case, Einstein was good, but Maxwell seems really superior in the sense that he could extract the essential information from the much more confuse sources that were available at his time.
The Treatise of Electricity of Magnetism of Maxwell and his collected papers are very instructive lectures even now and there are fragments of them that remain better than the equivalent fragments in many modern textbooks, despite the huge additional information that we have today.
Most of the original texts written by Einstein are now of interest only for understanding the details of how his thinking and that of the contemporaneous physicists have evolved during those important years, but those texts have little other importance except for history, because later better expositions of the same material have appeared. Nonetheless, it is important to also read Einstein's originals, because in many later books there are some inappropriate interpretations attributed to Einstein, but in reality Einstein had said something else and which was more correct than what some later author misunderstood.
Wasn't this the case with Einstein's relativity as well? From what I understand the maths for it was already developed for many years, but Einstein was the first to tie it all together into a cohesive theory. Ideas like there being no absolute time, speed of light being constant in all frames, seem like a pretty big conceptual leap. I don't know how scientists at the time reacted to it, but defying Newtonian laws at a time when it was considered absolutely sacrosanct, must have been a big deal.
We see the same with general relativity, with ideas like saying gravity and acceleration is the same thing. I think what really impresses people is the radical nature of these ideas, which completely upheaval long-standing, prevailing notions. To do something like that once would be pretty remarkable by itself. But to do it twice, and in both cases, resulting in a new framework that will become the basis for nearly everything else, surely must put them in the conversation for the greatest at the very least.
That's a rather wild claim, given that the article says
> The survey was conducted among 100 of today's leading physicists.
Science rightfully recognizes the mind that doesn't just first describe the idea but provides a robust framework to test it and communicate it.
> A few reflections on my "LLMs Can’t Jump" paper:
> My position paper recently got some traction here, so I wanted to share a few thoughts and clarify a few things.
> First things first: some people are framing this as "DeepMind is throwing cold water on AI for science" or claiming the paper argues LLMs can never make real scientific discoveries. This is NOT the case.
> This is a personal position paper, not the company's view on AI for science. This is also not my position. As a core contributor to AlphaProof (the first AI system to win an IMO medal), I know firsthand that my colleagues at DeepMind, other frontier labs, and academia have made amazing discoveries with LLMs and will continue to do so. This paper is NOT an "LLMs are a dead end" kind of thing.
> Rather, the paper is the result of a deep dive I took to study the invention of General Relativity. I wanted to explore what it would take for a modern AI system to make that exact kind of jump. Specifically, I focused on the equivalence principle—a key axiom that Einstein formulated through thought experiments grounded in his physical intuition. I was trying to figure out what it would take to give modern AI systems that sort of thinking.
> Giving AI this specific capability isn't necessarily the most urgent thing to do next. It is very likely that improving our current recipes will lead to many exciting discoveries in the near future. In fact, that is what I am personally working on these days (sorry to disappoint you!). It is also quite possible that I am wrong, and that simply scaling our current systems will lead to new inventions in physics and elsewhere.
> Nevertheless, this was my position last winter when I wrote the paper, and I'm sticking to it. I think that there are a few interesting ideas to explore in this space which could influence the next generation of AI systems. I was very lucky to receive a lot of interesting feedback about this position—thank you for all the messages!
1. Read the last sentence of the abstract, and
2. Reflect that frontier reasoning agents already increasingly integrate multimodal models.
I simply point out here, that fully accepting the paper’s premise, the paper’s conclusion isn’t limiting on frontier AI reasoning agents. The paper posits the necessity of multimodal world models and the limitations of LLMs. Frontier agents aren’t simply LLMs and do increasingly integrate increasingly capable multimodal models.
> Be kind. Don't be snarky. Converse curiously; don't cross-examine. Edit out swipes.
> Comments should get more thoughtful and substantive, not less, as a topic gets more divisive.
> When disagreeing, please reply to the argument instead of calling names. "That is idiotic; 1 + 1 is 2, not 3" can be shortened to "1 + 1 is 2, not 3."
It's weird because the equivalence principle is very unintuitive. Aristotle's Mechanics does not have it. It took almost two thousand years to discover inertia that is the most simple version of the equivalence principle. Einstein understood the idea of the the equivalence principle because he had a physics degree, not because he feel that in real life.
Moreover, if you ever have to study or teach Quantum Mechanics, physical intuition gets in the way. A lot of properties contradict the physical intuition but after a while you get use to them. If we continue with Einstein, the photoelectric effect does not aperar in real life.
(Note that the HP Chipmunk 9836 also had a scroll selector wheel in 84)
Most people are UI bigots. Once they get used to a first something, they expect everything to work that way and hate learning anew. They get stuck on keyboards, mice, trackpoint nubs, trackpads, trackballs, scroll wheels, or touchscreens and refuse to move on. Of course there are 'objective' performance tests for each including Fitt's test of accuracy and latency, as well as, cognitive load. I guess once you have a hammer, every screw looks like a nail.
So I'd be a bit careful with the "of course!". It may well be obvious only because that is the first mental model you latch onto.
Of course it does. How do you think your phone camera works?
Another bad idea for a movie is a watermillpunk universe, where during a practice on a hot day the best ever curling player discover inercia.
TFA was actually about leaps of intuition, sadly.
One of the experiments I've heard proposed around here is to somehow create an LLM from all text up to 1980 or 1990 and see if it can get back to making itself.
My feeling is that a prompt would have to provide a vague description of a program that meaningfully passes something like a Turing test, an API to conform to, an expectation of novel construction (no 'ifs all the way down'), and then a requirement to search broadly and pursue promising ideas and not get hung up on the philosophy. Anything more precise feels like it would corrupt the test, but as it is that description feels doomed to loop before even trying the interesting parts.
IMO, the mechanism isn't the important thing, the behaviour is. If you look at the step-by-step, we are also looking for the next word or motor action (and for whoever is about to suggest that we humans plan ahead, Transformer-based LLMs have been shown to also do this); as this is not a useful description of what it means to be a living brain, I'd say it's also not a useful description of what makes everything post-InstructGPT different from what came before.
<noob> Where do birds go when it rains?
<expert> They
I know the answer: because it leads to model collapse. But why is that? Wouldn't a smart model not collapse? It's seeming like they keep getting smarter because we keep pouring more of our own knowledge into them, not because they are actually getting smarter. And yes, sometimes a dumb but persistent bruteforcer can make new discoveries.
and i think this is exactly the crux;
the really big models need really big datasets
and current gen LLMs get a lot of training data beyond "all books + all of the internet"
the objection is then that producing this additional data would already confound it with pre "virtual cutoff date" knowledge (since the training data probably implies mathematical and SWE concepts that were developed post "virtual cutoff date")
But to prevent model collapse you need a way to pump down the entropy. Much like in thermo, it's an expensive and slow process.
Chessboxing was the invention of comics book artist Enki Bilal (and he's credited with this in Wikipedia). I first saw it in his Nikopol trilogy. Because life is weird, it then became a real thing.
It's unrelated to computers playing chess. It predates Kasparov's first defeat by Deep Blue. I don't remember any mention of computers being good at chess in the trilogy, either. Or any computers, for that matter.
Could be, but preventing leakage from more modern stuff can be challenging.
This was attempted with Victorian public domain content: https://www.estragon.news/mr-chatterbox-or-the-modern-promet...
I can't find the citation right now, but I think people found it was leaking anachronisms? So this probably wasn't as well filtered as the creator had hoped?
At a minimum, yes. IIRC, the sum total of all compute manufactured over history only reached the minimum needed to train an OK LMM in the mid 00s.
> How much could it infer from it?
Only way to find out is to try.
This is an interesting experiment but I wonder if it would be possible to prevent some sort of retrospective bias. For example, I’d expect the experiments that lead to relativity to be over-represented in our catalogue of scientific literature prior to 1900, just because in retrospect they were important, so the records about them were preserved.It would have to be a very intentionally constructed corpus, I think.
https://didof.dev/blog/talkie-1930-llm-reasoning/ seems like a decent overview
It's actually possible to answer this question rigorously:
1. Define a scientific result which qualifies as a "jump". They should be frequent enough that they happen every year - otherwise one might say humans can't jump either.
2. Identify all such "jumps" in articles published in 2026, and use LLM with 2025 knowledge cut-off to re-derive these results with minimal amount of information.
It really irks me that people boost these low-effort articles just because they confirm pre-conceived notion that LLMs are limited
To Yann LeCun it was "obvious" that you can't build a world model from a text, so he predicted that even "GPT 5000" can't predict that an object sitting on the table would move together with the table. But GPT-4 could solve this task and provided correct explanation. GPT-5 can solve much more complex tasks, code physics simulation, etc.
Now people talk about invention of General Relativity as something LLMs can't do... something almost no human was able to do either.
The idea that everything needs to be proven with evidence is what an LLM excels at. If you give it a task such as "analyze this massive dataset and demonstrate that is proves or disproves this concept" or "find a counter argument to this theory", it will dutifully run through the information and give an answer. If you give it the task of "invent a new style of music", it will do poorly.
It will not do poorly from of lack of ability, nor lack of originality, it will do poorly as "style of music" is a human concept, "new" is a human concept and "invent" is a human concept. Being human concepts, these are things without quantitative value. There is no truth or certainty to what is "invented", as others may call it "discovered" or "developed" (developed in the sense of progressing from a previous work). While a trained ear my call a song "Breakcore" and not "Gabber", another would clump it into "EDM".
The only solution to this is to have human origin. As humans will tell the tale of how an idea came about, inspiring similar thought patterns in peers and earning the title of "new". This to me seems to happen more often with art than science or engineering. Although, it is hard to tell if it is due to increased creativity, or ability to convey thoughts.
All of this is impossible for a non-human to do, as they will give evidence to support the conclusion rather than the tale. This will turn into a possible addition to the knowledge of humanity, but can never earn acceptance as a "jump", as humans care little for evidence. Humans are inherently social beings and, as a side effect, will never acknowledge a non-human contribution.
An easy way to put this is a literal jump. The humans at Merriam-Webster define it in part as "to spring into the air". However, a quick peruse of the comments on a video about the "World's Highest Jumping Robot" will find disagreements about the definition of "jumping" and "robot". The device does spring into the air, but it does so without human legs, human movement, or human consideration for safety. Should you instead look at a video on the highest jumping bipedal robot, you will instead find nearly full acceptance that what happened was a "jump" (it should be noted that there is much disagreement in comments about if such videos are AI generated). By my understanding of the world, this is due to human ability to excite brain paths of an action when viewing another entity preforming the action. So if the viewer's brain does not fire the signal of a jump motion happening, the viewer will seek any rationale to explain the action as not being a jump.
TL;DR: It is all based on vibes, and LLMs give the wrong vibe.
"In abstract domains such as Mathematics or Computer Science, the Sense Experience (E) may be grounded in high-dimensional topology or have other goals such as generality or minimality."
But if such sense experience is possible in abstract domains via some high-dimensional topology, why could a sufficiently advanced LLM not develop an equivalent high-dimensional topology for domains like physics and use it to make creative leaps?
But we have no idea at how good humans are at that. Given the appalling failures of humans to handle even basic statistical situations like identifying that the same thing happens over and over, it might be that they are hilariously bad at creative leaps in abstract fields, it is just we have had nothing better available to measure against. We've spent about as long as decision theory existed trying to convince people to use it instead of flailing. Limited success, usually in exceptional cases.
And the paper seems a bit dodgy, we have models created with sensory data available. No reason a LLM can't be trained on more sensory data than a human can accumulate in one lifetime. There is a lot of visual data on YouTube.
The case study they chose is literally Albert Einstein coming up with General Relativity, something most scientists of his time were not able to do
One interesting (albeit sad) area which might be related are humans who are never raised with a first language. They seem to never developer abstract reasoning and even seem to lose the ability to develop it later in life. This might indicate there is some 'real world senses' -> 'direct language' -> 'indirect language' -> 'abstract abduction' hierarchy that develops, perhaps related to more real world abductions as a necessary side chain to developing abstract ones.
One of the obvious problems with this is just how difficult we find it to study intelligence purely in humans. We are measure a LLMs by a yardstick that is already known broken, but maybe this is still the right path.
Until LLMs have some 0% error humans will have to be in the loop (even if they only serve to take responsibility of the process).
Saying you want to make workers more productive and provide better tools for people just isn't that sexy.
An LLM in isolation from its environment might as well be a brain in a vat in some dark cave. You need an external environment to sample from and act upon to make forward progress.
That said, here's an experiment conducted by some Soviet psychologist, I forgot the name. The man wanted to study intuition. So he invented an experiment that was supposed to trigger it in laboratory conditions. (Take a moment to marvel at that; how would you approach such a task?) He gave people a few puzzles. One was to place some sticks according to some rules. Yet another was to find a path in a maze. The secret was that the path in the maze was the same figure as the solution to the stick puzzle.
And he observed interesting results. People who solved the maze after the sticks found the path much faster than the control group. If a subject was asked to comment how he was solving the maze, at the start or halfway through, the speed dropped to typical. Subjects normally didn't notice the similarities.
So there is something to study here, although it is obviously a case of pattern matching, only subconscious. This is a jump of sorts, but not the one I mean. What I mean is a Zen jump.
I see similar thinking in stories of how humanity got here. Religion has thousands of years adapting to this problem, every time we explain something, the goal post moves. Catholics today accept evolution (or least the church does), but it is the "jump" from monkeys to humans where God is the only explanation.
Just 5 years ago we didn't have a technology that knows more about everything than even most experts. We keep coming up with benchmark after benchmark and LLM/AI keeps destroying them. Now we've moved the benchmark to "the jump". Again, maybe it's LLMs or the way we currently do them that can't do this, but eventually something will.
And we still don’t. What we have are simply very advanced search results aggregators with delusions of personality. Just because your fridge says “I” doesn’t mean it is a person.
We still don't have that. LLMs have shown time and time again that they don't know a single thing and are incapable of reasoning.
For example "..ARC captures the logical leap, it misses the manipulative component—the physical sensation and embodied simulation..." makes lots of assumptions on how such a discovery must occur, e.g. through "physical sensation and embodied simulation". Results matter, not the path there.
For example, quantization of energy, at the core of QM, wasn't discovered through "physical sensation and embodied simulation" at all. Planck simply found that if energy is quantized, then one obtained the observed black-body radiation spectrum. There was no "physical sensation and embodied simulation".
https://news.ycombinator.com/item?id=49136070 https://news.ycombinator.com/item?id=49096837 https://news.ycombinator.com/item?id=46890333 https://news.ycombinator.com/item?id=46870562
All of these are titled “LLMs Can’t Jump”
LLMs are inherently probabilistic, and there's currently no mechanism for producing an orthogonal directional change in the path traced through a latent space which is also contextually relevant (landing on a punch line).
In other words, LLMs are fundamentally incapable of making intuitive/orthogonal leaps in context.
It might be possible to add this capability with a new architectural component like transformers, but specifically for making "left turns"/intuitive leaps.
"Probabilistic next word prediction" and "humor" sit about as far apart as "modulating airflow with meat flaps" and "humor" do. One is an interface through which an action is performed and the other is a highly abstract capability.
Would you claim that a podcast comedian is fundamentally incapable of being funny because all he ever does is wiggle the air with his throat meat flaps? Probably not.
Absolutely nothing about "probabilistic next word prediction" forbids "making intuitive/orthogonal leaps in context". The interface is expressive enough.
And empirically? The "sense of humor" in LLMs is yet another "a function of model scale" capability. GPT-4.5 was reportedly funnier than both GPT-4o and o1. Fable 5 is reportedly funnier than Opus 4.x. It's one of those ever-elusive "big model smell" signs that are hard to measure with anything other than vibes.
Under the "humor as an opposed social intelligence test" family of hypothesis, what "being funny" reflects is the funny guy's ability to model and predict you and your reactions. For the comedian to be able to make the audience laugh, he must know his audience well, model it accurately enough to be able to spot the "breaking points" of humor, things they'd find unexpected and clever and thus "funny", and then weave those things into the jokes.
Then, a bigger LLM gets better at humor because it has a more accurate model of how humans think of things - including the "ha-ha" gaps. It's a "theory of mind" capability. It's not "special", it's just hard.
I think you're saying that you can eventually train models to arrive at that destination by training on existing jokes, effectively encoding these leaps as probabilities.
In that case, the model isn't actually making an intuitive/comedic leap; they're just following new probability chains in attempting to approximate examples they've seen in training.
I'm suggesting that something architecturally different is necessary to create a model which can make intuitive/comedic leaps.
Try to get a frontier model to write a clever, funny joke which hasn't been seen before. Or, try to get it to make an intuitive leap that leads to a novel discovery.
You can use them to guide your own efforts along these lines, as a sounding board. But with current architecture I just don't think either is possible for an LLM to do on its own.
This is what I refer to when it comes to larger models like Fable 5 being funnier. They are more capable of doing that. They can deliver that "sudden orthogonal leap from context" of yours more reliably.
It's not a "fundamental inability" and never was. If you crank the scale up and a capability appears, "current architecture" was never the problem.
Just another emergent capability - one that follows exactly the same patterns any other emergent capability does.
idk the harnesses adding in "but wait -- " or some variation every few lines in the thinking trace seems to do great at this.
For this paper specifically, after reading the abstract [2], I felt almost certain that the author would have used Judea Pearl's ladder of causation (https://web.cs.ucla.edu/~kaoru/3-layer-causal-hierarchy.pdf) but they did not. Would have probably been a better argument to make.
[1] paper in quotes because it may never get published (it is over 20 pages atm). the core argument is that lack of native adjacency resolution makes problems harder and sample inefficient, not impossible
[2] "Using Einstein’s formulation of General Relativity as a case study, we demonstrate that LLMs are structurally incapable of creating new foundational axioms, particularly when observational data is scarce. "
Also, the claim that 'LLMs are structurally incapable of creating new foundational axioms' is provably false depending on where you place 'fundamental'.
Every "can't" of this nature was followed by a discovery of "they can, just poorly", and then by that "poorly" improving steadily generation to generation.
The paper doesn't provide a way to measure or quantify this elusive "jumping" capability, not even as an approximation. It just throws "can't jump" out there, as if "abduction" is an established class of problem with known computational properties and requirements that the LLM architecture fails to satisfy. It's none of those things - and the paper makes the claim without backing it by anything but rhetoric attempts at persuasion.
The proposed solution is also dubious. The empirical track record of dedicated "world models" for reasoning and problem-solving is, frankly, downright abysmal. Even integrating multimodal data into LLMs has failed to yield general reasoning capability gains.
LeCun's misadventures in the field aside, the main frontier lab that pushes in favor of "improving reasoning via multimodal fusion" is GDM - and Gemini isn't exactly a paragon of frontier reasoning capabilities. It has strong multimodal capabilities, but lags behind both OpenAI and Anthropic in performance outside that - while Anthropic is the lab that always treated multimodal grounding as an afterthought, and still trades blows with OpenAI at the very edge of the performance frontier. Multimodal grounding seems to work great as a way to improve an AI's ability to deal with those specific modalities, but it falters outside that.
Now, it's not impossible that everyone who tried multimodal world models for reasoning is just doing it wrong, and there is an undiscovered recipe for multimodal grounding that results in a step change in AI capabilities. But the results we have so far suggest it to be unlikely.
What LLMs are "fundamentally incapable" of doing has striking parallels to https://en.wikipedia.org/wiki/God_of_the_gaps
My opinion of claims like "LLMs need memory to manage codebases" has also hit the dumpster bin a while ago.
Why would knowing how to make a maintainable change to a codebase require any more "memory" than knowing how to play an optimal chess move? The codebase is the memory. A sufficiently capable LLM can ingest it, figure out what changes to make, and make them.
Were that the case LLMs would’ve been phenomenal code monkeys from the get go. They were not. They still are not.
2/ SWE are doing way more than "writing code", and since those areas are less "computational verifiable" (what is a good architecture that will stand in 5y?) and more "connected to real world" (what are the requirements?), LLMs struggle with them
This is also tied to halucinations: it is something that humans do (for writing fiction, and for "jumps") - but what LLMs currently lack is intellectual honesty. Coming up with bullshit is fine (and in this context valuable) - the important bit is putting those ideas through some form of rigor, or just immediately turn around and admit to talking shit.
So I'd arge that hallucinations are what prevent LLMs from doing this in a useful way.
If we could make progress in that area, maybe CoT could gradually decrease as it approaches its limit, or maybe the LLM could control the temperature of the next token itself (how this would be trained, I have no idea).
Interesting!
Induction Vs. Deduction Vs. Abduction!
(You know, if you like Logic, Philosophy, Law, or... just plain different ways to think/reason about something! :-) )
Turns out temperature is pretty bad too, you can find ways to sample from deeper in the distribution without distorting it. Great example is XTC (exclude top choices), In a few weeks/months it'll also have a proper scholarly paper with peer review.
In math its simple: does the verification say its okay.
If its mechanical: is any property better than what we have already.
etc.
I just don't see any of the LLM users around at all. Clearly some force is guiding them all away from thinking any of the "leap of faith" thoughts that I am thinking.
You may think this is not a good test because an older (or say a smaller) LLM can study from the knowledge on the Internet and build. But we are like that - we can access the Universe through our senses.
Can we ever produce anything that is beyond this Universe? I think an LLM that is lacking in knowledge can build more complex systems as long as it can access more data.
The only way a LLM can come up with new ideas if the "idea" appeared as a generalisation durring training or if it was achieved using reason in chain of thought.
A jump in intuition comes from automatic processes reorganising the relational structure of conceptual models. There is no reorganisation of the model durring inference.
one could argue that model can reorganize / interact with prior knowledge captured in text form (edit files) hence there can be reorganisation
Ordinarily in mathematics there’s a TON of papers to write just combining low level problems with different techniques. Better still, and often considered groundbreaking is borrowing techniques from other fields and adapting them or creating analogous methods to solve problems. A lot of landmark papers have been written this way. This is also what transformers are sort of good at within other contexts. They have super human breadth so I’m hopeful they’ll become real assets in math for a long time. Though the leaps necessary to adapt a technique in a nonobvious way might be too much for a while longer. We’ll see.
Truly novel techniques are quite rate indeed and I don’t know if LLMs can represent them faithfully in their embedding space or not. My inclination is that they probably can most of the time, but I don’t know. Mathematicians would describe such thins as “alien”.