Worth reposting a follow-up tweet from the author Tom Zahavy [1] after this made the rounds on X/Twitter recently:
> 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!
Came for:
"A computer once beat me at chess, but it was no match for me at kick boxing."
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.
The curious case here is how much of a description do we give it of itself? That would almost certainly dominate success rates.
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.
I wonder if we could just tell it to invent itself without any description and see if it can I introspect enough through its own interface to figure out what it is.
The popular retelling of how Einstein created Special Relativity to "Resolve the contradictions of Michelson-Morly experiments" is very reductive to the history of the question. The epitome is the quote from the paper:
> 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.
I think this is more of a function of the harness and the environment than the LLM. I've seen some LLM interactions over complex environments like Godot and Unity that challenges the notion that there is no "jumping" going on at all.
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.
It's a very shaky position, and the empirical track record of "LLMs can't..." is in itself a reason to call it into doubt.
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.
I found this paper really thought provoking, but I think the conclusion of “world models are the solution” leaves something to be desired. People are already equipping agentic systems with physical simulation tools and exploring action-conditioned world models. This is cool because you can change the rules of the simulation and observe what happens, but it doesn’t address the core question of what to change the rules to, or even what the goal should be in the first place.
I have been writing a 'paper' [1] on an adjacent topic for months now. At some point, I decided to make it an empirical paper vs position paper. I am still chasing the experiments (when I get some free time waiting for agentic loops)
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'.
Clearly LLMs cant do leaps of intuition since their "intuition" is locked after training ends.
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.
The early physics background is messy and incorrect. I didnt read the full position paper, but from its start: The Lorentz transformations were by Lorentz, well before Einstein’s paper on special relativity; the principle of relativity also existed before the Einstein paper. The math was all there, with steps taken by Maxwell, Voigt, Larmor, Lorentz, and Poincare. Einstein supplied a clean physical interpretation, making all inertial frames equivalent, making simultaneity frame dependent, and explaining length and time deformations without the need of the concept of ether. Skimming the end of the paper with the arguments about lack of abduction or inability to make the analogy without sensory experience, I see that this paper is unfounded speculation rather than solid/hard philosophical logic. As a position paper it is OK to appear, but i think it misses the point of how LLMs or other autoregressive learners of future states can build analogies and intuition that can help them formulate new theories of the world. Soon it will be more obvious to everyone, so I am not very worried about these writings.
The theory is that creative leaps in theoretical physics require a grounding in sensory experience, but the obvious counter-argument is that humans can make creative leaps in abstract fields without such sensory grounding. They do address this at the end, saying
"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 the obvious counter-argument is that humans can make creative leaps in abstract fields without such sensory grounding...
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.
That's an interesting analogy. My gut sense is that theoretical mathematics requires a high level of intelligence versus more grounded domains. That may imply that deficiency in grounding can be made up for with intelligence and basically reverse engineering the gaps in grounding from first principles/limited grounding. The ultimate question would then be what is the tradeoffs between grounding and raw intelligence for the same outcome.
If there is enough cross over between real world knowledge engrams and abstract knowledge engram, would this allow for the jump?
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.
I have a weird thought experiment: If you give a GPT-2/3 level LLM tools to search the internet - any document, can it build bigger, better LLMs?
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.
This tracks for me as someone making keeps of intuition in little-explored areas.
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.
I feel like you could just add some noise or randomness to the LLM and start approximating the leaps that the human mind uses to solve and understand unrelated things. Maybe that’s naive, it’s just coming from my organic computer in my skull.
Technically true, but the counter argument would be that the probability of this working would be ~ 2^(-(entropy_of_leap)) for an LLM (presumably intractable) and a human would succeed at a higher probability.
Hahaha you just derived temperature from first principles.
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.
Worth reposting a follow-up tweet from the author Tom Zahavy [1] after this made the rounds on X/Twitter recently:
> 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] https://x.com/TZahavy/status/2082401499628376180
Came for: "A computer once beat me at chess, but it was no match for me at kick boxing."
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.
I think this can't work because an LLM needs too much data, and before the internet there probably just wasn't enough to get close to what we have now
Even simpler: Can GPT-2 anticipate and build Gwen/Deepseek? I think the answer is almost trivially "no", so I wonder what changed?
Is that how chessboxing was invented? Genuinely asking.
The curious case here is how much of a description do we give it of itself? That would almost certainly dominate success rates.
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.
I wonder if we could just tell it to invent itself without any description and see if it can I introspect enough through its own interface to figure out what it is.
Article was plenty interesting to me.
The popular retelling of how Einstein created Special Relativity to "Resolve the contradictions of Michelson-Morly experiments" is very reductive to the history of the question. The epitome is the quote from the paper:
> 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.
I think this is more of a function of the harness and the environment than the LLM. I've seen some LLM interactions over complex environments like Godot and Unity that challenges the notion that there is no "jumping" going on at all.
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.
Why does the cave need to be dark if its just a brain in a vat?
It's a very shaky position, and the empirical track record of "LLMs can't..." is in itself a reason to call it into doubt.
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.
I found this paper really thought provoking, but I think the conclusion of “world models are the solution” leaves something to be desired. People are already equipping agentic systems with physical simulation tools and exploring action-conditioned world models. This is cool because you can change the rules of the simulation and observe what happens, but it doesn’t address the core question of what to change the rules to, or even what the goal should be in the first place.
I have been writing a 'paper' [1] on an adjacent topic for months now. At some point, I decided to make it an empirical paper vs position paper. I am still chasing the experiments (when I get some free time waiting for agentic loops)
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'.
Previous: https://news.ycombinator.com/item?id=49162791
Also:
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”
Also https://news.ycombinator.com/item?id=49136070
You posted this twice now. Why not just edit your original comment?
Clearly LLMs cant do leaps of intuition since their "intuition" is locked after training ends.
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.
The early physics background is messy and incorrect. I didnt read the full position paper, but from its start: The Lorentz transformations were by Lorentz, well before Einstein’s paper on special relativity; the principle of relativity also existed before the Einstein paper. The math was all there, with steps taken by Maxwell, Voigt, Larmor, Lorentz, and Poincare. Einstein supplied a clean physical interpretation, making all inertial frames equivalent, making simultaneity frame dependent, and explaining length and time deformations without the need of the concept of ether. Skimming the end of the paper with the arguments about lack of abduction or inability to make the analogy without sensory experience, I see that this paper is unfounded speculation rather than solid/hard philosophical logic. As a position paper it is OK to appear, but i think it misses the point of how LLMs or other autoregressive learners of future states can build analogies and intuition that can help them formulate new theories of the world. Soon it will be more obvious to everyone, so I am not very worried about these writings.
The theory is that creative leaps in theoretical physics require a grounding in sensory experience, but the obvious counter-argument is that humans can make creative leaps in abstract fields without such sensory grounding. They do address this at the end, saying
"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 the obvious counter-argument is that humans can make creative leaps in abstract fields without such sensory grounding...
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.
That's an interesting analogy. My gut sense is that theoretical mathematics requires a high level of intelligence versus more grounded domains. That may imply that deficiency in grounding can be made up for with intelligence and basically reverse engineering the gaps in grounding from first principles/limited grounding. The ultimate question would then be what is the tradeoffs between grounding and raw intelligence for the same outcome.
Isnt the sensory grounding even in abstract cases some (limited) intuition that simulates in a mental world model?
If there is enough cross over between real world knowledge engrams and abstract knowledge engram, would this allow for the jump?
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.
Best comment on this from 6 months ago: https://news.ycombinator.com/item?id=46870575
I have a weird thought experiment: If you give a GPT-2/3 level LLM tools to search the internet - any document, can it build bigger, better LLMs?
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.
This tracks for me as someone making keeps of intuition in little-explored areas.
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.
I feel like you could just add some noise or randomness to the LLM and start approximating the leaps that the human mind uses to solve and understand unrelated things. Maybe that’s naive, it’s just coming from my organic computer in my skull.
Technically true, but the counter argument would be that the probability of this working would be ~ 2^(-(entropy_of_leap)) for an LLM (presumably intractable) and a human would succeed at a higher probability.
Hahaha you just derived temperature from first principles.
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.