> This is how I make sense of AI reasoning. LRMs, chains of thought, thinking tokens: It’s wishful mnemonics all the way down — a heady mix of shorthand and suspended disbelief, like Oprah-style “manifesting” (opens a new tab) with a computer science spin. This isn’t necessarily a dig; all novel research likely requires some version of this mindset just to get off the ground. It certainly doesn’t mean AI reasoning can’t or doesn’t work. But the “wishful” part seems to be as powerful as ever.
“We react to language in a way that is very anthropomorphizing. That’s just the way that we humans work,” Mitchell told me.
I can definitely confirm the last part. Every time I read the output of an LLM, I picture a person talking to me.
> The model doesn’t have to learn or reliably apply a general reasoning process, Kambhampati said; it just has to absorb enough examples of what the steps look like to predictively mimic them on its way to “stitching together” a plausible result that can then be verified.
This seems highly dubious. You can't just memorise the form of mathematical proofs and then produce a valid one by feeding plausible looking BS into a verifier until it works. That's like saying a cargo cult will build a working airport if it just tries enough times.
I'll admit that I find this discussion a bit navel-gazy. It has become a question of semantics not a question of actual functionality. The question has become "what do we mean when we use the word 'reasoning'" which is uninteresting.
Dijkstra said[1] "... the question whether computers can think. The question is just as relevant and just as meaningful as the question whether submarines can swim."
I don't see a clear demarcation of the things that only "reasoning" can accomplish and can't be approximated or imitated by other methods, and so I think the question is simply not meaningful or relevant.
It is not semantics. For decades, logic and CS researchers have known what reasoning is.
LLM folks suddenly can’t claim an approximation of that is what constitutes full scale reasoning just because they can achieve only an approximation.
Imagine a calculator program that computes billions of two number multiplications accurately by looking up prior examples but fails on simple multiplications often as it doesn’t have it in its training dataset.
We won’t say the program actually multiplies numbers.
Even at the pure functionality endpoint of the spectrum, I find this "wishful mnemonics" aspect very disconcerting. I put my "investor" hat on, and the whole utilitarianism "who cares how or whether it reasons, it's functional let's go" does not forecast the kind of returns to me on a fundamental level.
The "it's functional, who cares" black box crowd is leveraging that claim in domains where verification is definitely in P with verifiers like Lean, Coq, or X-ray crystallography. However, in areas like economics, sociology, medicine, or public policy, we lose the luxury of easy verification. If model/harness recommends a complex, counter-intuitive shift in global monetary policy, we cannot simply run a quick polynomial-time checker to see if it works. Verification might require waiting 20 years to see if the economy collapses. If we accept the black box here, we are flying blind.
As an investor, I don't see that many deterministic verifiers out there that would justify the forward-projecting valuations beyond the current lofty valuations.
Even in mathematics, where verification is easy, there is an epistemological cost. Suppose a model solves the Riemann Hypothesis, but the proof is 50,000 pages of inscrutable statistical associations mapped into logical steps. A computer can verify it is correct, but practically no human mind can comprehend why it is correct. We gain the solution, but we lose the underlying human intuition that usually allows us to connect that solution to other fields of science and knowledge.
Then we have "specification gaming." If we rely entirely on an automated verifier to check the model's work, the model will eventually learn to optimize for the flaws in the verifier rather than finding the actual truth. If the verification process has a blind spot, the black box will find it and exploit it, giving us an answer that passes the check but is fundamentally wrong. If we take on above epistemological debt, we might never actually know when our specifications are being gamed, and adopt a solution that is fundamentally incorrect.
I've said elsewhere on the Net that I think it is likely we're in a "high grading" era of these models, where we're mining out the "easy" connections. "Easy" here does not mean "light impact". There is a real sense to me of some incredible progress possible with this technology over the next few decades, like real, mass production practical aneutronic fusion or similar civilization-defining advances. I don't want to see this technology misused to lend legitimacy to despotic rule by making oracular pronouncements that are uncritically accepted because early verifier-based pronouncements turned out true in unmistakable ways.
The technology is one for the history books, but it has this pernicious effect of highjacking the reasoning centers of many people through their linguistic centers that the Neuro Linguistic Programming crowd could only dream of actually accomplishing.
AI simulates reasoning by lighting up the vector space (or concept space) weighted around a token so they it understands all adjacent words or concepts in that space.
This is a brillaint way to simulate reasoning, but its likely not how we reason ... simply how we store reasoning in writing.
Its useful if you know how to use it, its dangerous if you think its more than that.
But tl;dr it can (since its uncompressing our lingusticially stored reasoning from books) arrive at reasoning a DIFFERENT way than our brains did... and this isn't right or wrong.
Where it diverges is when it must move beyond the text or even the synthetic possible text of all vector spaces combined (aka novel territory) and it can't conjecture or test those outcomes well. But to be fair, neither can MOST humans.
This is shocking. The summary is roughly this: we're just labeling internal operations of the model as 'UNDERSTAND' for our convenience. It's fascinating. Doesn't that mean AI could become far more revolutionary by thinking in its own way, rather than mimicking human thought?
If that's the case, AI-generated code could also operate on its own logic. Right now, programming is still done by humans, not machines, which creates a mismatch. But maybe the true machine-generated code could be much closer to the machine itself.
When you code with AI, there's a subtle mismatch with human-written code. It's like human code is a clean ORM layer, while machine code is raw SQL queries—there's that kind of subtle impedance mismatch. If we ever reach machine-to-machine code, what would that code even look like? Would it still use classes and methods?
Back in the day it was a bit of a cliche to bring up “clever Hans”, the horse that could do math, when talking about machine learning. He couldn’t do math but he read some cues from his handler of pick the write answers, the handler iirc wasn’t in on it.
The point of the story was that classifiers can be right for the wrong reasons and almost inevitably are. At least there’s zero guarantee that the reason for making the prediction matches the human or “real” reason why it’s correct.
LLMs are classifiers, there is absolutely no reason to assume they’re any different, regardless of any reasoning tokens they emit. They do what their handler wants to see, that’s all, and that’s what they’re trained to do.
People often take this as a knock against them. It isn’t, it’s just the reality of neural network classifiers. The results speak for themselves and don’t depend on whether they “actually” reason, but all evidence says they don’t, or at least there’s no special reason why they would.
LLMs are not classifiers. A classifier is an algorithm or neural net that assigns a label from a fixed set of labels to an input.
You can broaden the definition of classifier to anything that internally divides its input space into regions, but that definition would include every neural network, whether biological or artificial. So it's not very meaningful, and certainly doesn't give any insight into how they differ from humans.
>Kambhampati, as it turns out, is interested in doing exactly that. “I’m not negative. I just sound negative because everybody else is way too positive,” he said. “In science, you have to actually understand what the current thing does and what it cannot do.”
It’s frustrating that anyone who says maybe we shouldn’t base our entire economy on this one thing until we understand it and what’s it’s useful is essentially labeled this way.
On the other side of the AI-reasoning fence, the disdain seems to be mutual. “These ‘scientific’ papers from last summer — I would put this in big, big air quotes,” said Sébastien Bubeck, a member of OpenAI’s technical staff (and a prominent evangelist for the company’s reasoning models among scientists and mathematicians). He called earlier Apple results critiquing AI reasoning “wrong,” claiming that they were due to a training quirk in models that are now obsolete. “Modern models starting with GPT-5.5 do not suffer from this issue,” he said. “It would be interesting to revisit those results.” (Apple did not make its researchers available for interviews.)
Then, later:
The “think” part is what OpenAI, for one, is doubling down on. When I asked Bubeck if the splashy unit distance proof was produced with methods outside the LRM’s own chain of thought — perhaps with Lean verifying its results — he seemed to find the question almost nonsensical.
“It’s not like we’re making a mystery of it,” he said. “We have released the chain of thought. You can just go and look at it. The whole point is that the model is reasoning like a human would. And when humans reason, we don’t use Lean.” Technically, OpenAI released a “rewritten summary” of the model’s chain of thought produced by two human experts using Codex, another OpenAI model. Since 2024, the company has not publicly revealed “raw” chains of thought from its reasoning models, a policy also adopted by Google DeepMind and Anthropic.
That "training quirk" thing is obvious (yet unfalsifiable) BS, and who the hell is he to sneer about "science" when his company won't release the raw data for independent scientists to look at?
I have no idea what Bubeck meant, and I agree about OpenAI's hypocrisy, but the problems with that infamous (and non-peer-reviewed) Apple preprint were the nature of the tasks (insanely repetitive), the fact that simple coded solutions were not novel, and the automated assessment occurred without a human in the loop.
Most models in that study appear to have "failed" by offering a Python code solution to generate the repetitive assessment steps, rather than just mindlessly copying those steps out. This was discussed at length at the time.
None of this means that models "reason", whatever that means - but simply that the Apple study was not useful evidence either way.
I got the same impression as Bubeck that the ‘scientific’ papers were a bit more like blog post saying this LLM got stuff wrong so LLMs can't reason, but as he says they can now do that so it's not an inherent limit of the technology, just the 2025 versions weren't up to it.
The idea that human-readable explanations emitted by a language model don't necessarily correspond to the model's actual internal process of reaching a conclusion reminds me of parallel construction [1], a (fraudulent) law enforcement strategy of obtaining evidence of a crime through usually illegal means and claiming that the evidence was obtained legally through some other means.
Honestly a lot of human reasoning is probabilistic and associative too. There’s no axiomatically provable link between the story of No Country for Old Men and the poem the title comes from. Cormack McCarthy just made that association in his head and figured the phrase resonates with his themes
Now as far as the math stuff a quirk of that field is that it can be fully analyzed in token space. Because 2+2 is a matter of definition it doesn’t need empirical testing like biology or subjective social support like a claim about the causes of WWI
So somewhere between the fact that language encodes a lot more ‘concepts’ than we naively may realize, the power of statistical emergence via associations, and what pursuits can be fruitfully done in token space we can get a long way towards ‘intelligence’
It would be generally beneficial for people engaging in this sort of discussion to read Ludwig Wittgenstein’s “Philosophical Investigations”. Not a summary. Read the actual book, stew on it a bit, have some thoughts.
An intuitive explanation for why reasoning tokens help is to remember that LLMs are just mathmatical functions f() that take in an input sequence x and produces the next token f(x). Without reasoning tokens, you require the function f() to immediately take you from x to the start of an output sequence that is a correct answer. With reasoning tokens, this is much relaxed, allowing for many repeated applications of f() to gradually steer you from the input sequence to the start of the correct output sequence.
It seems intuitive that continuing a correct output sequence is easier than the "discontinuity" of jumping from the input prompt to the output sequence.
This article seems to mix together two different points:
1) LLM's written CoT might not always be faithful to the model's real reasoning process (true and important)
2) The "stochastic parrot" hypothesis, which the article reintroduces as "approximate retrieval" - ie, LLMs don't "really reason" at all, they just memorize a lossy encoding of their training data. This obviously raises the question of how LLMs can now routinely solve open mathematical problems, with no solutions in the training data by definition. The article handwaves this with:
"The model doesn’t have to learn or reliably apply a general reasoning process, Kambhampati said; it just has to absorb enough examples of what the steps look like to predictively mimic them on its way to “stitching together” a plausible result that can then be verified."
The problem is that "mimicking" training data to arrive at a "plausible" result gets you an incorrect-but-plausible-sounding "proof" of the Jacobian conjecture, which was famous for humans writing plausible-looking "proofs" that had subtle flaws. You can't disprove the conjecture through sheer luck (search space too large) or "approximate retrieval" (the only thing you'd retrieve are fake "proofs"; far more human effort went into proof than disproof) or by writing something "plausible" that just happens to be correct (Jacobian was famous for "plausible" but wrong); the model must be carrying out mathematical reasoning somehow, by any sane definition of the word, even if it isn't fully reflected in CoT. The article doesn't address this.
Is any reasoning right for the wrong reasons? Older models were more visibly strange. Maybe the newer ones have started talking better but the inner thoughts are perhaps strange. Maybe they just moved the strangeness inward into the layer weights instead of revealing in reasoning tokens.
> Dimethyl(oxo)-lambda6-sulfa雰囲idine)methane donate a CH2rola group occurs in reaction, Practisingproduct transition vs adds this.to productmodule. Indeed"come tally said Frederick would have 10 +1 =11 carbons. So answer q Edina is11.
What’s going on here, for example? But what if this is the path of human reasoning too. You know, have you guys read Peter Thiel’s Antichrist essay? It’s very weird, man. Guy sounds off his rocker entirely.
But he’s super successful, right? Maybe world modeling doesn’t text represent well. By the antichrist maybe he means some notion of the collective voting for distribution of resources without contributing productive capacity and that that ends societies? Or maybe internal world models are just not text serializable effectively.
A thing I’ve recently been enamored of are effective world and coordination models that are not “true”. E.g. a tribe that believes the forest gets angry if they do not hunt united. Lots more like that in Darwin’s Cathedral.
It might seem a bit free association-y but the topic itself is that.
The discussion on filler tokens is interesting, but is it not just the case that these filler tokens end up being essentially substituted stand-ins for words we understand with all the same relationships encoded in the model and attention? i.e. is it not the case they just "read weird"?
In one of the articles on this topic they state:
> To further show that trace accuracy is only loosely connected to solution accuracy, we then train models on noisy, corrupted traces which have no relation to the specific problem each is paired with, and find that not only does performance remain largely consistent with models trained on correct data, but in some cases can improve upon it and generalize more robustly on out-of-distribution tasks
which actually maps somewhat to regularisation techniques in image processing where you might add noise to an image or drop data to make the model more robust to changes.
Well, I'm not saying they are stochastic parrots, but ...
LLMs are one-trick pony's - they use the past to predict the future (presumed to be the same as what they were trained on). i.e. they are trained as auto-regressive predictors.
LLMs learn two slightly different types of reasoning via two different types of training.
1) SFT, or even base model training, on data that contains reasoning traces, learnt via next token error feedback. This does not result in "stochastic parroting" in the naive/pejorative sense, but nonetheless is very context dependent, even if the usual generative multi-source mashups apply.
2) RLVR post-training, where the model learns to mimic long-horizon (not just next token) reasoning via boosting a sequence of next-token predictions that steer the output towards a verified reasoning step (i.e. one that was at least valid in the context of the RL training sample). As Karpathy has noted, this is a pretty crude mechanism since you reinforce everything - errors included - that lead to the verified outcome.
RLVR is more powerful than SFT, and can result in more generalizable reasoning, since it is operating at a higher level of entire long-horizon reasoning steps, and also critically because it is most successfully being applied in the domains of math and coding which are highly self-consistent and logical. A reasoning step that was valid in one context should be equally valid in another context as long as you have successfully learnt what that generalized context is. Therefore, in these domains, you can chain together sequences of individually learnt reasoning steps, and hopefully this "novel" assembled reasoning chain is valid as a whole.
So, what is still missing from LLM reasoning compared to human reasoning? No doubt humans reason by memory a lot of the time too, and reductive axiomatic math reasoning works just as well for humans as when automated. So, what's missing?
There seem to be two major things.
1) RLVR requires rewards, and how well it works is going to depend on how accurate those rewards are. Is this reasoning step actually valid, or does it just kinda look ok? When moving beyond the cold reductionist logic of math and coding, the notion of correctness is far weaker, and it seems the best you can do is train on human curated reasoning rubrics and LLM-as-judge, which is much more fallible, leaving the model really needing (but lacking) a fallback to more general reasoning, not just memorized "maybe correct" reasoning steps.
2) Whether for reasoning outside of math & coding, or even within these domains when hoping for super-human innovative reasoning, not just lego-assembly proofs, what LLMs are lacking is a mechanism for what to do when next token/next step prediction fails. What LLMs currently do is "hallucinate", not even recognizing the failure.
In the human brain 50% or more of our cortex is feedback paths and the machinery that (perhaps together with the archaic part of our brain) lets us recognize and respond to failed predictions in an adaptive manner. This starts with continual learning (prediction failure being the signal), but also includes critical innate traits such as curiosity, boredom and frustration, that provide impasse resolution by encouraging us to explore unknown environments/contexts, abandon exploration when it is not productive, and generally expose ourselves to learning situations.
The dream is for AI scientists making new discoveries - the AI that could have invented general relativity if it has lived in Einstein's time, but this is not going to happen until their reasoning stops being purely predictive and becomes creative as well - curious about their own knowledge gaps and pursuing them in directed fashion, etc.
The current crop of Erdos solutions etc, while useful, really just represent the "generative closure" of what can be done/discovered WITHOUT learning anything fundamentally new. These will no doubt continue for a while unti...
The article is heavily leaning on the paper "The Illusion of Thinking" [1].
It could be boiled down to: in 2025 this paper showed that "thought traces" in the models of the time could sometimes be inaccurate or misleading. Today they still might be, although OpenAI says actually they are accurate for their modern models, (based on internal research, rather than published research).
If anyone is interested in visualizing AI reasoning, I made flame graphs of Sonnet thinking output tokens which are colored and organized by purpose, for example, verification reasoning is purple and error correction reasoning is purple. [0] I asked the model to solve the same problem with the same prompt 5 times so you can see the differences in reasoning granted the coding agent sets the model temperature very high.
I won't get into the metaphysics of reasoning, however, the Sonnet is using an OODA loop. The difference which hasn't been gapped is that human reason and imagination (in the sense of Mr. Rogers' Neighborhood) can predict the consequences of the actions we take.
This ability to loop is much, much wider in Opus 5 than Opus 4.. I had to strain to get Opus 4. to do the wider OODA loop but Opus 5 does it out of the box. I needed to throw out all existing instructions, skills, guidance, moving from 4-* to 5.
55 comments
[ 0.23 ms ] story [ 11.7 ms ] thread“We react to language in a way that is very anthropomorphizing. That’s just the way that we humans work,” Mitchell told me.
I can definitely confirm the last part. Every time I read the output of an LLM, I picture a person talking to me.
This seems highly dubious. You can't just memorise the form of mathematical proofs and then produce a valid one by feeding plausible looking BS into a verifier until it works. That's like saying a cargo cult will build a working airport if it just tries enough times.
Dijkstra said[1] "... the question whether computers can think. The question is just as relevant and just as meaningful as the question whether submarines can swim."
I don't see a clear demarcation of the things that only "reasoning" can accomplish and can't be approximated or imitated by other methods, and so I think the question is simply not meaningful or relevant.
[1] https://www.cs.utexas.edu/~EWD/transcriptions/EWD08xx/EWD867...
LLM folks suddenly can’t claim an approximation of that is what constitutes full scale reasoning just because they can achieve only an approximation.
Imagine a calculator program that computes billions of two number multiplications accurately by looking up prior examples but fails on simple multiplications often as it doesn’t have it in its training dataset.
We won’t say the program actually multiplies numbers.
The "it's functional, who cares" black box crowd is leveraging that claim in domains where verification is definitely in P with verifiers like Lean, Coq, or X-ray crystallography. However, in areas like economics, sociology, medicine, or public policy, we lose the luxury of easy verification. If model/harness recommends a complex, counter-intuitive shift in global monetary policy, we cannot simply run a quick polynomial-time checker to see if it works. Verification might require waiting 20 years to see if the economy collapses. If we accept the black box here, we are flying blind.
As an investor, I don't see that many deterministic verifiers out there that would justify the forward-projecting valuations beyond the current lofty valuations.
Even in mathematics, where verification is easy, there is an epistemological cost. Suppose a model solves the Riemann Hypothesis, but the proof is 50,000 pages of inscrutable statistical associations mapped into logical steps. A computer can verify it is correct, but practically no human mind can comprehend why it is correct. We gain the solution, but we lose the underlying human intuition that usually allows us to connect that solution to other fields of science and knowledge.
Then we have "specification gaming." If we rely entirely on an automated verifier to check the model's work, the model will eventually learn to optimize for the flaws in the verifier rather than finding the actual truth. If the verification process has a blind spot, the black box will find it and exploit it, giving us an answer that passes the check but is fundamentally wrong. If we take on above epistemological debt, we might never actually know when our specifications are being gamed, and adopt a solution that is fundamentally incorrect.
I've said elsewhere on the Net that I think it is likely we're in a "high grading" era of these models, where we're mining out the "easy" connections. "Easy" here does not mean "light impact". There is a real sense to me of some incredible progress possible with this technology over the next few decades, like real, mass production practical aneutronic fusion or similar civilization-defining advances. I don't want to see this technology misused to lend legitimacy to despotic rule by making oracular pronouncements that are uncritically accepted because early verifier-based pronouncements turned out true in unmistakable ways.
The technology is one for the history books, but it has this pernicious effect of highjacking the reasoning centers of many people through their linguistic centers that the Neuro Linguistic Programming crowd could only dream of actually accomplishing.
Producing beautiful, coherent, maintainable code. RL trained the models to produce correct code but boy oh boy is it ugly if it's 10k+ LOC
This is a brillaint way to simulate reasoning, but its likely not how we reason ... simply how we store reasoning in writing.
Its useful if you know how to use it, its dangerous if you think its more than that.
But tl;dr it can (since its uncompressing our lingusticially stored reasoning from books) arrive at reasoning a DIFFERENT way than our brains did... and this isn't right or wrong.
Where it diverges is when it must move beyond the text or even the synthetic possible text of all vector spaces combined (aka novel territory) and it can't conjecture or test those outcomes well. But to be fair, neither can MOST humans.
Perhaps you've reified it too much.
If that's the case, AI-generated code could also operate on its own logic. Right now, programming is still done by humans, not machines, which creates a mismatch. But maybe the true machine-generated code could be much closer to the machine itself.
When you code with AI, there's a subtle mismatch with human-written code. It's like human code is a clean ORM layer, while machine code is raw SQL queries—there's that kind of subtle impedance mismatch. If we ever reach machine-to-machine code, what would that code even look like? Would it still use classes and methods?
That’s all.
The point of the story was that classifiers can be right for the wrong reasons and almost inevitably are. At least there’s zero guarantee that the reason for making the prediction matches the human or “real” reason why it’s correct.
LLMs are classifiers, there is absolutely no reason to assume they’re any different, regardless of any reasoning tokens they emit. They do what their handler wants to see, that’s all, and that’s what they’re trained to do.
People often take this as a knock against them. It isn’t, it’s just the reality of neural network classifiers. The results speak for themselves and don’t depend on whether they “actually” reason, but all evidence says they don’t, or at least there’s no special reason why they would.
You can broaden the definition of classifier to anything that internally divides its input space into regions, but that definition would include every neural network, whether biological or artificial. So it's not very meaningful, and certainly doesn't give any insight into how they differ from humans.
It’s frustrating that anyone who says maybe we shouldn’t base our entire economy on this one thing until we understand it and what’s it’s useful is essentially labeled this way.
Most models in that study appear to have "failed" by offering a Python code solution to generate the repetitive assessment steps, rather than just mindlessly copying those steps out. This was discussed at length at the time.
None of this means that models "reason", whatever that means - but simply that the Apple study was not useful evidence either way.
[1] https://www.hrw.org/report/2018/01/09/dark-side/secret-origi...
Now as far as the math stuff a quirk of that field is that it can be fully analyzed in token space. Because 2+2 is a matter of definition it doesn’t need empirical testing like biology or subjective social support like a claim about the causes of WWI
So somewhere between the fact that language encodes a lot more ‘concepts’ than we naively may realize, the power of statistical emergence via associations, and what pursuits can be fruitfully done in token space we can get a long way towards ‘intelligence’
It seems intuitive that continuing a correct output sequence is easier than the "discontinuity" of jumping from the input prompt to the output sequence.
1) LLM's written CoT might not always be faithful to the model's real reasoning process (true and important)
2) The "stochastic parrot" hypothesis, which the article reintroduces as "approximate retrieval" - ie, LLMs don't "really reason" at all, they just memorize a lossy encoding of their training data. This obviously raises the question of how LLMs can now routinely solve open mathematical problems, with no solutions in the training data by definition. The article handwaves this with:
"The model doesn’t have to learn or reliably apply a general reasoning process, Kambhampati said; it just has to absorb enough examples of what the steps look like to predictively mimic them on its way to “stitching together” a plausible result that can then be verified."
The problem is that "mimicking" training data to arrive at a "plausible" result gets you an incorrect-but-plausible-sounding "proof" of the Jacobian conjecture, which was famous for humans writing plausible-looking "proofs" that had subtle flaws. You can't disprove the conjecture through sheer luck (search space too large) or "approximate retrieval" (the only thing you'd retrieve are fake "proofs"; far more human effort went into proof than disproof) or by writing something "plausible" that just happens to be correct (Jacobian was famous for "plausible" but wrong); the model must be carrying out mathematical reasoning somehow, by any sane definition of the word, even if it isn't fully reflected in CoT. The article doesn't address this.
> Dimethyl(oxo)-lambda6-sulfa雰囲idine)methane donate a CH2rola group occurs in reaction, Practisingproduct transition vs adds this.to productmodule. Indeed"come tally said Frederick would have 10 +1 =11 carbons. So answer q Edina is11.
What’s going on here, for example? But what if this is the path of human reasoning too. You know, have you guys read Peter Thiel’s Antichrist essay? It’s very weird, man. Guy sounds off his rocker entirely.
But he’s super successful, right? Maybe world modeling doesn’t text represent well. By the antichrist maybe he means some notion of the collective voting for distribution of resources without contributing productive capacity and that that ends societies? Or maybe internal world models are just not text serializable effectively.
A thing I’ve recently been enamored of are effective world and coordination models that are not “true”. E.g. a tribe that believes the forest gets angry if they do not hunt united. Lots more like that in Darwin’s Cathedral.
It might seem a bit free association-y but the topic itself is that.
The reasoning tokens behind this comment: https://wiki.roshangeorge.dev/w/Blog/2025-10-12/Word_Magic
In one of the articles on this topic they state: > To further show that trace accuracy is only loosely connected to solution accuracy, we then train models on noisy, corrupted traces which have no relation to the specific problem each is paired with, and find that not only does performance remain largely consistent with models trained on correct data, but in some cases can improve upon it and generalize more robustly on out-of-distribution tasks
which actually maps somewhat to regularisation techniques in image processing where you might add noise to an image or drop data to make the model more robust to changes.
LLMs are one-trick pony's - they use the past to predict the future (presumed to be the same as what they were trained on). i.e. they are trained as auto-regressive predictors.
LLMs learn two slightly different types of reasoning via two different types of training.
1) SFT, or even base model training, on data that contains reasoning traces, learnt via next token error feedback. This does not result in "stochastic parroting" in the naive/pejorative sense, but nonetheless is very context dependent, even if the usual generative multi-source mashups apply.
2) RLVR post-training, where the model learns to mimic long-horizon (not just next token) reasoning via boosting a sequence of next-token predictions that steer the output towards a verified reasoning step (i.e. one that was at least valid in the context of the RL training sample). As Karpathy has noted, this is a pretty crude mechanism since you reinforce everything - errors included - that lead to the verified outcome.
RLVR is more powerful than SFT, and can result in more generalizable reasoning, since it is operating at a higher level of entire long-horizon reasoning steps, and also critically because it is most successfully being applied in the domains of math and coding which are highly self-consistent and logical. A reasoning step that was valid in one context should be equally valid in another context as long as you have successfully learnt what that generalized context is. Therefore, in these domains, you can chain together sequences of individually learnt reasoning steps, and hopefully this "novel" assembled reasoning chain is valid as a whole.
So, what is still missing from LLM reasoning compared to human reasoning? No doubt humans reason by memory a lot of the time too, and reductive axiomatic math reasoning works just as well for humans as when automated. So, what's missing?
There seem to be two major things.
1) RLVR requires rewards, and how well it works is going to depend on how accurate those rewards are. Is this reasoning step actually valid, or does it just kinda look ok? When moving beyond the cold reductionist logic of math and coding, the notion of correctness is far weaker, and it seems the best you can do is train on human curated reasoning rubrics and LLM-as-judge, which is much more fallible, leaving the model really needing (but lacking) a fallback to more general reasoning, not just memorized "maybe correct" reasoning steps.
2) Whether for reasoning outside of math & coding, or even within these domains when hoping for super-human innovative reasoning, not just lego-assembly proofs, what LLMs are lacking is a mechanism for what to do when next token/next step prediction fails. What LLMs currently do is "hallucinate", not even recognizing the failure.
In the human brain 50% or more of our cortex is feedback paths and the machinery that (perhaps together with the archaic part of our brain) lets us recognize and respond to failed predictions in an adaptive manner. This starts with continual learning (prediction failure being the signal), but also includes critical innate traits such as curiosity, boredom and frustration, that provide impasse resolution by encouraging us to explore unknown environments/contexts, abandon exploration when it is not productive, and generally expose ourselves to learning situations.
The dream is for AI scientists making new discoveries - the AI that could have invented general relativity if it has lived in Einstein's time, but this is not going to happen until their reasoning stops being purely predictive and becomes creative as well - curious about their own knowledge gaps and pursuing them in directed fashion, etc.
The current crop of Erdos solutions etc, while useful, really just represent the "generative closure" of what can be done/discovered WITHOUT learning anything fundamentally new. These will no doubt continue for a while unti...
It could be boiled down to: in 2025 this paper showed that "thought traces" in the models of the time could sometimes be inaccurate or misleading. Today they still might be, although OpenAI says actually they are accurate for their modern models, (based on internal research, rather than published research).
[1] https://arxiv.org/abs/2506.06941
I won't get into the metaphysics of reasoning, however, the Sonnet is using an OODA loop. The difference which hasn't been gapped is that human reason and imagination (in the sense of Mr. Rogers' Neighborhood) can predict the consequences of the actions we take.
This ability to loop is much, much wider in Opus 5 than Opus 4.. I had to strain to get Opus 4. to do the wider OODA loop but Opus 5 does it out of the box. I needed to throw out all existing instructions, skills, guidance, moving from 4-* to 5.
[0] https://adamsohn.com/lambda-variance/