What this covers

Yann LeCun sits with Lex Fridman to lay out a critique of current large language models and a blueprint for what comes next. The conversation centers on a fundamental limitation: autoregressive LLMs achieve fluency in language without possessing the deeper capacities that constitute human intelligence — understanding the physical world, persistent memory, reasoning, planning. LeCun argues this is not a gap that scale alone will close, but rather a structural problem requiring a different architecture. He proposes joint-embedding predictive architectures as a path forward, systems that learn abstract world models by predicting representations rather than pixels. The conversation also branches into governance: LeCun warns that proprietary concentration of AI systems poses a greater threat than other technical risks, and that open-source development is essential to preserve diversity of thought and democratic resilience.

The discussion moves across several territories. LeCun characterizes intelligence not as a linear metric but as a high-dimensional collection of skills, and invokes Moravec's Paradox to explain why tasks humans find trivial remain difficult for machines. He draws on evidence that children acquire core knowledge — gravity, animacy, stability — through visual observation alone, suggesting self-supervised learning on embodied perception data holds more promise than text alone. On safety, he departs from "doomer" framings, comparing AI control to iterative engineering of turbojets rather than solving a single mathematical principle. He addresses bias not as a technical problem to eliminate but as inevitable — different groups will always perceive systems differently, so the solution mirrors journalism and science: diversity and openness. Throughout, he contests the assumption that intelligence automatically breeds domination, and argues that breakthroughs in AI disseminate too rapidly for any single actor to maintain decisive advantage.

Sharpest takeaway

LeCun argues that autoregressive LLMs are a fundamentally limited path to human-level intelligence because they lack grounded world models, and that the future requires non-generative joint-embedding architectures (JEPA) plus open-source development to prevent dangerous concentration of AI power.

  • LLMs cannot understand the physical world, have persistent memory, reason, or plan, and learn from far less rich data than embodied perception provides
  • Joint-embedding predictive architectures that predict abstract representations (not pixels) can learn world models that enable planning
  • Open source AI is the only way to ensure diversity and prevent a few companies from controlling humanity's information diet

The argument · threads12 threads · 49 claims
0.85

Current LLMs are linguistic pattern-matchers, not reasoning systems; they lack world models, persistent memory, planning, and grounded physical understanding.

7 pointscentrality 5/5
  • A world model predicts the state of the world at time T+1 given the current state and a candidate action, which enables planning by allowing a system to imagine the outcomes of action sequences and select those that satisfy an objective — something LLMs cannot do.

    “Here is my idea of the state of the world at time T. Here is an action I’m taking. Here is a prediction of the state of the world at time T plus one... If you have a model of this type, you can use it for planning.”

  • Cats and dogs understand the physical world far better than any LLM and can plan complex actions despite having no language, demonstrating that grounded world understanding is achievable and necessary independent of language, and should be solved before combining with language.

    “we’re not going to get to the level of even the intelligence or level of understanding of the world of a cat or a dog, which doesn’t have language. They don’t have language and they understand the world much better than any LLM.”

  • The space of all possible prompts is so gigantic that fine-tuning can only cover a tiny subset; there is a long tail of prompts the system was never trained on, so an LLM can always be broken into producing nonsense, making it effectively a giant lookup table rather than a reasoning system.

    “within the set of all possible prompts, the proportion of prompts that have been used for training is absolutely tiny, it’s a tiny, tiny, tiny subset of all possible prompts”

  • LLMs can answer planning questions only for scenarios similar to their training data, essentially regurgitating templates they were trained on; they cannot plan for situations never encountered before, and cannot represent the high-bandwidth physical experience needed for low-level actions like standing up from a chair.

    “They’re not going to be able to plan for situations where that they never encountered before. They basically are going to have to regurgitate the template that they’ve been trained on.”

  • Autoregressive LLMs cannot achieve human-level intelligence because they lack or only primitively possess four essential characteristics of intelligent behavior: understanding the physical world, persistent memory, reasoning, and planning.

    “the capacity to understand the world, understand the physical world, the ability to remember and retrieve things, persistent memory, the ability to reason, and the ability to plan. Those are four essential characteristics of intelligent systems”

  • We are fooled by LLMs' linguistic fluency into assuming they possess all the characteristics of human intelligence, but that impression is false; fluency in manipulating language does not imply general intelligence.

    “We just assume that if a system is fluent in manipulating language, then it has all the characteristics of human intelligence, but that impression is false. We’re really fooled by it.”

  • A four-year-old child receives roughly 10^15 bytes of visual information through the optic nerve in 16,000 waking hours, vastly more than the ~2×10^13 bytes of all public internet text used to train LLMs, suggesting text alone is insufficient data to build a world model.

    “a four-year-old has been awake for 16,000 hours in his or her life, and the amount of information that has reached the visual cortex of that child in four years is about 10 to 15 bytes”

0.72

AI doomers lack faith in human goodness and institutions; optimists trust people will use AI to amplify intelligence and capability for widespread benefit.

4 pointscentrality 5/5
  • Many AI doomers are doomers precisely because they don't believe people are fundamentally good and don't trust people or institutions to do the right thing, whereas a belief that people are fundamentally good underlies optimism that empowering them with AI amplifies goodness.

    “a lot of doomers are doomers because they don’t think that people are fundamentally good, and they either don’t trust people or they don’t trust the institution to do the right thing”

  • AI will amplify human intelligence, giving everyone a staff of smart AI assistants, which is intrinsically good because intelligence is the most in-demand commodity and most of humanity's mistakes stem from lack of intelligence or knowledge — analogous to how the printing press made everyone smarter by democratizing access to books.

    “We can make humanity smarter with AI. AI basically will amplify human intelligence. It’s as if every one of us will have a staff of smart AI assistants.”

  • The assumption that an intelligent AI system will necessarily want to take over is false; the link between intelligence and the desire to dominate observed in some species does not transfer to AI, and there are strong incentives to build AI systems that are submissive to humans with guardrails encoded in their objectives.

    “there is another fallacy, which is the fact that because the system is intelligent, it necessarily wants to take over.”

  • AI doomer catastrophe scenarios rely on mostly false assumptions: that superintelligence will arrive as a single event, that an intelligent system necessarily wants to dominate, and that we cannot iteratively correct mistakes; in reality intelligence and the desire to dominate are unrelated, and we can build many controllable AIs to police rogue ones.

    “AI doomers imagine all kinds of catastrophe scenarios of how AI could escape or control and basically kill us all, and that relies on a whole bunch of assumptions that are mostly false.”

0.72

Proprietary AI concentration threatens democratic diversity and information freedom far more than technical AI risks.

4 pointscentrality 5/5
  • The inevitable direction of history is that the vast majority of AI systems will be built on top of open source platforms, because only open source enables an ecosystem of companies and groups to fine-tune specialized systems that are diverse in language, culture, and value systems for vertical applications.

    “the direction of inevitable direction of history is that the vast majority of AI systems will be built on top of open source platforms”

  • The solution to AI bias is the same as the solution found in liberal democracy for the press: AI systems must be free and diverse so that our information does not come from a single source, because a unique source of information is opposite to democracy, like science which progresses through disagreement and consensus formation.

    “the answer is the same answer that we found in liberal democracy about the press, the press needs to be free and diverse. We have free speech for a good reason, is because we don’t want all of our information to come from a unique source”

  • The concentration of power through proprietary AI systems is a much bigger danger than other AI risks, because a future where all our information diet is controlled by a small number of companies with proprietary systems threatens the diversity of ideas and the preservation of democracy.

    “I see the danger of this concentration of power through proprietary AI systems as a much bigger danger than everything else.”

  • Meta can open source its foundation models without impairing revenue because its business model derives revenue through services, ads, and business customers given its huge existing user base; open sourcing also accelerates progress as thousands of people improve the models and build applications Meta can buy or benefit from.

    “if you have a big enough potential customer base and you need to build that system anyway for them, it doesn’t hurt you to actually distribute it to the open source”

0.67

World models trained via joint-embedding predictive architectures learn by predicting abstract representations, not reconstructing pixels.

6 pointscentrality 4/5
  • Building world models by training generative models to predict future video frames fails because you cannot represent probability distributions over high-dimensional continuous spaces (all possible frames) the way you can over a finite dictionary of text tokens; the world is too rich in unpredictable detail.

    “what you would have to do is predict the distribution of all possible frames in a video, and we don’t really know how to do that properly. We do not know how to represent distributions over high-dimensional, continuous spaces in ways that are useful.”

  • V-JEPA, applied to video by masking a temporal tube across 16 frames and predicting its representation, learns representations good enough that a supervised head can classify actions accurately, and preliminary results suggest it can detect whether a video is physically possible or impossible based on object appearance/disappearance.

    “the representation allow our system to tell whether the video is physically possible or completely impossible, because some object disappeared or an object suddenly jumped from one location to another”

  • Self-supervised prediction algorithms learn more structure when input data is more redundant; perceptual/sensory inputs like vision have far more redundancy than text (which is already compressed and less redundant), so self-supervision works better on perception than on language.

    “those self-supervised algorithm that learn by prediction, even in representation space, they learn more concept if the input data you feed them is more redundant. The more redundancy there is in the data, the more they’re able to capture some internal structure of it.”

  • Training a system to reconstruct images from corrupted versions (denoising autoencoders, masked autoencoders) does not produce good generic image features; supervised training with labels yields much better representations, proving reconstruction-based self-supervision is the wrong approach for vision.

    “the thing that does not work is training the system to learn representations of images by training it to reconstruct a good image from a corrupted version of it”

  • JEPA architectures can be trained with non-contrastive methods (distillation-based like BYOL and I-JEPA, or regularized like VICReg) rather than contrastive methods, avoiding the need for a gigantic number of negative contrastive samples when the answer space is large.

    “the techniques we’re using are non-contrastive. So not only is the architecture non-generative, the learning procedures we are using are non-contrastive.”

  • Joint-Embedding Predictive Architecture (JEPA) trains by encoding both a full input and a corrupted version, then predicting the representation of the full input from the corrupted one in abstract representation space, rather than reconstructing pixels; this lets the encoder discard unpredictable noise and preserve only predictable, modelable structure.

    “In a JEPA, you’re not trying to predict all the pixels, you’re only trying to predict an abstract representation of the inputs. And that’s much easier in many ways.”

0.63

True human-level AI requires gradually solving component problems: world models, memory, hierarchical planning, reasoning over a decade or more.

5 pointscentrality 4/5
  • Achieving ubiquitous advanced machine intelligence will require hardware innovation because current compute is far from matching the human brain — a GPU uses half a kilowatt to a kilowatt versus the brain's ~25 watts, and you would need 100,000 to a million GPUs to match brain power, so power efficiency must improve dramatically.

    “A GPU today is half a kilowatt to a kilowatt. Human brain is about 25 watts, and a GPU is way below the power of the human brain. You need something like 100,000 or a million to match it”

  • Human-level AI (AMI/AGI) will not be a single event where someone discovers a secret and turns on a machine; it will be gradual progress over at least a decade and probably much more, building up component capabilities (world models, memory, reasoning, hierarchical planning) that must be made to work together, with unforeseen problems likely.

    “it’s not going to be an event. The idea somehow... that somebody is going to discover the secret to AGI... and then turn on a machine and then we have AGI, that’s just not going to happen.”

  • The robotics industry has failed to emerge for 10-20 years beyond pre-programmed behavior because of Moravec's paradox; significant progress and capable domestic robots require world models that let systems train themselves to understand how the world works and plan actions, which is still some distance away.

    “until we have, again, world models, systems that can train themselves to understand how the world works, we’re not going to have significant progress in robotics.”

  • Complex actions require hierarchical planning that decomposes high-level goals into progressively lower-level subgoals down to muscle control, but nobody knows how to train an AI system to learn the appropriate multiple levels of abstract representation needed for hierarchical planning.

    “nobody really knows how to do this in AI. Nobody knows how to train a system to learn the appropriate multiple levels of representation so that hierarchical planning wor”

  • Babies learn core knowledge — gravity, inertia, stability, the animate/inanimate distinction, and why people want things — mostly by observation in the first months and years of life, before language appears around age one, and this observationally-acquired knowledge is exactly what current AI systems lack.

    “by the time you are nine years old, you’ve learned about gravity, you know about inertia, you know about gravity, the stability... there’s a lot of things that you learn mostly by observation, really not even through interaction.”

0.53

LLMs hallucinate because autoregressive token-by-token generation compounds error probability; gradient descent on objectives is far more efficient.

6 pointscentrality 3/5
  • LLM hallucination arises because in autoregressive prediction each token has a non-zero probability of taking the output outside the set of correct answers, and assuming these errors are roughly independent, the probability of remaining correct decreases exponentially with the number of tokens produced.

    “every time you produce a token, the probability that you stay within the set of correct answer decreases and it decreases exponentially”

  • The transformational effect of RLHF on LLMs comes from the human feedback itself, much of which is actually supervised learning rather than reinforcement learning; reinforcement learning should be minimized because it is incredibly sample-inefficient.

    “What’s had the transformational effect is human feedback, there is many ways to use it, and some of it is just purely supervised, actually, it’s not really reinforcement learning.”

  • The blueprint for future AI systems is to plan answers by optimizing an objective (energy) function over an abstract representation space via gradient descent, then decode that optimized abstract thought into language — making reasoning independent of the output language and Turing-complete, unlike one-token-at-a-time LLMs.

    “they will think about their answer, plan their answer by optimization before turning it into text, and that is turning complete”

  • LLM reasoning is primitive because the amount of computation spent per token produced is constant — proportional to network size times number of tokens — regardless of whether the question is simple, complex, or impossible; humans, by contrast, spend more time and computation on harder problems.

    “the amount of computation that is spent per token produced is constant... it doesn’t matter if the question being asked is simple to answer, complicated to answer, impossible to answer”

  • Optimizing an answer in continuous representation space via gradient descent is far more efficient than the discrete-space approach of generating many candidate token sequences and selecting the best, which requires running the LLM for every possible sequence and is incredibly computationally wasteful.

    “it’s much better to do an optimization in continuous space where you can do gradient and descent as opposed to generate tons of things and then select the best, you just iteratively refine your answer to go towards the best, that’s much more efficient”

  • An energy-based model is a function with a scalar output that takes two inputs X and Y and outputs zero if Y is compatible with X and a positive number if it is not, providing an alternative to probabilistic models for measuring whether an answer fits a prompt.

    “Energy based model is a function with a scaler output, just a number, you give it two inputs, X and Y, and it tells you whether Y is compatible with X or not.”

0.50

AI progress is continuous and disseminates rapidly; unlike nuclear weapons, no single breakthrough gives permanent advantage to first movers.

2 pointscentrality 4/5
  • Unlike nuclear weapons, AI will not be a single decisive breakthrough that matters who gets it first, because progress is continuous and any breakthrough disseminates extremely quickly through industry — even AlphaGo was reproduced within three months — and government/military are far behind industry.

    “it’s not going to be an event. It’s going to be continuous progress. And whenever one breakthrough occurs, it’s going to be widely disseminated really quickly.”

  • The success of the US tech industry and Silicon Valley comes from the very rapid circulation and dissemination of information — knowing something is possible motivates investment, and talent moving between companies spreads knowledge — keeping the whole region ahead.

    “What makes the success of the US tech industry and Silicon Valley in particular is exactly that, is because the information circulates really, really quickly and disseminates very quickly. And so the whole region is ahead because of that circulation of information.”

0.50

AI doomer predictions follow a recurring historical pattern of fearing new technologies without evidence they cause catastrophe.

2 pointscentrality 4/5
  • There is a recurring, knee-jerk human pattern of fearing new technology and cultural phenomena and blaming them for society's problems — electricity, trains, jazz, comic books were all predicted to cause catastrophe — and the AI doomer reaction follows this same pattern documented in the Pessimist Archive.

    “any technological revolution or cultural phenomenon was always accompanied by groups or reaction in the media that basically attributed all the current problems of society to that particular change.”

  • The AI community decided many years ago that the Turing test is a really bad test of intelligence.

    “This is what the AI community has decided many years ago that the Turing test was a really bad test of intelligence.”

0.50

AI safety emerges through iterative design producing better, more useful, more controllable systems, not from discovering a single principle.

2 pointscentrality 4/5
  • AI safety will be achieved like turbojet reliability — not through a single general principle or mathematical proof, but through decades of progressive, iterative design where a better, more useful, more controllable AI is intrinsically a safer one, refuting the doomer claim that a slight error means extinction.

    “How did we figure out how to make turbojet so unbelievably reliable?... it took decades to fine tune the design of those systems so that they were safe... a better turbojet is also a safer turbojet”

  • In objective-driven architectures where output results from inference satisfying an objective, guardrails can be embedded directly into that objective, allowing open source systems to ship with a minimum set of universally-agreed safety guardrails while communities add their own fine-tuned guardrails for gray areas like hate speech.

    “that objective can include guardrails, and we can put guardrails in open source systems... we can put guardrails in those systems that guarantee that there is a minimum set of guardrails that make the system non-dangerous and non-toxic”

0.45

Intelligence is multi-dimensional and contextual, not a single linear quantity, so superintelligence concepts are incoherent.

4 pointscentrality 3/5
  • Moravec's paradox observes that high-level abstract tasks like playing chess and solving integrals are easy for computers, while sensorimotor tasks humans take for granted, like driving or grasping objects, are extremely hard to automate.

    “how is it that with computers, it seems to be easy to do high-level complex tasks like playing chess and solving integrals and doing things like that, whereas the thing we take for granted that we do every day, like, I don’t know, learning to drive a car or grabbing an object, we can’t do with computers”

  • Intelligence cannot emerge without grounding in some reality (physical or simulated) because the environment is far richer than what can be expressed in language; most human knowledge derives from interaction with and mental modeling of the physical world, not from language.

    “I’m clearly in the camp that yes, intelligence cannot appear without some grounding in some reality. It doesn’t need to be physical reality. It could be simulated, but the environment is just much richer than what you can express in language.”

  • Much of human thinking occurs at an abstract level of representation that is independent of language; bilinguals think the same way regardless of language, and tasks like imagining a rotated object or planning a physical construction have nothing to do with language.

    “if I ask you to imagine what this water bottle will look like if I rotate it 90 degrees, that has nothing to do with language. And so clearly, there is a more abstract level of representation in which we do most of our thinking”

  • Intelligence is not a linear quantity measurable by a single number; it is a collection of skills plus the ability to acquire new skills efficiently, occupying a high-dimensional space, so you cannot meaningfully compare whether one entity is more intelligent than another across different skill sets (e.g., humans vs orangutans).

    “intelligence is a collection of skills and an ability to acquire new skills efficiently... Because it’s a multidimensional thing, the set of skills is a high dimensional space, you can’t measure, you cannot compare two things as to whether one is more intelligent than the other.”

0.31

System bias stems from customer acceptability constraints and diversity gaps, not designer ideology; no unbiased system is possible because bias is in the eye of the beholder.

4 pointscentrality 2/5
  • The ideological lean in systems like Gemini stems not from the political leanings of the designers but from the acceptability constraints of the customer base — a big company cannot afford to offend too many people, so it makes products 'safe', and overdoing de-biasing produces non-factual outputs like Black Nazi soldiers.

    “It has to do with the acceptability or political leanings of their customer base or audience. So a big company cannot afford to offend too many people”

  • Meta has built speech-to-speech translation systems that work directly through an internal representation of discrete speech units (Textless NLP) without going through text, enabling translation even for languages that have no written form.

    “We don’t go through text, it goes directly from speech to speech using an internal representation of speech units that are discrete, but it’s called Textless NLP.”

  • LLMs' success is an impressive demonstration that self-supervised learning works — taking text, corrupting it, and training a neural net to reconstruct the missing parts has produced systems that understand language, translate hundreds of languages in any direction, summarize, and answer questions.

    “those things are a incredibly impressive demonstration that self supervised running actually works”

  • It is impossible to produce an AI system that is unbiased, not for technological reasons but because bias is in the eye of the beholder — different people have different ideas of what constitutes bias, so any system will be perceived as biased by some group no matter how it is tuned.

    “is it possible to produce an AI system that is not biased? And the answer is, absolutely not... it’s because bias is in the eye of the beholder.”

0.22

LLMs pose minimal bioweapon or chemical weapon risks beyond existing search engines because real-world lab expertise is the hard part.

2 pointscentrality 2/5
  • A persuasive AI weapon (e.g., built by a hostile state to manipulate voters) will not be able to reach people directly because all future digital interactions will be mediated by one's own AI assistant, which will be at least as smart and act as a spam filter, screening out manipulation before it reaches the human.

    “They’re not going to be talking to you, they’re going to be talking to your AI assistant, which is going to be as smart as theirs... it’s going to be like a spam filter.”

  • Studies suggest LLMs do not make it meaningfully easier to design or build a bioweapon or chemical weapon beyond what a search engine and library already provide, because the hard part is the real-world lab expertise to actually build it, not the list of instructions — analogous to needing to blow up many rocket engines before one works.

    “having an LLM doesn’t help you design or build a bioweapon or a chemical weapon if you already have access to a search engine and their library. So the increased information you get or the ease with which you get it doesn’t really help you.”