ClaimsPeople

Yann LeCun

Yann LeCun is a computer scientist who shared the 2018 ACM Turing Award, is a Jacob T. Schwartz Professor at NYU's Courant Institute, was Meta's Chief AI Scientist and founding director of FAIR, and is Executive Chairman of AMI Labs.

1,175 claims collected, 197 researched, 84 resolved. Report of 4 November 2026, method 0.1.0+c876030c.

Scores by kind of claim

Each kind of claim is scored on its own, out of 100: the share that held up, with the range it likely falls in. Fewer resolved claims give a wider range.

  • Factual claims

    Statements about the present or the past, checked against the record: did they hold up?

    78 72 83

    High confidence, 79 factual claims resolved. From 16 documents, 2006 to 2026; 19 documents sampled from 33 candidates found.

    Not scored: 43 unresolvable.

  • Predictions

    Statements about events outside the person's control, checked once they fell due: did they come true?

    64 43 83

    Low confidence, 5 predictions resolved. From 4 documents, 2010 to 2024; 19 documents sampled from 33 candidates found.

    Not scored: 66 unresolvable. 3 more fall due later.

    Does not yet allow for how likely each prediction was when made.

  • Promises

    Commitments to the person's or organisation's own future action, with something to deliver: was it delivered?

    Insufficient record: 0 promises resolved, 5 needed for a score.

    Not scored: 1 unresolvable.

All types combined: 77 (likely 71 to 82, high confidence), from 84 claims: 79 factual claims, 5 predictions. It mixes kinds of claim in different proportions for different people, so compare people kind by kind.

Yann LeCun's claims

55 claims

55 claims, sorted by newest said. Choose a claim to read it in full.
VerdictKindSaidClaimTopic
Mostly true Factual 29 Mar 2024 Yann LeCun gave a lecture hosted by the Center of Mathematical Sciences and Applications at Harvard on 2024-03-28 to a full house. Factual · 29 Mar 2024 Their own work or organisation
Mostly true Factual 18 Mar 2024 People can still think even if their language abilities are destroyed by a stroke. Factual · 18 Mar 2024 Philosophy, ethics and society
Mostly true Factual 18 Mar 2024 Transformers are equivariant to permutations, as ConvNets are equivariant to translations. Factual · 18 Mar 2024 AI capabilities
Mostly true Factual 14 Mar 2024 There are currently no AI systems that can teach themselves to drive a car in about 20 hours of practice, as a 17-year-old can. Factual · 14 Mar 2024 AI capabilities
Mostly true Factual 10 Mar 2024 The white-noise-sounding flat-spectrum signal is the pseudo-random bit sequence used to pre-train an adaptive equalizer that cleans up an analog signal into clean bits. Factual · 10 Mar 2024 Technology
Mostly true Factual 8 Mar 2024 Training an encoder to reconstruct images does not lead it to produce good generic image features, though the encoder architecture is good. Factual · 8 Mar 2024 AI capabilities
Mostly true Prediction 8 Mar 2024 Search engines will increasingly be replaced by dialogue systems that answer questions and perhaps point to appropriate references. Prediction · 8 Mar 2024 AI capabilities
Mostly true Factual 8 Mar 2024 Rocket trajectories have been planned using model-based optimization of command sequences since computers existed, essentially since the early 1960s. Factual · 8 Mar 2024 Other
Mostly true Factual 8 Mar 2024 Current LLMs can do none of understanding the physical world, persistent memory, reasoning and planning, or only in a very primitive way. Factual · 8 Mar 2024 AI capabilities
Mostly true Factual 8 Mar 2024 LLMs, including BERT-style bidirectional ones, exploit self-supervised learning, and LeCun has been a very strong advocate of self-supervised learning for many years. Factual · 8 Mar 2024 AI capabilities
Mostly true Factual 8 Mar 2024 Training a base LLM is currently very expensive and difficult, so only a few companies can do it properly. Factual · 8 Mar 2024 Economics and markets
Mostly true Factual 28 Jun 2023 The US is a clear outlier among developed countries in healthcare cost and outcomes. Factual · 28 Jun 2023 Other
Mostly true Factual 28 Jun 2023 The audience YES/NO split at the Munk Debate was 67/33 before the debate and 63/37 after. Factual · 28 Jun 2023 AI safety and alignment
Mostly true Factual 28 Jun 2023 In the official result of the Munk Debate on AI existential threat, the NO side (Mitchell and LeCun) won by 4% over the YES side (Tegmark and Bengio). Factual · 28 Jun 2023 AI safety and alignment
Mostly true Factual Jun 2023 Masked autoencoders are less successful with images and other continuous domains than with discrete domains such as language. Factual · Jun 2023 AI capabilities
Mostly true Factual Jun 2023 Probabilistic models require normalization and may become intractable for high-dimensional data. Factual · Jun 2023 AI capabilities
Mostly true Factual Jun 2023 With probabilistic models, training is confined to loss functions derived from the negative log-likelihood, such as cross entropy. Factual · Jun 2023 AI capabilities
Mostly true Factual Jun 2023 All modules of the proposed architecture are fully differentiable. Factual · Jun 2023 AI capabilities
Mostly true Factual 4 Mar 2021 Facebook has deployed XLM-R, a model based on its RoBERTa architecture, to improve hate speech classifiers in multiple languages across Facebook and Instagram. Factual · 4 Mar 2021 AI capabilities
Mostly true Factual 4 Mar 2021 LeCun said Facebook AI had created and open sourced SEER, a billion-parameter self-supervised computer vision model that he described as proven to work efficiently with complex, high-dimensional image data. Factual · 4 Mar 2021 AI capabilities
Mostly true Factual 4 Mar 2021 As of March 2021, self-supervised learning had not yet brought computer vision the same improvements it has brought to natural language processing. Factual · 4 Mar 2021 AI capabilities
Mostly true Factual 4 Mar 2021 The Barlow Twins method tries to minimize redundancy between the individual components of the embedding vectors. Factual · 4 Mar 2021 AI capabilities
Mostly true Factual 2021 A cat has fewer than 1 billion neurons, while a dog has 2 billion. Factual · 2021 Other
Mostly true Factual 2021 Since the 1990s, algorithms exist that learn oriented edge detectors fully unsupervised with the equivalent of a few minutes of real time. Factual · 2021 AI capabilities
Mostly true Factual 2021 Self-supervised training of visual representations from video has not yet been successful. Factual · 2021 AI capabilities
Mostly true Factual 2021 In reinforcement learning the machine receives only a single scalar feedback, and only occasionally. Factual · 2021 AI capabilities
Mostly true Factual 2021 The amount of image training data available for self-supervised learning is essentially unlimited, with billions of new images per day. Factual · 2021 AI compute, hardware and data
Mostly true Factual 28 Jun 2019 Jumping spiders can count to three, with categories for one prey, two preys and many preys. Factual · 28 Jun 2019 Other
Mostly true Factual 25 Jun 2019 Nick Clegg, Facebook's VP of global affairs and communications, wrote an op-ed about the future of the Internet. Factual · 25 Jun 2019 AI policy and governance
Mostly true Factual 3 Jun 2019 The Beijing Academy of Artificial Intelligence's published AI ethics guidelines mention the protection of individual privacy. Factual · 3 Jun 2019 AI policy and governance
Mostly true Factual 2019 The ConvNet architecture was loosely inspired by the ventral pathway in the visual cortex. Factual · 2019 AI capabilities
Mostly true Factual 2019 ReLU nonlinearities and dropout regularization enabled purely supervised backprop training of very deep networks with a dozen layers. Factual · 2019 AI capabilities
Mostly true Prediction 2019 Usage of graph neural networks will increase for a wide variety of applications likely to violate the assumptions of current deep learning hardware. Prediction · 2019 AI compute, hardware and data
Mostly true Prediction 2019 As DL is used for high-level cognitive tasks, memory modules will become commonplace and very large, requiring hardware support. Prediction · 2019 AI compute, hardware and data
Mostly true Factual 28 Jul 2016 The check reading system LeCun and a colleague built at Bell Labs in the early 1990s used structured prediction on top of convolutional nets, called Graph Transformer Networks. Factual · 28 Jul 2016 AI capabilities
Mostly true Factual 27 May 2015 Deep learning uses the backpropagation algorithm to indicate how a machine should change the internal parameters that compute each layer's representation from the previous layer's. Factual · 27 May 2015 AI capabilities
Mostly true Factual 27 May 2015 The acknowledgements, dated 2015, state that the authors Y.L. and Y.B. are fellows of the Canadian Institute For Advanced Research (CIFAR). Factual · 27 May 2015 Field building and careers
Mostly true Factual 18 Feb 2015 Deep Learning systems augmented with a memory module are starting to show impressive results in natural language processing. Factual · 18 Feb 2015 AI capabilities
Mostly true Factual 18 Feb 2015 The brain is much better than current artificial models at unsupervised learning. Factual · 18 Feb 2015 AI capabilities
Mostly true Factual 18 Feb 2015 Shallow learning systems have one or two layers, while deep learning systems typically have five to 20 layers. Factual · 18 Feb 2015 AI capabilities
Mostly true Factual 18 Feb 2015 The system would probably not be more accurate, but it can potentially recognize people among hundreds of millions. Factual · 18 Feb 2015 AI capabilities
Mostly true Factual 18 Feb 2015 A typical Deep Learning system is like a box with 500 million knobs, 1,000 light bulbs and 10 million training images. Factual · 18 Feb 2015 AI compute, hardware and data
Mostly true Prediction 18 Feb 2015 Natural language understanding will be the next frontier for Deep Learning. Prediction · 18 Feb 2015 AI timelines and forecasting
Mostly true Factual 18 Feb 2015 A trained classifier may also correctly classify cars and dogs it has never seen before. Factual · 18 Feb 2015 AI capabilities
Mostly true Factual 15 May 2014 The notions of partition functions, entropy, free energy and variational methods prevalent in the graphical models literature come from statistical physics. Factual · 15 May 2014 Science and research methods
Mostly true Factual 15 May 2014 The first commercially viable check reading system, deployed by AT&T/NCR in 1996, used a ConvNet rather than an SVM. Factual · 15 May 2014 AI capabilities
Mostly true Factual 15 May 2014 No one uses generalization bounds to do model selection; practitioners use cross-validation. Factual · 15 May 2014 Science and research methods
Mostly true Factual 15 May 2014 Speech is a domain with access to ridiculously large amounts of data and a very large number of categories. Factual · 15 May 2014 AI capabilities
Mostly true Factual 15 May 2014 Facebook AI Research uses Torch7 for many projects, as do Deep Mind and several groups at Google. Factual · 15 May 2014 Technology
Mostly true Prediction 2010 In the next decade, a major challenge of computer vision, machine learning and AI will be devising methods that learn good feature hierarchies from unlabeled and labeled data in an integrated fashion. Prediction · 2010 AI capabilities
Mostly true Factual 2010 A face detection application on the CNP system was demonstrated achieving 10 fps on VGA images. Factual · 2010 AI compute, hardware and data
Mostly true Factual 2010 Convolutional Networks are a biologically inspired trainable architecture that can learn invariant features. Factual · 2010 AI capabilities
Mostly true Factual 2006 The hinge loss with G_W(X)=W^T X and a ||W||^2 regularizer gives the linear support vector machine. Factual · 2006 AI capabilities
Mostly true Factual 2006 The hinge loss systematically chooses the most offending incorrect answer as the one whose energy is pulled up. Factual · 2006 AI capabilities
Mostly true Factual 2006 Support vector Markov models, max-margin Markov networks and conditional random fields are all sequence modeling systems that use linearly parameterized energy factors. Factual · 2006 AI capabilities
Back to the list

Yann LeCun

Open its page

Factual AI capabilities

“LLMs can do none of those or they can only do them in a very primitive way and they don’t really understand the physical world.”

Yann LeCun, Meta chief AI scientist, interviewed 8 March 2024 Transcript for Yann Lecun: Meta AI, Open Source, Limits of LLMs, AGI & the Future of AI (Lex Fridman Podcast)

Current LLMs can do none of understanding the physical world, persistent memory, reasoning and planning, or only in a very primitive way.

Mostly true True in substance, wrong in a detail.

The second judge would not resolve this claim; the verdict shown is the first judge's.

What would count as true or false

Judge as of 2024-03-08 against the state of large language models at that time. The four capabilities are understanding the physical world, persistent memory, reasoning and planning. True: benchmarks, evaluations and expert assessments from around that date show that the leading LLMs (for example GPT-4, Claude 3, Gemini) either lacked each of the four capabilities or showed only rudimentary, unreliable versions of them, and that they did not really understand the physical world. Mostly true: this holds for most of the four, but for one capability (for example, some forms of reasoning) the evidence shows clearly more than primitive ability. Mixed: about half of the capabilities show substantial, non-primitive performance. Mostly false: most of the capabilities were shown at a substantial level, and only one or two were primitive or absent. False: the evidence shows that LLMs then performed all four capabilities at a clearly non-primitive level. Contested expert opinion alone does not settle the claim; weigh documented evaluation results.

Evidence

  1. On the Planning Abilities of Large Language Models : A Critical Investigation arXiv (Valmeekam, Marquez, Sreedharan, Kambhampati)
    “Our findings reveal that LLMs’ ability to generate executable plans autonomously is rather limited, with the best model (GPT-4) having an average success rate of \(\sim\)12% across the domains.”
  2. PlanBench: An Extensible Benchmark for Evaluating Large Language Models on Planning and Reasoning about Change NeurIPS Datasets and Benchmarks track (proceedings.neurips.cc)
    “Overall, the performance of these LLMs on our benchmark shows that, as of right now, LLMs are pretty ineffective in reasoning about actions and change.”
  3. GRASP: A Novel Benchmark for Evaluating Language GRounding and Situated Physics Understanding in Multimodal Language Models IJCAI-24 (International Joint Conference on Artificial Intelligence) · 2024
    “At the same time, all models perform below or at the chance level of 50% in the Intuitive Physics tests, while human subjects are on average 80% correct.”
  4. GRASP: A Novel Benchmark for Evaluating Language GRounding and Situated Physics Understanding in Multimodal Language Models IJCAI-24 (International Joint Conference on Artificial Intelligence) · 2024
    “In comparison, GPT-4 achieves an accuracy of 75.0% on this task.”

Reasoning

Independent evaluations from around the claim date show limited planning: GPT-4 produced executable plans in only about 12% of autonomous instances (E1), and PlanBench found LLMs 'pretty ineffective in reasoning about actions and change' (E2). On GRASP, multimodal LLMs scored at or below chance on intuitive physics, against about 80% for humans (E3), which supports the claim about the physical world. The evidence does not address persistent memory, and E4's 75% GPT-4 figure has no stated task, so the claim is supported for most but not all of the four capabilities.

Cite this claim

Yann LeCun, “Current LLMs can do none of understanding the physical world, persistent memory, reasoning and planning, or only in a very primitive way.” (said 8 March 2024). Verdict: mostly true. Reliability API, claim c_861fee986cde, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_861fee986cde

Claims without a verdict

Is a verdict here wrong, or do you speak for Yann LeCun? Reply on the claim's own page, with the form at its foot (how replies work). Verdicts are made automatically; how verdicts and scores are made.