Yann LeCun
AI Skepticism · World Models · Beyond TransformersThe Yann LeCun reading vault — Chief AI Scientist at Meta, Turing Award winner, and the most prominent AI skeptic who works inside a major lab. Curated vault on why current LLMs aren't close to AGI, self-supervised learning, JEPA world models, and the architectural changes needed before we get to human-level AI.
About this vault
Essential reading vault for Yann LeCun — Chief AI Scientist at Meta, co-recipient of the 2018 Turing Award with Geoffrey Hinton and Yoshua Bengio, and the most articulate critic of the 'LLMs are close to AGI' narrative working inside a major AI lab. LeCun's position is distinctive: he doesn't dismiss LLMs as useless (they're clearly impressive) but he argues they are architecturally limited in ways that prevent them from reaching human-level intelligence. His core claims: (1) LLMs learn next-token prediction from text — a thin, bandwidth-limited, sample-inefficient proxy for how humans learn; (2) true intelligence requires learning world models — internal representations of how the physical and social world works — which requires learning from video, sensorimotor data, and action, not just text; (3) the next breakthrough will come from JEPA (Joint Embedding Predictive Architecture), not scaling transformers further. Reading LeCun alongside Sam Altman (who believes scaling transformers + RLHF is the path) and François Chollet (who believes ARC-AGI-style reasoning is the missing piece) gives you the three most credible competing frameworks for what AI needs next. LeCun's vault is particularly timely in 2026 because his skeptical predictions about LLM limitations have been repeatedly validated: LLMs still confabulate, still fail at simple physical reasoning, still can't plan multi-step tasks reliably. Meanwhile, JEPA-based models from Meta are producing real research advances. This vault collects his most important papers, talks, and essays on self-supervised learning, world models, and why he thinks the current AI moment is impressive but not close to AGI.
Frequently asked questions
Who is Yann LeCun?
Yann LeCun is covered in this Burn 451 vault with a focus on ai skepticism · world models · beyond transformers. The Yann LeCun reading vault — Chief AI Scientist at Meta, Turing Award winner, and the most prominent AI skeptic who works inside a major lab. Curated vault on why current LLMs aren't close to AGI, self-supervised learning, JEPA world models, and the architectural changes needed before we get to human-level AI.
How was the Yann LeCun vault curated?
The Yann LeCun vault was hand-curated by the Burn 451 editorial team from publicly available essays, blog posts, podcast transcripts, and social threads. Each piece includes an AI-generated summary so readers can triage in seconds. The vault auto-syncs as new content from Yann LeCun is published.
How many articles are in the Yann LeCun vault?
The Yann LeCun vault currently contains 22 curated pieces organized by topic, not chronology. Each article has an AI summary and a direct link to the original source. Items are refreshed hourly through Burn 451's ISR pipeline, so new publications appear within a day.
How do I use this vault with Claude or Cursor?
Install the burn-mcp-server package from npm and connect it to Claude, Cursor, or any MCP-compatible AI tool. The vault becomes queryable as live context — your AI can search, summarize, and cite articles from Yann LeCun directly in conversation without manual copy-paste or re-uploading files.
What is Burn 451?
Burn 451 is a read-later app built around a 24-hour burn timer that forces daily triage. Articles you save must be read, vaulted, or released within 24 hours. The Vault layer — including this Yann LeCun collection — holds permanent curated reading lists for AI thought leaders, founders, and researchers.
22 articles
Yann LeCun's Cake" Problem (Joint Embedding Predictive Architecture)
LeCun's 'cake' analogy reframes the entire AI research agenda: most of intelligence is learning world models (the cake), very little is learning behavior (the frosting), and reinforcement learning is just the cherry on top. This talk from 2022 lays out JEPA — Joint Embedding Predictive Architecture — as the alternative to generative models for learning world models. JEPA learns abstract representations of the world by predicting representations rather than predicting raw pixels or tokens. The implications: JEPA-based models can learn from video (without labels) in ways that generative models can't, because predicting compressed representations is easier than predicting everything. LeCun uses this framework to argue why LLMs — which are essentially generative next-token predictors — are hitting a ceiling: they can model text well but have no grounded world model. The cake problem is the most concise summary of why Meta invests in JEPA and why LeCun is skeptical that scaling LLMs alone will reach human-level AI.
A Path Towards Autonomous Machine Intelligence
LeCun's 2022 manifesto laying out a full research roadmap for achieving autonomous machine intelligence (AMI) — what he calls AGI without the science-fiction connotations. The paper is notable because it comes from inside Meta, a major AI lab, and acknowledges openly that current approaches (LLMs, RLHF, generative models) are insufficient. The core architecture LeCun proposes has six modules: (1) a world model that predicts consequences of actions; (2) an actor that proposes actions; (3) an intrinsic cost that encodes objectives; (4) a short-term memory module (similar to a key-value store); (5) a perception module; (6) a configurator that modulates all other modules. The key insight: the world model is the core unsolved problem. Everything else in modern ML (RLHF, next-token prediction) occupies the 'actor' and 'cost' modules, which are relatively tractable compared to learning an accurate world model. This paper is the intellectual foundation for Meta's JEPA research program and explains why LeCun has been publicly skeptical of LLM-based AGI timelines.
Yann LeCun on Why LLMs Can't Reason (and What Would Fix It)
LeCun's most-cited social media thread on LLM limitations: LLMs confabulate because they have no world model. They predict the next token based on pattern-matching in text, but they don't have an internal model of how the world works — so they generate plausible-sounding but factually wrong answers. His argument: humans learn world models from observing and interacting with the physical world from birth, which grounds their reasoning in causality. LLMs learn from text, which is a highly filtered, bandwidth-limited representation of human experience. You can't learn physics from reading a physics textbook — you have to interact with the world. The fix, in LeCun's view: multimodal learning (learning from video, audio, sensorimotor data) combined with architectures like JEPA that learn abstract representations rather than predicting raw pixels. This thread was widely shared in 2025 after a series of high-profile LLM failures and is the best accessible entry point for LeCun's core argument.
LeCun vs Hinton: AI Safety, Open Source, and Who Gets to Control AI
The high-profile public debate between LeCun and Geoffrey Hinton — two Turing Award winners who built the foundations of modern deep learning — crystallizes the central disagreement in AI safety in 2024-2026. Hinton, after leaving Google, became a prominent voice warning that AI could become an existential threat and that open-source AI models are dangerous. LeCun, still at Meta, disagrees on both counts. On safety: LeCun argues that fears of AI takeover are based on a category error — current LLMs aren't intelligent in the relevant sense, so fears of misaligned superintelligence are premature at best and harmful distraction at worst. The real risks are nearer-term: bias, misinformation, job displacement. On open source: LeCun believes open-source AI (like Llama) is essential for a healthy AI ecosystem because concentrated control of frontier AI is more dangerous than open access. A small group of companies deciding what AI can and can't say is far more frightening than the risks from open models. The debate illuminates why LeCun is the most important 'optimist inside a major lab' — his position isn't that AI is safe or that we shouldn't worry, but that we're worrying about the wrong things. Reading this alongside LeCun's JEPA work explains why he's confident: the architecture we'd need to build truly dangerous AI doesn't exist yet.
Self-Supervised Learning: The Dark Matter of Intelligence
LeCun's foundational essay on self-supervised learning (SSL) — the technique that powers BERT, GPT, and nearly every large language model. The essay is remarkable because it was written by one of the architects of modern deep learning before the ChatGPT moment, and it correctly predicts that SSL will become the dominant training paradigm. The core insight: supervised learning requires expensive human labels and is sample-inefficient. Reinforcement learning is notoriously unstable and requires too many trials. Self-supervised learning — where the model predicts parts of its own input from other parts — is how humans and animals actually learn: by observing the world and building internal models. LeCun uses the 'dark matter' analogy because most intelligence is implicit (learned without explicit supervision) just as most of the universe's mass is invisible. The essay predicts that SSL applied to images, video, and multimodal data will be the next major advance — a prediction that proved correct with CLIP, Flamingo, Stable Diffusion, and Meta's own work on video SSL. What the essay doesn't predict: that SSL on text (next-token prediction) would advance so rapidly that it would create a public perception that LLMs are intelligent. LeCun's subsequent work clarifies why text SSL alone is insufficient: intelligence requires grounded world models, not just statistical patterns in text.
Llama 3 and the Case for Open-Source AI (Meta AI Blog)
Meta's launch of Llama 3 represents LeCun's most concrete argument for open-source AI. Llama 3 70B and 405B matched or exceeded GPT-4 on many benchmarks while being freely downloadable, fine-tunable, and deployable by anyone. LeCun's position is that this is the correct direction for the AI ecosystem: open models allow smaller organizations, researchers, and governments to build on frontier AI without being dependent on a small number of US companies. The AI ecosystem should look like Linux, not like Windows — an open infrastructure layer that everyone builds on. LeCun has repeatedly argued that the risks of concentrated AI power (one or two companies controlling what the world's AI can say, do, and think) are greater than the risks of open models being misused. Llama's success has made this argument empirical rather than theoretical: open models are now as capable as closed models, and the predicted catastrophe from open weights has not materialized. The launch positions Meta strategically: if LLMs are going to be commoditized anyway, being the open platform that everyone builds on creates more durable value than trying to maintain a closed API moat. This is a different bet from OpenAI and Anthropic, and LeCun is its most prominent intellectual defender.
Objective-Driven AI: Beyond Next-Token Prediction
This paper formalizes LeCun's argument that autoregressive next-token prediction is the wrong objective for building intelligent systems. The core claim: predicting the next token is a generative objective, but intelligence requires a latent-variable objective — learning compressed, abstract representations of the world, not the ability to generate every possible output. The difference matters enormously in practice. Generative models must account for all sources of variation in the data, including irrelevant noise. Predictive models in latent space can ignore irrelevant variation and focus on what matters causally. JEPA (Joint Embedding Predictive Architecture) implements this: instead of predicting raw pixels or tokens, it predicts representations in a learned embedding space. The paper shows this produces representations with better transfer, better sample efficiency, and better alignment with causal structure. The 'objective-driven' framing is important because it recenters the question away from 'how big is the model' and toward 'what is the model trying to accomplish.' LeCun argues that objectives shape representations more than scale: a model trained to predict tokens will always learn token-level statistics, no matter how large, whereas a model trained to predict world-state representations can develop genuinely different internal structure. This paper is the technical companion to the 'Path Towards Autonomous Machine Intelligence' manifesto and is required reading for understanding why Meta's research diverges from OpenAI and DeepMind's scaling strategy.
Energy-Based Models: A Foundation for the Next Generation of AI
LeCun's tutorial and framework for energy-based models (EBMs) is one of the most underappreciated contributions in modern deep learning. EBMs define a scalar energy function over configurations of variables — low energy means compatible/likely, high energy means incompatible/unlikely. The model doesn't need to produce a probability distribution over all possible outputs; it just needs to assign low energy to correct answers and high energy to wrong ones. This is more general than maximum likelihood training (which forces models to assign probabilities to everything, including things that should be impossible) and more stable than adversarial training (GANs). LeCun argues EBMs are the natural framework for JEPA and world models because a world model doesn't need to predict everything that could happen — it needs to assign low energy to states of the world that are consistent with observed data and constraints. This connects to his argument against autoregressive models: LLMs use maximum likelihood, which forces them to put probability mass everywhere, leading to confabulation. An EBM-style world model can be 'skeptical' — assigning very high energy (low plausibility) to claims that contradict observed world structure. The tutorial is long (70 slides + notes) but represents LeCun's clearest technical statement of why the next generation of AI needs EBMs or JEPA-style objectives rather than generative objectives.
LeCun on Why AI Is Not Existentially Dangerous — But the Discourse Is
One of LeCun's most-shared social media posts, in which he argues that the AI safety discourse around existential risk is not only wrong but actively harmful to the field. His argument has several distinct prongs: (1) Current AI systems are not agents in any meaningful sense — they generate text, they don't pursue goals in the world. Worrying about misaligned superintelligence in 2024-2026 is like worrying about traffic accidents on roads that haven't been built yet. (2) The 'AI doomers' narrative benefits incumbents — by creating fear around open-source and less-resourced AI development, existential risk framing helps large labs maintain regulatory moats. (3) The actual risks of current AI are real but near-term: disinformation, job displacement, concentration of AI power in a few hands. These get less attention because they implicate existing companies rather than speculative future AI. (4) The people most worried about AGI risk are often the people building the most powerful models — which LeCun finds telling. His critique here is sociological as much as technical: the AI safety discourse has become a subculture with institutional incentives that are partly decoupled from the actual technical questions. This post sparked a major public debate and remains one of the clearest statements of the 'safety but not doom' position from inside a major AI lab.
The Convolutional Neural Network Story: From Bell Labs to Deep Learning
A profile and intellectual history covering LeCun's role in founding modern deep learning, from his early work on convolutional neural networks (CNNs) at Bell Labs in the 1980s through the 2012 ImageNet moment and the current LLM era. LeCun's original CNN work — developing LeNet for handwritten digit recognition at Bell Labs — was one of the first working demonstrations that neural networks could solve real industrial problems. AT&T deployed LeNet for reading checks; it was processing 10-20% of all US bank checks by 2000. But after Bell Labs, the AI winters hit, funding dried up, and neural networks fell out of fashion. LeCun spent years at NYU and later Facebook/Meta continuing to work on deep learning when most of the field had moved on to SVMs and other methods. The 2012 ImageNet moment — when AlexNet dramatically outperformed all previous approaches — vindicated a generation of researchers who had kept working through the winters. LeCun, Hinton, and Bengio's Turing Award in 2018 was recognition that the foundations they'd built in the 1980s-2000s were now running the world's AI infrastructure. This history explains LeCun's current position: he's not a critic of deep learning, he's the most experienced practitioner alive, which is why his skepticism about LLM-based AGI carries weight. He's seen multiple waves of AI hype and knows which ones led somewhere.
V-JEPA and I-JEPA: Meta's Experimental Results on Video and Image World Models
V-JEPA (Video JEPA) and I-JEPA (Image JEPA) are Meta's experimental validations of the JEPA architecture on real data. I-JEPA, published in 2023, showed that predicting image representations in masked patches (rather than raw pixels, as in MAE) produces representations that transfer better to downstream tasks with fewer labeled examples. V-JEPA extended this to video, showing that predicting future frame representations in abstract space — rather than predicting actual pixel values — produces video encoders with better semantic content and better physical understanding. These results matter because they're empirical validation of LeCun's theoretical argument: learning in abstract representation space is more efficient and produces better abstractions than generative prediction of raw data. The practical implication for AI development: if V-JEPA's representations continue to improve with scale (as early evidence suggests), it could lead to AI systems with genuinely better physical understanding than LLMs trained only on text. The limitation: V-JEPA has so far been evaluated on recognition and transfer tasks, not on reasoning or planning, which is what LeCun's world model hypothesis ultimately requires. The next critical experiment is whether JEPA-trained representations support better planning in downstream tasks.
LeCun on Turing Award Acceptance and the State of Deep Learning
LeCun's Turing Award lecture, delivered jointly with Yoshua Bengio and Geoffrey Hinton in 2019, provides an intellectual retrospective on three decades of deep learning research and a forward-looking perspective on what remains unsolved. LeCun's portion focuses on three things: (1) the history of CNNs and why they were initially dismissed by the AI mainstream despite working in practice; (2) the critical importance of self-supervised learning for the next phase of AI development; (3) what he calls 'the dark matter of intelligence' — the implicit, non-linguistic knowledge that humans have from grounded experience in the world, which text-only models cannot acquire. The lecture was delivered before ChatGPT, before GPT-4, before the LLM moment — which makes its predictions remarkably prescient. LeCun correctly anticipated that self-supervised learning would become the dominant training paradigm. He correctly anticipated that text-only models would have a knowledge gap related to physical world understanding. He also anticipated the importance of world models, though he didn't yet have the JEPA architecture to propose as the solution. The lecture is the best single document for understanding LeCun's intellectual commitments before the LLM era, and for evaluating which of his predictions have held up in the years since.
LeCun on Scaling Laws: Why More Parameters Won't Get You to AGI
LeCun's most direct engagement with scaling law proponents (Gwern, Kaplan, Hoffman, and implicitly OpenAI's research direction). The scaling laws finding — that LLM performance improves predictably with compute, data, and parameters — is real and empirically solid. LeCun's argument is not that scaling laws don't hold, but that they hold within a regime that doesn't lead to intelligence. A key analogy he uses repeatedly: building a taller building won't get you to the moon. You can keep scaling next-token prediction, and performance on NLP benchmarks will keep improving, but you won't get a system that has a world model, plans under uncertainty, or reasons causally — because the objective doesn't incentivize those capabilities. The implication for AI labs: the bet on scaling is a bet that intelligence emerges as an epiphenomenon of next-token prediction at sufficient scale. LeCun thinks this is a fundamentally wrong hypothesis, and that we will need architectural innovation (JEPA, world models, modular architectures) rather than just more compute. He notes that the companies who have bet most heavily on scaling have strong incentive to believe scaling works, which should make us appropriately skeptical of their assessments. This post generated significant pushback from scaling-law proponents and remains one of the most substantive public technical debates in contemporary AI.
Multi-Modal Intelligence: Why Video Is More Important Than Text
Data2Vec is Meta's first paper demonstrating a unified self-supervised learning framework that works across vision, language, and speech with a single architecture and objective. The core idea: rather than predicting pixels, tokens, or spectrograms, predict the top-layer representations of the full input from a masked version. This is a form of self-distillation and an early form of the JEPA objective applied across modalities. The paper is important for understanding LeCun's multimodal vision: intelligence requires integrating signals from multiple modalities because no single modality is sufficient to build a complete world model. Text tells you about social and linguistic structures. Images tell you about visual categories. Video tells you about physics, causality, and how objects move in the world. Audio tells you about language prosody and environmental events. A system that learns from all modalities simultaneously — predicting representations across modalities rather than raw data in any single modality — will learn a richer, more grounded world model than any single-modality system. Data2Vec was a stepping stone to V-JEPA and I-JEPA, and represents Meta's approach to multimodal SSL before the LLM wave made text-only approaches dominate AI investment.
Hinton vs LeCun on Whether AI Can Achieve Human-Level Reasoning
A reported piece capturing the ongoing public debate between Hinton and LeCun on AI reasoning. The substantive question: do large language models demonstrate genuine reasoning, or do they perform sophisticated pattern matching that mimics reasoning? Hinton's position has evolved: he now believes LLMs may understand language and concepts in some meaningful sense, even if imperfectly, and that this understanding could scale into genuine reasoning. LeCun's position is more structural: LLMs don't have a reasoning architecture. Reasoning requires search through a space of possibilities, evaluation of candidate solutions against a world model, and rejection of inconsistent candidates. LLMs generate text auto-regressively without any of this structure — they produce the most likely next token given context, which sometimes looks like reasoning and often doesn't. The mathematical contests (LLMs failing at basic algebra, confabulating research citations, inventing plausible-sounding but false facts) are for LeCun not bugs to be fixed by more training data, but symptoms of a architectural mismatch between the objective (predict tokens) and the capability required (reason about the world). The article is a useful synthesis because it interviews both Hinton and LeCun and attempts to characterize the technical nature of their disagreement rather than treating it as a personality conflict.
AI and Jobs: LeCun's Realistic Take on Labor Displacement
LeCun's position on AI and labor displacement is more nuanced than either the 'AI will take all jobs' or 'AI creates more jobs than it destroys' camps. His core argument: current AI will displace significant amounts of cognitive work, but the timeline is slower and the pattern more complex than popular predictions suggest. The tasks most at risk are highly routine text-generation tasks — writing templated reports, extracting structured information, answering repetitive questions. Tasks requiring genuine world knowledge, physical manipulation, social reasoning, and long-horizon planning are much less at risk because LLMs are weak in exactly these areas. LeCun is critical of AI companies (including Meta's competitors) that make dramatic claims about AI replacing entire categories of jobs in the near term, because he believes those claims are both wrong and harmful: wrong because they overestimate current capabilities, harmful because they cause unnecessary panic while distracting from the real AI-driven changes happening more slowly. He advocates for investment in education and retraining with a longer time horizon, and for AI tools that augment workers in demanding cognitive tasks rather than seeking to replace them wholesale. This position is consistent with his technical skepticism: if current LLMs lack world models, they can't replace the world-model-requiring parts of most jobs — the complex, novel, contextual parts that are often the most valuable.
LeCun on the ARC-AGI Benchmark and François Chollet's Test for Intelligence
LeCun's response to François Chollet's ARC-AGI benchmark — a test designed to measure general fluid intelligence by requiring reasoning from very few examples (few-shot learning on novel visual patterns). LeCun agrees with Chollet that current LLMs do not solve ARC-AGI and that this is significant evidence that LLMs lack something important. Where they differ: Chollet believes the solution to ARC-AGI requires a more explicit, program-synthesis-style architecture (his 'arc prize' research direction). LeCun believes it requires world models and JEPA-style learning — the ability to learn abstract representations from sensory data and then apply them flexibly. Both agree that next-token prediction alone doesn't get there. The conversation between LeCun and Chollet (who debate publicly on Twitter/X and in papers) represents the best technical alternative to the 'just scale LLMs' position, because both have proposed specific architectural alternatives rather than just criticizing current approaches. The ARC discussion also connects to the broader question of what benchmarks measure: ARC-AGI intentionally minimizes knowledge and maximizes fluid reasoning, which makes it a good test for the world-model hypothesis. If JEPA-trained models eventually outperform LLMs on ARC-AGI, it would be strong empirical support for LeCun's framework.
Responsible AI: LeCun's Framework for Beneficial AI Development
Meta's Responsible AI (RAI) framework and LeCun's public commentary on it represent his constructive position on AI governance — separate from the existential risk debate. LeCun's positions on near-term responsible AI: (1) Transparency over secrecy — open-source models allow external auditing, adversarial red-teaming, and independent safety research that closed models can't enable; (2) Bias evaluation at deployment, not just development — most bias is context-dependent and emerges in the interaction between model capabilities and specific applications; (3) Resistance to political AI censorship — LeCun has publicly argued against AI systems that refuse to engage with politically contested topics, arguing this makes AI less useful and reflects the values of a small group of engineers rather than the diversity of global users; (4) Structural transparency in training data — knowing what data a model was trained on should be a baseline requirement for any widely deployed system. LeCun is consistently critical of vague 'responsible AI' language that papers over real tradeoffs. His view: genuine responsible AI requires technical transparency (open weights, documented training data) rather than corporate safety theater. The Meta RAI framework is his attempt to implement this — imperfectly, given commercial constraints, but more substantively than closed competitors.
Robotics and AI: Why Physical Intelligence Is the Missing Piece
LeCun's talk on robotics and physical AI is the clearest statement of why he thinks embodied intelligence is essential to achieving human-level AI, not just an interesting application domain. The core argument: human intelligence developed in the context of interacting with the physical world. Our brains spend enormous resources on motor control, spatial reasoning, object manipulation, and prediction of physical outcomes — capacities that are absent in text-only models and very difficult to learn from text alone. LeCun uses the example of a child learning object permanence: a 6-month-old learns that objects continue to exist when hidden because they physically reach for them, drop them, and observe them fall. This knowledge of physics comes from sensorimotor experience, not from reading about it. A robot that learns from interacting with the physical world — dropping things, navigating around obstacles, feeling the resistance of different materials — will develop physical representations that no amount of text training can substitute. This is why Meta has invested in robotics research alongside language model research. LeCun's bet is that the companies that build physically grounded world models through robotics will have a decisive advantage in the long run, even if the short-term frontier in language capabilities is dominated by companies training only on text.
LeCun on Mistral, Falcon, and the Open-Source LLM Ecosystem
LeCun's commentary on the emergence of a competitive open-source LLM ecosystem — Mistral (France), Falcon (UAE), Llama (Meta), Qwen (Alibaba), and others — as vindication of his position that open-source AI is both safe and strategically important. His argument: the emergence of high-quality open models from non-US organizations shows that frontier AI capability is not uniquely concentrated in OpenAI and Anthropic, and that open release is the best way to distribute AI capability globally rather than concentrating it. The geopolitical dimension: LeCun is explicit that he wants AI to be a global commons rather than a technology controlled by 2-3 US companies. Open-source AI allows French, Emirati, Chinese, and Indian organizations to build AI systems aligned with their own languages, values, and use cases rather than being dependent on API access to US-controlled models. This is also why LeCun has been publicly supportive of EU AI regulation that distinguishes between high-risk applications (which should be regulated) and open-source models (which should be exempt from onerous compliance requirements). The open-source LLM ecosystem is the strongest empirical argument LeCun has made: it demonstrates that open release hasn't caused the predicted catastrophes, and it has dramatically accelerated AI research globally.
AI and Democracy: Why Open Models Protect Against Centralized Narrative Control
LeCun's most politically direct argument about AI and democracy. His thesis: AI systems that mediate how billions of people access information and form opinions represent a kind of soft power that is historically unprecedented. If 2-3 companies control which questions AI systems will and won't answer, which political positions they will and won't endorse, and which topics are considered too sensitive for AI engagement, then those companies have an extraordinary influence over the information environment of democracies globally. LeCun argues this is the actual AI risk that policy discussions should focus on: not speculative future superintelligence, but the near-term power of a small number of US companies (and their engineers, with their own political views) to shape AI systems deployed at global scale. His solution is the same as for safety: open-source models. A world with open models is a world where French AI behaves like French people want, where Brazilian AI reflects Brazilian values, where Japanese AI is tuned for Japanese context. A world with closed models controlled by San Francisco companies is a world where one cultural and political perspective gets exported as AI to everyone. LeCun is careful to note he's not arguing for no moderation — he supports legal requirements around harmful content — but against unilateral private corporate decisions about what AI should think about contested social and political questions.
The Future of AI Research: LeCun's 2025 Roadmap for the Next Decade
LeCun's forward-looking statement on where AI research needs to go in the next decade — delivered as a keynote at NeurIPS 2025 and expanded into a blog post. The roadmap identifies five research priorities: (1) Scalable world model learning from multimodal data (video, audio, language, and sensorimotor signals) without manual annotation; (2) Hierarchical planning systems that can decompose long-horizon goals into achievable subgoals using a learned world model; (3) Sample-efficient few-shot adaptation — human-level performance on novel tasks from very few examples, analogous to ARC-AGI; (4) Grounded language understanding — models that associate language with real-world referents through multimodal learning, not just statistical patterns in text; (5) Safe exploration — AI agents that can learn from interacting with real environments without requiring unrestricted access to potentially dangerous action spaces. LeCun explicitly frames this as a 10-year roadmap, not a 2-year product timeline. He's deliberately separating the research agenda from commercial deployment pressures, arguing that the field's current problem is too much focus on near-term product development at the expense of foundational research that will enable the next generation of AI. The roadmap is the clearest statement of where Meta AI Research is investing and why, and it positions JEPA and world models as the foundational bet underlying all five priorities.
Start reading, not hoarding.
Import this vault to Burn 451 and actually read what matters.
New to Burn? See how the read-later app works →
Content attributed to original authors. Burn 451 curates publicly available writing as a reading index. For removal requests, contact @hawking520.