Yann LeCun on Open AI Models, Safety and the Future of Research

Yann LeCun predicts that AI will amplify people’s abilities and make new forms of creativity possible. The Meta chief AI scientist discusses his research career, open models, safety and employment with Harry Stebbings. His predictions and disputed claims reflect his views in 2023 and his position at Meta.

As an engineering student in France, LeCun encountered a debate between Jean Piaget and Noam Chomsky on how language is acquired. Seymour Papert’s discussion of the perceptron led him to research learning machines and neural networks.

He developed an early multilayer learning method, met Geoffrey Hinton in 1985 and later did a postdoc with him. He describes developing convolutional networks during that period before joining Bell Labs, work for which he became known.

Interest in neural networks waned around the mid-1990s. LeCun spent several years on document image compression at AT&T, then returned to learning research. He says he, Hinton and Yoshua Bengio worked to revive interest in neural networks in the early 2000s, a process that took years.

Recent progress looks both continuous and surprising to him. Researchers saw improvements over years, but self-supervised learning with transformers and the capabilities arising from scale exceeded many expectations. Public demonstrations such as ChatGPT make a gradual series of developments look like a sudden jump.

LeCun says predicting the next word at scale produced unexpected language capabilities. He is interested in what comes next: in his assessment, fluent language models still have limited understanding of the world, planning and physical interaction, because much human and animal knowledge is not linguistic.

He expects future systems to learn from observations or simulations and to plan, rather than rely solely on autoregressive language models. He argues current text-trained models can reproduce patterns for some physical or planning questions without having a robust world model. This is his research interpretation and forecast, not a settled assessment of every system.

LeCun says then-current autoregressive models are difficult to steer because their responses depend on training and prompting, without a direct mechanism guaranteeing factuality or other constraints. He predicts they will be replaced by more controllable planning systems within a few years; that timing is speculative.

He challenges the assumption that intelligence necessarily produces a desire to dominate. In his view, human dominance motives reflect social evolution rather than intelligence alone. Critics of advanced AI risk may dispute the inference that a system can be safely controlled simply by choosing its objectives.

LeCun proposes future systems that plan actions against multiple objectives, including factuality, appropriateness and hard safety constraints. A domestic robot, for example, might be constrained when wielding a knife near people. He acknowledges that precisely implementing such safety constraints is not solved and suggests testing and correcting systems progressively.

Stebbings asks who should set those objectives. LeCun anticipates vetting, testing and potentially regulatory processes for high-risk uses such as healthcare and transportation. He also imagines assistant infrastructure and knowledge being developed openly with community review, somewhat like Wikipedia; the governance design remains hypothetical.

He predicts assistants will become a primary interface to the digital world and argues people will want transparent infrastructure rather than a private company controlling it. This is a projection. At the time, Meta was developing and distributing language models, giving the company an interest in the open-versus-closed debate.

LeCun says FAIR’s LLaMA code was open but the model was then distributed for research and noncommercial use because of legal questions around training data. He argues open development can bring together ideas and contributors beyond any one organization.

As an analogy, he cites Linux and Apache becoming foundational internet infrastructure despite competition from proprietary systems. His argument is that broad participation can improve common infrastructure; the historical analogy does not establish the outcome for future AI models.

Stebbings asks how Meta benefits from openness. LeCun cites React, PyTorch and hardware designs and argues that others having access to the technology does not prevent Meta from using it in its own services. He says the value of a tool can be captured in applications such as translation or content moderation.

LeCun points to LLaMA as evidence that useful models need not always be the largest. Pretraining still takes substantial compute, but smaller models can be fine-tuned and run on more accessible hardware. He predicts improved architectures could be more efficient, without claiming that every small model matches the frontier.

He cites approximately 1.4 trillion training tokens for LLaMA and contrasts the volume of text with how efficiently humans learn. He argues observation, world models and hierarchical planning may be missing ingredients for more efficient systems. His comparison is an intuition about learning, not a controlled equivalence between human experience and tokens.

On incumbents versus startups, LeCun favors an open base-model ecosystem with businesses building specialized applications and services on top. He predicts a broad contribution process could improve models faster than proprietary approaches and create jobs; both the market outcome and employment effect are forecasts.

Why did a smaller OpenAI release ChatGPT before Meta or Google? LeCun argues that large companies faced reputational and legal risks. He recounts the backlash to Meta’s experimental Galactica demo and says the team withdrew it. He considers the reaction disproportionate; the risks of misleading scientific text are contested, and his account comes from a Meta executive.

Stebbings suggests established search advertising also creates an innovator’s dilemma. LeCun believes companies must build AI assistants even if they cannibalize existing products, comparing the decision with Meta’s shifts toward mobile and short-form video. He expects assistant interfaces to become widespread, though the trajectory is uncertain.

On employment, LeCun points to past transitions from agriculture and manufacturing to new occupations. He says economists he consulted did not believe society would run out of work and predicts AI will create at least as many jobs as it displaces. That is his interpretation and forecast, not an established consensus or guarantee.

He also recognizes that productivity gains can initially benefit a narrow group and says distribution depends on political and social choices. When pressed to name future jobs, he says he cannot know precisely, but expects creative work and personal services to remain important.

Stebbings worries that adoption is now fast enough to create significant short-term unemployment. LeCun responds that learning and organizational change can slow diffusion, citing economists’ accounts of delayed productivity effects from computers. He estimates effects may take ten to fifteen years; the pace and distribution of disruption remain uncertain.

He speculates that people pay attention to surprising or threatening events because this helps them learn and respond to danger. He offers an infant’s reaction to an unsupported object as an illustration, while noting he is not a social psychologist or sociologist.

Asked about Geoffrey Hinton leaving Google and warning about AI risks, LeCun says they had not yet discussed it. He disagrees with Hinton’s emphasis on extinction risk and says he can speak publicly at Meta with unusual freedom. He acknowledges potential bias when discussing colleagues’ work.

LeCun argues AI can help address problems on social networks. In his account, early feed ranking contributed to information bubbles and clickbait; Meta later changed its systems and used improved models for moderation. His claims about the extent of those corrections and moderation effectiveness are company-side assertions, not independently established by the interview.

Stebbings raises Elon Musk’s concern that powerful AI cannot simply be recalled after release. LeCun rejects a rapid, uncontrollable takeover scenario, arguing that intelligence does not inherently create a desire to dominate and that resources and agency would be needed. This is a contested safety judgment, not a demonstration that future systems cannot pose such risks.

He opposes pausing research and compares restrictions to historical resistance to the printing press. Later he distinguishes that position from regulating AI products, particularly those used for consequential decisions. His analogy and preferred policy are his own arguments.

In the quick-fire round, LeCun criticizes research incentives in China while acknowledging strong Chinese work, praises accessible European undergraduate education but sees weak research career incentives in much of Europe, and praises Switzerland’s pay and research resources. These regional assessments are his opinions.

He describes US strengths in research funding, university salaries and willingness to back risky ideas, alongside the high cost of education. He also observes growing startup activity in Europe while saying capital access can still be harder there.

His main message is optimistic: AI may expand access to knowledge and creative tools. He acknowledges risks but rejects claims that catastrophe is inevitable. He supports regulation for consequential AI products while opposing a slowdown of basic research.

LeCun observes researchers leaving large labs to build startups as commercialization becomes possible. He names Meta’s FAIR and the combined Google DeepMind and Brain teams as well positioned to pursue world models and common sense; his role at FAIR is relevant to that assessment.

Looking ahead, LeCun says he remains motivated by understanding and building intelligence. He values the complementarity of academic and industrial research and intends to keep contributing while he can. Stebbings closes the conversation.