Yann LeCun at Sciences Po: Why LLMs Will Never Be Enough, and the Bet on World Models

On September 16, 2026, Sciences Po opened its academic year with its first Grande Conférence: Yann LeCun, Turing Award laureate and executive chairman of Advanced Machine Intelligence (AMI Labs), in conversation with Éric Hazan for nearly 90 minutes on the real state of artificial intelligence. The core message: large language models have hit a structural ceiling, and the next revolution will come from world models — systems able to understand and predict the physical world, not just text.

The full conference is available on Sciences Po's YouTube channel: Yann Le Cun: where is artificial intelligence heading? (in French).

Against the AI "pause"

Asked upfront about calls to slow down AI development, LeCun is blunt: he doesn't buy it. He recalls that in 2019, Dario Amodei (now head of Anthropic, then at OpenAI) refused to release GPT-2 as open source on safety grounds — only to publish it six months later, then keep GPT-3 closed not for safety reasons but to protect a commercial advantage. For LeCun, existential-risk rhetoric mostly serves to justify regulation that would choke off open source, while the same companies keep racing ahead at full speed.

Why text will never be enough

The talk's central argument rests on a single number. All the text available on the internet used to pretrain today's LLMs amounts to roughly 20 trillion words — about 10^14 bytes of raw information. A 4-year-old child has accumulated a comparable amount of information through vision alone, across roughly 16,000 waking hours (the optic nerve carries about 2 million bytes per second).

In other words: an LLM trained on the entire available text of the internet doesn't hold more information than a 4-year-old who has simply opened their eyes. A large part of reality — physics, causality, common sense — simply isn't present in text. That, per LeCun, is why LLMs can pass the bar exam or prove theorems, yet no car is truly self-driving and no home robot can clear a table.

JEPA and world models: AMI Labs' bet

To move past this limit, LeCun has since 2022 championed an architecture he calls JEPA (Joint Embedding Predictive Architecture): rather than trying to predict every pixel of a video — which doesn't work, since much of that information is inherently unpredictable — the system learns an abstract representation of the world and predicts within that abstract space. Trained at scale on observation → action → consequence sequences, such a system can plan: predict the state of the world after an action, and therefore search for the sequence of actions that achieves a goal.

This architecture sits at the core of AMI Labs. According to LeCun, it also carries an economic advantage: unlike LLMs, which must memorize a massive amount of declarative knowledge (hence the enormous GPU memory requirements), a world model replaces memorization with understanding — meaning smaller, less compute-hungry models.

European sovereignty: two doors, not one

Asked about Europe's place against US and Chinese investment levels, LeCun points to two levers:

1. The battle isn't over. It was lost on LLMs, but a new revolution is coming with world models — a field where no one yet holds a decisive lead, since Silicon Valley's industry remains locked into improving LLMs. 2. Open source as sovereign infrastructure. LeCun is backing Project Tapestry, an effort to federate countries and institutions (India, Japan, Vietnam, and others already committed) to train an open, free model fed by each participating country's non-publicly-available cultural archives (national libraries, archives) — a shared repository of human knowledge, as an alternative to proprietary US or Chinese models.

Robotics: "these robots are completely useless"

Asked about the current wave of commercialized humanoid robots, LeCun doesn't hold back: no company in the field currently knows how to make them genuinely useful. The missing data isn't video (abundant and cheap), but data from physical interaction — observation, action, consequence — particularly touch, for which current sensors remain poor and short-lived. The result: robots can run or jump (ground interaction being simple to model), but can't manipulate an object or clear a table.

On risk: "a good James Bond movie plot"

On cybersecurity and bioweapon risks often raised by existential-risk advocates, LeCun is equally direct: in practice, defense always stays ahead of attack, and building a dangerous biological weapon requires skills, equipment, and logistics no LLM can provide. On environmental concerns, he dismisses the data-center water-consumption argument (closed loop, recycled) and notes that the economic incentive to cut energy use already exists without further regulation, since energy is the bulk of a data center's operating cost.

His position on regulation is nonetheless nuanced: he backs regulating use and deployment (healthcare, autonomous driving) — not research itself, which he considers impossible to regulate upfront before the systems in question have even been built.

What this means for businesses

The underlying message, for any organization investing in AI today: LLMs remain a powerful tool for anything text-related — writing, classification, summarization, code — but they aren't the path to systems that understand and act on the real world. Use cases requiring genuine physical or operational understanding (robotics, complex automation) will stay limited until world models mature. In the meantime, the talk confirms a trend already well underway around AI agents: today's real value is built on what LLMs already do well — understanding a request, orchestrating tools, qualifying, responding — not on a general intelligence that doesn't exist yet.

Full conference (in French): Yann Le Cun: where is artificial intelligence heading? — Sciences Po, September 16, 2026