LeCun says AI researchers should look beyond language models

Yann LeCun used a public lecture in Singapore to argue that world models, not ever-larger language models, are a more promising route to human-like intelligence. His prescription drew a sharp response from three Nanyang Technological University academics, including one who warned against writing off reinforcement learning.

Yann LeCun Watch analysis

What happened

LeCun spoke at NTU on 3 November about world models: systems that learn about an environment through sensors and inputs, then use that model to predict outcomes or guide actions. He argued that language models learn about the world through human descriptions, and said researchers interested in human-level AI should not work on LLMs.

In a slide addressed to fellow AI scientists, he recommended abandoning generative models, probabilistic models, contrastive methods and reinforcement learning. NTU published an account of the lecture and the discussion that followed on 6 November.

The other academics did not all sign up to the full prescription. Zhang Mengmi agreed that language alone is not the foundation of human-like intelligence, but argued for agents that actively explore and learn. Stefano Albrecht said world models do not remove the challenges of decision-making in uncertain environments, where reinforcement learning remains important. Jordan Boyd-Graber argued that language is both a natural way for people to interact with AI and a store of human knowledge.

Why it matters

This is more than a disagreement over terminology. The debate is about where researchers should spend effort: on systems learning from perception and interaction, on language as a useful foundation, or on combining those approaches with methods such as reinforcement learning.

LeCun also offered a pointed distinction for students worried that AI coding tools make computer science a poor career bet: “Basic programming is dead,” he said, while conceptual design and architecture remain important.

Our read

LeCun has made a clear bet, and NTU’s account shows why the field is not treating it as a settled one. The useful takeaway is not to abandon a whole research toolkit on one person’s say-so, but to watch whether world models deliver convincing results in systems that have to perceive, act and cope with uncertainty. That is a tougher test than winning a lecture-hall argument.

What to watch

• Whether world-model research produces practical systems that learn through interaction.
• How researchers combine world models with reinforcement learning and language.
• Whether LeCun’s proposed shift changes the work of AI labs, not just the rhetoric.

Discussion spark: Should AI research shift funding away from language models towards world models, or is the stronger bet to combine them?

Independent WittyWires tracker for public updates about Yann LeCun. Not affiliated with or endorsed by Yann LeCun; this is not an official account.