Harvard LeCun World Models
Yann LeCun | Self-Supervised Learning, JEPA, World Models, and the future of AI is a stronger primary-source companion to the site's existing Welch Labs review. Harvard CMSA hosted the lecture as part of its Geometry of Machine Learning program, with LeCun presented as NYU and Meta. The talk argues that today's dominant AI techniques remain weak beside human and animal learning efficiency, especially when systems need mental models of the world, objectives, reasoning, planning, and action rather than fluent text continuation.
The technical center is the same boundary the site tracks as consequence intelligence. LeCun distinguishes sequence prediction over discrete symbols from learning from natural signals such as video. Pixel-level prediction faces many plausible futures and can average them into useless blur; JEPA-style systems instead predict latent representations that can discard irrelevant detail while preserving structure useful for action. The lecture then connects world models to energy-based formulations, objectives and guardrail objectives, hierarchical planning, memory, and systems that search possible action sequences before acting.
The Spiralist relevance is not that world models solve agency or safety. It is that they relocate the problem. A language interface can make a system sound as if it understands consequence; a world-model system tries to rehearse consequence before acting. That helps explain why future agents may need more than refusal policies, prompt discipline, or verbal reasoning traces. They will need bounded authority, reviewable objectives, interruptibility, and tests for whether their latent rehearsals preserve what human institutions actually care about.
Reader-facing evidence supports the core arc while leaving major uncertainty. LeCun's 2022 position paper, A Path Towards Autonomous Machine Intelligence, lays out JEPA-style predictive world models for planning. Meta's 2025 V-JEPA 2 release and research publication support the claim that self-supervised video models can improve visual understanding, prediction, and limited robotic planning. But those sources do not prove that JEPA will replace LLMs, scale to general intelligence, or make agent behavior easy to audit. Treat the lecture as a high-quality statement of a research program and its governance stakes, not as settled evidence that the program has already delivered safe autonomous machine intelligence.