Yann LeCun publishes 'A Path Towards Autonomous Machine Intelligence'

Yann LeCun, Meta’s chief AI scientist, circulated a position paper, “A Path Towards Autonomous Machine Intelligence,” setting out an architecture he argued could get machines closer to the way humans and animals learn: efficiently, from a small number of examples, with an internal model of how the world works. Rather than predicting the next token in a sequence — the approach behind GPT-3 and the large language models then dominating the field — LeCun proposed a “joint embedding predictive architecture” (JEPA) that would learn to predict abstract representations of future states of the world, discarding the parts of raw sensory input that are unpredictable or irrelevant.

The paper combined several components: a configurable world model, a hierarchical structure for reasoning and planning at different levels of abstraction, and behaviour driven by intrinsic motivation rather than only external reward. LeCun argued this design would avoid a weakness he attributed to generative, autoregressive models — that they waste capacity trying to predict every low-level detail of their input, including noise that cannot be predicted at all — and that it offered a plausible route to planning and common-sense reasoning that scaling text prediction alone would not.

The paper was explicitly a position piece, not a working system, and LeCun labelled it version 0.9 to signal a research proposal rather than a result. It marked LeCun’s most public break from the industry’s dominant scaling-and-autoregression approach, a stance that put him in direct tension with researchers, including Eliezer Yudkowsky, who argued that current scaling trends were themselves the danger; LeCun was a prominent public sceptic of near-term AI extinction risk. Meta’s research division spent the following years building on the proposal, releasing I-JEPA and V-JEPA world models, before LeCun left Meta in 2025 to found a company built around the approach.