A Path Towards Autonomous Machine Intelligence (Version 0.9.2)
Canonical position paper proposing JEPA, configurable predictive world models, hierarchical planning, intrinsic motivation, and SSL as the blueprint for Autonomous Machine Intelligence
A Path Towards Autonomous Machine Intelligence (LeCun, 2022)
Overview
The canonical position paper from Yann LeCun (NYU Courant + Meta FAIR) proposing a cognitive-architecture blueprint for Autonomous Machine Intelligence (AMI). This paper is the load-bearing primary source for everything LeCun has said publicly about world models and the insufficiency of LLMs since 2022.
Key Architectural Proposals
1. Configurable Predictive World Model
- An internal simulator the agent queries at inference time to predict outcomes of candidate actions
- Distinct from a reactive policy; lets the agent "imagine" before acting
2. Joint Embedding Predictive Architecture (JEPA)
- Non-generative in the sense that it does not attempt to predict
yfromxpixel-by-pixel - Instead captures dependencies between
xandyin an abstract embedding space - Avoids the cost and blurry-average failure modes of generative video prediction
3. Hierarchical Planning
- Multi-level abstraction — high-level (go to airport) composes into mid-level (hail taxi) into low-level (motor control)
- Enables long-horizon planning that would be intractable at a single level
4. Intrinsic Motivation + Trainable Critic
- Agent has a hand-designed intrinsic cost + a learnable extrinsic cost (critic)
- Shapes exploration and learning without requiring dense external rewards
5. Self-Supervised Learning
- The paradigm for training everything — encoders, predictors, world models — on vast unlabeled sensory data
- The bandwidth argument (10^14 bytes sensory vs text) is articulated here
Why This Matters
This is the document to cite when tracing claims back to their source. Every subsequent JEPA paper (I-JEPA, V-JEPA, V-JEPA 2, LeWM, C-JEPA) is an operationalization of one or more proposals from this paper. LeCun's public statements — "LLMs are a dead end for AGI," "we need world models," "open-source is essential for sovereignty" — derive from the framework laid out here.
Published Context
- Version 0.9.2, June 27, 2022
- Posted to OpenReview rather than a traditional venue — explicitly a position paper, not peer-reviewed research
- Follow-up: arXiv 2306.02572 (June 2023) introduces Latent Variable Energy-Based Models as a more mathematical formalization
Notes
Filling the biggest gap in the world-models KB section — this is the primary source behind the LeCun lecture that seeded this whole research thread on 2026-04-22.
Source: A Path Towards Autonomous Machine Intelligence by Yann LeCun, NYU Courant + Meta FAIR