Writing

essay · understand it, then run it

JEPA, from first principles to your Mac, with mlx-tune

Yann LeCun's bet on how machines should learn from observation — built up from first principles, then put to work on your own Mac with mlx-tune. What JEPA is, why it works, and how to actually run it.

updated: june 2026
length: 2 parts · 15 chapters · understand + hands-on
covers: latent prediction · collapse · I-JEPA / V-JEPA 2 · LeJEPA & SIGReg · mlx-tune on Apple Silicon
for: ML-literate readers — new to JEPA, wanting both the intuition and the code

sources & attribution Part I synthesizes Yann LeCun's research program — the position paper A Path Towards Autonomous Machine Intelligence (2022), I-JEPA (Assran et al., 2023), V-JEPA 2 (Meta, 2025), and LeJEPA (Balestriero & LeCun, 2025) — cited inline. Part II uses mlx-tune, my open-source library for training and fine-tuning these models on Apple Silicon. My contribution is the synthesis, the ordering, and the implementation; any errors are mine.

contents

loading the guide