JEPA Wiki — Index

Where to start

If you're new to JEPA: read overview first for the full story, then concepts/latent-prediction for the core principle, then concepts/jepa-vs-alternatives to understand how JEPA differs from LLMs and diffusion models.

If you're looking for a specific paper: use the paper table below — each page has an architecture diagram or video, key results, and cross-references.

If you want concrete numbers: go to concepts/benchmarks-and-results.

Concepts

Core principles

• concepts/latent-prediction — The defining JEPA principle: predict in representation space, never in pixel space
• concepts/masking-strategies — Patch, multi-block, sequencer, object-level, dense predictive loss, deep self-supervision
• concepts/collapse-prevention — EMA, SIGReg, deep self-supervision, VICReg, distillation

Architecture & training

• concepts/vision-transformers — Encoder scales (5M to 2B), tokenization, positional encoding, predictor designs
• concepts/training-recipes — Datasets, training schedules, hyperparameters, scaling laws, compute costs

Comparisons

• concepts/jepa-vs-alternatives — JEPA vs LLMs, diffusion, DINO, MAE. The capacity waste argument, why diffusion can't plan, energy-based planning, representations as coarse-grained data.

Applications

• concepts/world-models-and-planning — Action conditioning, CEM planning, robot results, planning speed
• concepts/object-centric-representations — Patches vs objects, causal inductive bias, efficiency gains
• concepts/benchmarks-and-results — Consolidated results across all tasks and models

Papers — JEPA Family (by date)

Papers — Related Work

Not yet ingested

| 2507.02915 | Audio-JEPA | 2025-07 | audio | not on HF Papers |

Foundational

• papers/lecun-position-paper — "A Path Towards Autonomous Machine Intelligence" (LeCun, 2022). The 62-page blueprint: cognitive architecture, JEPA, H-JEPA, energy-based planning, Mode-1/Mode-2, four training criteria. What was validated and what remains open.
• papers/saining-xie-interview — 7-hour interview with Saining Xie (AMI Labs co-founder). Three stages of understanding JEPA, the case against LLMs, DiT/REPA origin story, AMI Labs mission, forward predictions.

Sources

• Turing Post: 14 JEPA Milestones — the article that seeded this wiki
• LeCun: A Path Towards Autonomous Machine Intelligence — the foundational position paper