CoolFace
Apppublic

mishig/jepawiki

sourceHugging Faceupdated 5mo agoView on Hugging Face
36likes
App README

JEPA Wiki

A Wikipedia-style knowledge base tracking the evolution of Joint-Embedding Predictive Architecture (JEPA) research — from I-JEPA (2023) through V-JEPA 2.1, ThinkJEPA, and beyond.

19 papers covered across the full JEPA family plus related work (DINO, DINOv2, NEPA), with 8 concept pages, 15 animated video explainers (manim), and 7 SVG architecture diagrams.

Built with Next.js, rendered from markdown wiki files with [[wikilink]] support.

How this was built

This wiki was built entirely through conversation with Claude Code, using the LLM Wiki pattern by Tobi Lutke (inspired by Karpathy's endorsement of the approach).

The workflow:

  1. 1.Source discovery: started from Turing Post: 14 JEPA Milestones as a Map of AI Progress article, extracted all arxiv paper IDs
  2. 2.Paper ingestion: used `hf papers read` (Hugging Face CLI) as the primary tool to fetch full paper markdown from arxiv HTML — the agent reads papers, extracts key concepts, and writes wiki pages
  3. 3.Incremental expansion: discovered 7 additional papers (V-JEPA, LeJEPA, IWM, LLM-JEPA, VL-JEPA, EB-JEPA) by searching for related work by Yann LeCun and Saining Xie
  4. 4.Cross-referencing: concept pages (masking strategies, collapse prevention, world models, etc.) synthesize information across all papers with concrete numbers and comparisons
  5. 5.Visualization: 21 short (5-12s) manim animations created using the Hermes Agent manim-video skill — a skill that provides structured guidelines for creating 3Blue1Brown-style educational animations with Manim Community Edition. The skill's visual design standards (opacity layering, monospace fonts, dark backgrounds, one-idea-per-scene) shaped the animation style. SVG architecture diagrams for static concepts.

Resources