A field guide to Yann LeCun's Joint-Embedding Predictive Architecture, in two halves. First the intuition from first principles — why JEPA predicts in latent space instead of pixels, how it beats representation collapse, and what I-JEPA, V-JEPA 2, and LeJEPA each add. Then a hands-on half: training LeJEPA from scratch and fine-tuning Meta's pretrained image and video encoders on your own Mac with mlx-tune, my MLX library for Apple Silicon.
A long-form personal field guide to JAX — the paradigm, the hardware substrate underneath (roofline, TPUs, GPUs, collectives), the modern ecosystem (Flax NNX, Grain, Orbax), sharded computation, the LLM stack (FlashAttention, KV caches, MoE, μP), and writing Pallas kernels. Synthesized from Google's scaling-book, the JAX docs, and a year of my own notes; kept up to date as the ecosystem moves.
A long read on the Abstraction and Reasoning Corpus — what makes it hard, why pattern-matching falls short, and why it might be the closest thing we have to a fluid-intelligence benchmark. Covers the rules, the leaderboards, notable attacks, and where the real difficulty lives.
