Session logs aren't memory: you can't grep ten thousand turns to find out why you moved off a parser. Funes, a single local binary, indexes those traces with vector search, BM25, and reranking, then gives Claude Code, Codex, pi, and Hermes a recall tool that returns raw turns with exact provenance. It stays on your machine by default, or syncs to a private Hugging Face dataset. One command adds it to the agent you already use.
Deepity is a from-scratch C++ library (with Python bindings) that swaps backprop for predictive coding: each layer adjusts locally to its own prediction error instead of shipping one signal end to end. It hits 97.73% on MNIST in 60 seconds on a laptop CPU, within 1% of a PyTorch backprop baseline and about 50x faster than the author's earlier version. A new variant, DKPPCN, uses Direct Kolen-Pollack feedback alignment to settle in a single step instead of the usual 20-30.
WebGPU runs models in the browser, but portable doesn't mean fast: shader behavior varies wildly across GPUs. The new @huggingface/kernels package ships 207 Apache-2.0 kernels, each in a versioned repo with correctness tests, benchmarks, and WGSL shader templates. A companion tool, Fleet, crowdsources benchmark evidence from real hardware to keep refining them. This is the performance foundation the browser-AI stack has been missing.