73 commits across five tracks — adding OpenAI alongside Gemini, replacing tone injection modes with a per-user model/product library, curating 0428/0429 tone pools, internal user tiers, and camera/lens picker polish
Google's LiteRT-LM v0.11.0 release adds Multi-token Prediction for Gemma 4 and native Windows support pushing on-device LLM inference forward
Five Claude Code plugins covered over two weeks — two of mine (harnesskit, claude-auto-permission) and three external (graphify, tene, trafficmonitor-ai-usage-plugin), each living on a different layer
A reversible tokenization vault layer on top of OpenAI Privacy Filter — pseudonymization vs anonymization, service boundaries, and why the vault is the load-bearing piece
A digest of five OpenAI announcements that landed on the same day spanning the model API product policy and infrastructure layers
Five parallel tracks landed in popcon over two weeks — payment infrastructure, Cloudflare R2 cutover, an anime-tuned matting model, a brutal UI redesign, and Korean i18n
A teardown of the albond recipe that stacks INT4 plus FP8 hybrid quant MTP speculative decoding INT8 LM head and TurboQuant KV to hit 80 percent speedup on one workstation
BiRefNet fine-tuned on 1228 anime images becomes ToonOut — how a generic matting model that breaks on anime hair gets repaired, plus BiRefNet's own dynamic-resolution training strategy
Notes from two Fly.io videos — Firecracker microVMs, why fctl is a fat client, and how the Machines platform shares the same primitive between fly launch and direct API users
Anthropic just leased the entire SpaceX Colossus 1 data center and turned it into higher rate limits for Claude Code and the Claude API. 220K GPUs and 300MW translate into doubled Claude Code windows, lifted peak throttles, and big Opus API bumps — but the real news is renting a frontier supercomputer from a direct rival
