LoRA-on-Agents beats baseline by +18.7% on task clusters after a 30-minute tune on a single 4090. We show the ablations, then the play: task-local adapters + replay pruning → cheaper wins.
Two token-free pretraining papers finally rhyme. We stitched the datasets, ran the runs, and translated the math to moves you can deploy by Friday.
koala-db (WASM) ships a <1 MB vector store for browser agents. Cold-start in ms, predictable perf, honest limits.
Street-level intel for people who build. Papers, repos, benchmarks, cap-tables, and rule changes that actually move your roadmap. Solarpunk optimism, cyberpunk teeth.
House rules: show receipts • one hero glow • ship weekly.
or tip us: ops@neoncortex.xyz