someone just broke googles unbreakable AI watermark with 200 black images and a math transform from 1965. billions in detection research, defeated by a technique older than the internet. at some point we gotta admit pixel-level watermarking is a dead end

marimo-pair drops agents into running python notebooks. they see your data, run code, build plots. no more copy pasting between chat and jupyter github.com/marimo-team/ma…

still got a couple of (niche?) usecases of gemini beyond the nano banana stuff: 1. dump long (10-20k+ loc) chat summaries from cursor chats and then tell gemini to summarize the core findings for handover to other agent(s) 2. dump entire pipeline logs to spot failures and

xt day

found archon. its like github actions but for ai coding, you write yaml workflows that force your agent through plan, code, test, review steps. no skipping github.com/coleam00/archon

MOSS-TTS-Nano. realtime text to speech on CPU, no GPU needed. 0.1B params, 20 languages, voice cloning tested locally and latency is wild for something this tiny github.com/OpenMOSS/MOSS-…

been testing caveman on claude code. it strips all filler from AI output, keeps code untouched. ~75% fewer output tokens, accuracy stays the same. one line install github.com/JuliusBrussee/…

been testing manifest for routing LLM calls. it scores each request in 2ms and picks the cheapest model that can handle it. self hosted, 300+ models. simple queries go to haiku instead of opus and you dont notice github.com/mnfst/manifest