Narayanan: recursive self-improvement may mean labs' self-extinction

A Reddit post demos Éloi's lip-sync control, designed to work across multiple languages, with testing so far in English, Korean and at least one more. No benchmark numbers or release details were shared.

Clown Context Suite adds ComfyUI nodes for chaining MiniMax H3 clips, addressing the detail loss that compounds when each clip starts from the previous clip's tail — by clip 10 output looks soft and washed out. The project has moved to a new repo, Clown-Prophecy-Experimental, with a revised technique.

Qodo's Codex plugin feeds company knowledge and coding standards into reviews so developers catch P0 issues across repositories before opening a pull request. OpenAI also published a walkthrough of reviewing pull requests, feedback, and private code discussions in one place.

Message threads, in public beta, let agents send only new messages plus the previous response ID instead of resending full history. With it on by default for most Claude Code users, 500k–1M token requests saw median time-to-first-token drop about 35%.

In an AI Engineer talk, Prime Intellect's Will Brown argues agents relearn the same things every session because company knowledge never gets written down. He makes the case for training models on a firm's own traces, feedback and workflows so they learn on the job like human coworkers.

A Reddit user with little prior experience used Claude to turn a hobby project about the Mongol invasions of Vietnam into a playable game, citing the lack of Vietnamese representation in games beyond Age of Empires II.