Aug 21, 2026

A one-page site asking people to stop pasting raw AI output as replies. It argues AI is fine as a drafting tool, but answers should pass through your own judgment and voice, and it offers a link to send to offenders

Why it matters: Etiquette for handing off AI output is a fight everyone has a stake in, and over 500 comments piled up from both the pasters and the pasted

A 125M-parameter transformer that suggests real-time continuations for a MIDI keyboard performance. INT8-quantized, it runs entirely on an iPhone at roughly 108 notes per second, trained on a few hundred thousand MIDI files (~300M note events)

Why it matters: Scaling the dataset 5x actually made it worse; cleaned data and DPO post-training did the most. A careful log of the build-small, run-on-device path

A context database for AI agents from Volcengine, ByteDance's cloud arm. It unifies memory, knowledge (RAG) and skills into one virtual filesystem the agent can browse with ls and find. AGPLv3

Why it matters: Memory the agent reads itself instead of querying a vector DB, with three tiers of abstraction to keep token use down