What's this about?
Today's top HuggingFace paper (499 upvotes). Agents doing ML research end-to-end keep "operational knowledge" — the know-how separating knowing a method from making it run — outside the agent, rediscovering it every run. DisCo distills GitHub repositories into compact, verified skills so it's reused instead.
What exactly is "operational knowledge"?
The unwritten stuff: which flags matter, what breaks, the workarounds — the difference between reading about baking and having baked. It lives in working code, not papers, which is why repositories are the distillation source.
Why is this worth noting?
499 upvotes made it the day's clear community favorite, and it formalizes what practitioners already do by hand (writing skills for their agents) into an automated pipeline: agents building their own skill libraries from the world's code.