Auto-autoresearch: self-improving agents on Karpathy's NanoChat benchmark

dooku820721

this is legit

jvdillon

The score is a big improvement but the self-improving harness is the most exciting part. If this can actually self-optimize the act of discovery itself then there could (and should) result in a revolution in science itself.

junpenglao

Amazing to see this out, great work Dan and team! Tools that makes all these possible: - Priml ( https://github.com/rekursiv-ai/priml ), strongly typed PyTorch (yes you read that right) ML library that allow you to iterate experiment ultra fast. - Sagent ( https://github.com/rekursiv-ai/sagent ), stop a Claude session and switch to Codex to continue the conversation, and that's a minor feature - Trackinizer ( https://github.com/rekursiv-ai/trackinizer ), epistemologically designed schema and knowledge process, not your regular knowledge graph.

sanjivjindia

Great work Josh and team!