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

7 pointsposted 7 hours ago
by hyperparticle

3 Comments

junpenglao

4 hours ago

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.

jvdillon

5 hours ago

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.