In my AI series, we kept circling back to one idea: AI that augments scientific judgment instead of replacing it. Feynman is one of the first tools where I can actually see that vision starting to materialise on my own machine.
What I like is not the "wow, it reads papers" part — it's the plumbing: local runs, explicit workflows (/deepresearch, /lit, /audit, /replicate), and every answer tied back to real citations. In other words, it behaves more like a very fast, very picky PhD student than a magical oracle.
The interesting shift
- "Deep research" becomes a reproducible workflow, not a late-night browser tab forest.
- Replication and audit are first-class citizens, not afterthoughts.
- Web search, code, and PDFs live in one loop instead of three disconnected tools.
How I'm using it
I've started playing with Feynman for exactly the kind of problems we discussed recently: stress-testing claims in methods papers, mapping open questions before an MD project, and designing more honest "replication plans" instead of hand-wavy future work sections.
Over to you
If agents like this become normal, how would you redesign your own literature and methods workflow — from "I skimmed a few PDFs" to something you'd actually trust six months later?