Six Hundred and Fifty Discarded Proofs
A model threw away six hundred and fifty invalid proofs before it found one, and seven healthcare engineers spent the same week explaining why they can't do that. The episode is about what a verifier costs in a domain where checking is as hard as answering.
Fireship on the summer's machine-generated proofs — a secondhand but detailed account of the Anthropic Riemann run: 36 hours, 60 sub-agents, 650 discarded attempts. Formal verification made brute-force search viable, which is the whole trick.
Dwarkesh Patel on research-and-development sufficiency — his claim is that you only need models…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.