@fchollet currently coincides with math + code because they have verifiers. But not intrinsically math + code. The frontier is whatever can be made verifiable
@DanRomik Yes! This puts the finger on the technical problem: can AI distill each result into the smallest verified, reusable insight, then compound those insights into less search next time? Do you consider that a plausible route beyond “distance one”?
@jdlichtman Yes ok.. this is very close to what I’m saying. Compression is what you will have needed after you discover which distinctions actually matter. Expansion can be used to find the consequential distinctions & verification tells us which survive.. compression removes the rest
@hillelogram Exactly. “Formally verified” always means verified in some formalism against some expressible property.. The interesting problem is connecting those partial views without pretending any one of them sees the whole system
@ilyasergey@leanprover Very cool! Keen to try Vermilion output checked by the MathGraph Lean Kernel, likely useful as an independent terminal-checker compatibility test..
@steida Yes. AI makes code cheap, Lean makes fake correctness expensive. The rare skill shifts from writing implementations to finding and stating the right invariants.