Notes on machines that do mathematics, and on the machines themselves.
Two timelines. ai/ traces how AI got here: the architectures, training recipes and ideas that each unlocked a step in capability, and what each of them leaned on. math/ tracks the mathematics: the theorems, constructions and counterexamples machines have actually found, with an honest account of what was proved, who checked it, and how much of it was the machine.
Both are dated to the day where the record allows, sourced, and cross-linked, so you can start at the Riemann-zeta result and walk back through subagents, chain-of-thought, PPO and the transformer to backpropagation.
timelines
- ai development 111 entries, 1943–2026
- ai in mathematics 68 entries, 1956–2026
recent
- math/finite-time blow-up with smooth forcing for IPM, boussinesq and 3D euler
- math/the navier–stokes rumour
- math/claude formalizes fermat's last theorem in eleven days
- ai/GPT-6 astra
- math/the long tail, and the profession takes stock
- ai/claude fable 5.1 and mythos 5.1
- math/erdős #4: first movement on large prime gaps since 2018
- math/palomar: a registry of lean-verified mathematics
posts
- the three modes of machine mathematics
constructions, conjectures and proofs are different epistemic objects, and the honest scoreboard sorts by which one you are looking at - every jump found a new cheap verifier
the ai timeline reads as a sequence of architectures; it reads better as a sequence of things that could not be fooled - how to read an ai mathematics announcement
a short checklist, calibrated on the october 2025 fiasco at one end and the zeta result at the other