Two critical updates re: Astra and mathematics

Substack · Read

• Follow-up to his Astra piece: reports a mathematician was able to replicate roughly half of Astra’s showcased results using an existing (non-Astra) model — undercutting how novel the capability jump actually is.
• Cites Terence Tao’s lecture on “proof indigestion” — the risk of AI-assisted mathematics producing results faster than the field can verify/digest them.
• Cautions against overestimating AI mathematical breakthroughs generally, consistent with his standing skepticism of narrow-benchmark-to-general-capability claims.