A Year of Unpublished Math Work Fed to an AI Coding Tool, and Training Can't Be Ruled Out
Scobleizer · x · 2026-09-10
Babak argues enterprises can keep using frontier models, but the closed loop—your data training your agents, agents improving at your business—should run on models you control.
The trigger: two mathematicians put a year of unpublished work into an AI coding tool; asked whether it trained the model, the best available answer was "unlikely, but we cannot rule it out." As Scoble notes, nobody did anything wrong, but an unverifiable promise is not a control—businesses should pay attention to data sovereignty in AI workflows.
Related event: Navier-Stokes dispute sparks debate over AI training on researchers' ideas(4 posts)→
More from AGI Musings
- Naval: Frontier labs' flywheel is distilling data from the smartest users of leading models — naval · 2026-09-10
- CMU Proposes Discovery Certification Protocol: Scores Alone Don't Prove AI Research Agent Discoveries — CarnegieMellonU · 2026-09-10
- levelsio: AI agents will vibe code SaaS features, replacing subscriptions with x402 micropayments — RileyRalmuto · 2026-09-10
- voooooogel not looking forward to where the AI safety mass movement heads — voooooogel · 2026-09-10
- x-risk researcher points newcomers to s-risks, the often-overlooked worse scenario — jacyanthis · 2026-09-10
- Proposal for safety researchers: quit in pairs with rival lab's capabilities staff — ESYudkowsky · 2026-09-10