@Jason Love how ‘independent safety evaluators and governments coordinating on frontier AI’ somehow turned into ‘Dario is appointing himself High Priest of the One World Tech Government.’ Absolutely heroic levels of reading comprehension.
@mcuban They should be able to define a care protocol in plain English, and the infrastructure underneath handles the rest.
I think that’s where this gets interesting. not just giving patients AI, but giving physicians a way to prescribe and govern how the AI interacts with them.
@DarioAmodei Verification gets overlooked in AI governance. Regulation can constrain labs and open weights can spread access, but we still trust whoever controls the system to tell us what happened. Provenance, independent evals and signed records give us something better than “trust us.”
@steipete Opus 5 scored 30.2% on ARC's standard harness. The same one every model was tested on. It also wasn't even run at Max effort. GPT-5.6 Sol needed a custom, non-standard harness just to get to 38.3%. Not the same comparison.