83% vs 22%. That is the gap Microsoft's MAI-DxO just posted against doctors on… · @harden_369741 · TwKit

@harden_369741

83% vs 22%. That is the gap Microsoft's MAI-DxO just posted against doctors on the hardest NEJM cases, in a paper Science published this week. The part nobody is quoting: swap in newer models, including GPT-5.5, and the gains hold. The model is interchangeable here. MAI-DxO wins by running like a panel of virtual doctors, one proposing, others challenging, each ordering only the tests it needs. The orchestration is the product. The model is the intern. I have stopped asking "which model" for my agent work. The question that keeps paying off is "which loop." Microsoft also open-sourced SDBench, the sequential-diagnosis benchmark behind the number. That is the part I actually care about. NEJM simulations are still simulations. A public benchmark is how we find out what breaks in a real ED. When your doctor orders a second opinion, how long before it is a panel of agents? Source: https://microsoft.ai/news/building-medical-superintelligence/

曝光 7 · 评论 0 · 点赞 0 · 书签 0