AI agents beat real doctors on diagnosis this week, then flunked the actual ward.

Run time: 6:12

In today's episode:

1. AI agents MIRA and AMIE out-diagnose doctors on test cases
2. Biggest real-world benchmark: top model flunks half of clinical tasks
3. o3 cracks 18 unsolved rare-disease cases
4. Utah lets AI renew prescriptions; oversight thin
5. FDA's cleared-AI list passes 1,500
6. Hetairos AI out-classifies five neuropathologists
7. Fable 5 leaves Anthropic's free plans tomorrow
8. GPT-5.6 and Gemini 3.5 Pro still stuck in preview

TL;DR:

• The week's whole story in one line: autonomous agents (MIRA, AMIE) beat doctors on constructed cases, but BRIDGE — the largest real-world clinical benchmark yet — shows the same model class scores 92% on exams and 44.8% on actual EHR tasks across nine languages. Lab brilliance, bedside gap.
• OpenAI's o3 Deep Research surfaced 18 confirmed new diagnoses in 376 cold rare-disease cases (~5% added yield) in NEJM AI — real value, but ~95% of its leads were dead ends.
• Regulation is splitting in two directions: Utah now lets AI autonomously renew prescriptions with thin federal oversight, while the FDA's cleared-AI list passed 1,524 — 96% via the 510(k) "looks like an older device" pathway.

Sources cited:

• Nature / MIRA
• Nature / AMIE
• Mass General Brigham
• OpenAI / NEJM AI
• Nature Medicine
• Cardiovascular Business
• The Imaging Wire
• Nature Cancer
• Anthropic
• Snowflake

Subscribe: YouTube

medAI Times is for educational and informational purposes only. The content does not constitute medical advice, diagnosis, treatment recommendation, or professional clinical guidance. Consult qualified healthcare professionals and refer to official sources before making clinical, research, regulatory, or business decisions.