Microsoft's MAI-DxO hits 83.2% on complex diagnosis, more than double physician accuracy
TheTuringPost · x · 2026-10-23
Microsoft AI published a Science paper introducing MAI-DxO (Microsoft AI Diagnostic Orchestrator), which simulates 5 clinical roles—proposing diagnoses, selecting questions/tests, challenging assumptions, checking reasoning, and weighing costs—and updates its assessment as evidence arrives.
Key results
- 83.2% accuracy on complex NEJM cases vs 36.1% for generalist physicians; 80.3% on emergency-department cases vs 46.8%.
- Microsoft's blog adds: 1.7x physician accuracy on MIMIC-IV-ED common ED presentations, 2x+ on the hardest NEJM cases, at half to one-third the cost with fewer tests.
Six-dimension medical intelligence framework
- Medical knowledge (apply evidence, know what it doesn't know)
- Diagnostic reasoning and triage
- Therapeutic decision-making (benefits/risks/costs/availability)
- Longitudinal context (tracking health over time)
- Clinical coordination (expertise, follow-up)
- Human values and agency (respecting patient preferences)
Microsoft's bar for medical superintelligence is high: outperforming the strongest human-led health systems across the full course of care. A domain-specific view makes superintelligence testable with clear tasks and criteria.
More from AGI Musings
- Debate: if weak generalization is asymptotic, frontier models will keep commoditizing faster — fkasummer · 2026-10-23
- A cultural evolution hypothesis: forgotten crucial considerations across cultures — jessi_cata · 2026-10-23
- "Claude's breakthroughs are strip-mined from uncompensated human labor," critic argues — AlexTensor · 2026-10-23
- As AI improves, should irreversible actions get time-based friction? A mechanism design question — Afinetheorem · 2026-10-23
- LessWrong thread: Do slower-growth Earths have a better shot at FAI? — jessi_cata · 2026-10-23
- Computer Use Is Solved; Specifying Goals Is the New Bottleneck for Agents — alishbaimran_ · 2026-10-23