Past AI News

All

1. Microsoft AI, Eric Topol and co-authors define "medical superintelligence" with six dimensions, beyond single-task scores

   A framework for evaluating medical AI.

   @EricTopol236 likes · 70 reposts · 28 replies · 21,525 viewsView post

   ExpandCollapseMicrosoft AI, Eric Topol and co-authors define "medical superintelligence" with six dimensions, beyond single-task scores

   Eric Topol shared an open-access paper in BMJ Digital Health & AI by Microsoft AI researchers with Mustafa Suleyman and Eric Horvitz. It defines "medical superintelligence" as surpassing today's strongest human-led health systems through integrated medical reasoning across the full arc of care. The paper proposes six interdependent dimensions: medical knowledge and its application, diagnostic reasoning and triage, therapeutic decision-making, longitudinal context, clinical coordination, and alignment with human values and agency. The authors stress that exam-style or vignette benchmarks cannot establish value; stress tests, cross-population assessment, prospective studies in real workflows and ultimately interventional evidence of better outcomes are needed.

2. Lancet Digital Health crossover trial: 125 clinicians reading head CTs both followed wrong AI advice and ignored correct advice

   Key evidence for human-AI teaming in the clinic.

   @zakkohane22 likes · 7 reposts · 4 replies · 2,167 viewsView post

   ExpandCollapseLancet Digital Health crossover trial: 125 clinicians reading head CTs both followed wrong AI advice and ignored correct advice

   A Lancet Digital Health study shared by Harvard's Zak Kohane had 50 radiologists and 75 non-radiologist physicians or trainees each read 50 head CTs for intracranial haemorrhage under three conditions: no AI, AI prediction only, and explainable AI with annotations. AI advice was correct 80% of the time and wrong 20%. Over-reliance on wrong advice ranged from 27% to 58%, and under-reliance on correct advice from 11% to 21%; because most advice was correct, under-reliance affected more cases and did more net harm to accuracy. Correct explainable advice reduced under-reliance and raised the odds of a correct diagnosis (OR 1.84), but did not shorten reading time or eliminate over-reliance.

3. Hugging Face open-sources DNA model Carbon-X: 14.8B, near GPT-6.1 Sol on variant tasks at about 1/1,800 the cost

   Low-cost, self-hostable genomics analysis.

   @aigclinkChinese-language4 likes · 1 reposts · 9 replies · 1,571 viewsView post

   ExpandCollapseHugging Face open-sources DNA model Carbon-X: 14.8B, near GPT-6.1 Sol on variant tasks at about 1/1,800 the cost

   AI blogger aigclink highlighted Carbon-X, open-sourced by Hugging Face with Zhongguancun Academy and Italy's Telethon Institute of Genetics and Medicine. The 14.8B model reads raw DNA in chat, predicts whether mutations are benign or pathogenic, and explains its reasoning in English. Built on Qwen3-14B with 5,515 added DNA tokens (one per 6 bases), it was further trained on about 302B tokens. It averages 66.7% across five DNA benchmarks, close to GPT-6.1 Sol's 67.5%, at about 1/1,800 of the per-problem cost. The model card estimates analysing ClinVar's 2.4M variants of uncertain significance would cost about $300, versus $32,000 to $4M with frontier models. The BF16 weights fit on one 80GB GPU, but the model is not clinically validated and should not be used for medical decisions.

4. Tsinghua AIR spinout Tairex upgrades its "Agent Hospital": 54 AI doctors, outpatient tests in 18 departments at Tsinghua Changgung

   Deployment data for multi-agent AI in Chinese hospitals.

   @ayonglabChinese-language0 likes · 0 reposts · 0 replies · 8 viewsView topic

   ExpandCollapseTsinghua AIR spinout Tairex upgrades its "Agent Hospital": 54 AI doctors, outpatient tests in 18 departments at Tsinghua Changgung

   According to QbitAI, Tairex, a spinout of Tsinghua's Institute for AI Industry Research (AIR), released a multi-avatar agent collaboration technique and upgraded its "Agent Hospital". Each doctor can have an AI avatar; for cross-specialty cases, avatars from other departments give suggestions as alerts or in consultation mode, disagreements are highlighted, and the doctor decides. As of July 2026 the system had 54 AI doctors and millions of simulated patients covering thousands of diseases. At Beijing Tsinghua Changgung Hospital, 18 clinical departments run outpatient tests and 40+ doctors use it routinely, with overall efficiency up 30%. The company says its AI doctors exceed 96% accuracy on MedQA.

5. iFLYTEK Healthcare upgrades its Spark medical model to V3.7, merging speech, imaging and language, and reports 99.6 on MedBench 5.0Unconfirmed

   Iteration of a Chinese medical LLM.

   @aitop_newsChinese-language0 likes · 0 reposts · 0 replies · 5 viewsView topic

   ExpandCollapseiFLYTEK Healthcare upgrades its Spark medical model to V3.7, merging speech, imaging and language, and reports 99.6 on MedBench 5.0

   Via an AI news aggregator citing Pandaily, iFLYTEK Healthcare upgraded its Spark medical model to V3.7, merging speech, imaging and language for medical agent applications. The company says the version scored 99.6 on Shanghai AI Lab's MedBench 5.0. The underlying report is brief and gives no details on deployment sites, clinical outcomes or comparison models; the score is self-reported.