medAI Times

Claude designed three hundred thousand proteins in sixty-four hours. Seventeen worked.Run time: 6:30In today's episode:Claude designs 17 scorpion toxin binders; no antivenom yetOpenAI swarm of 2,000 agents mines mouse brain atlasFace-photo AI predicts cancer survival in Lancet Digital HealthFDA clears TumorSight Risk for breast recurrence estimatesFDA clears Elucid FFR-IQ for CT-based coronary flowAstraZeneca-insitro deal: $40M upfront, $1.5B headlineFDA generative AI device comment window closesAnthropic opens Claude Duet agent to waitlistAnthropic draft IPO papers leak revenue and lossTL;DR:Claude's scorpion-toxin binders are real lab hits (17 of 211) but only binding data: no neutralization, no venom, no animals.OpenAI's agent swarm produced a DCTN1-cholesterol lead from a mouse brain atlas; preprint, company-authored, unconfirmed at the bench.Two FDA clearances (TumorSight Risk, Elucid FFR-IQ) rest on company validation that shows consistency, not outcome benefit.Sources cited:AnthropicThe DecoderbioRxiv preprintpress releaseTech TimesNews-MedicalFox affiliate report of company releaseCardiovascular BusinessFierce BiotechFDASubscribe: YouTubemedAI 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.

Eight states just took ChatGPT to court, and medical advice is in the complaint.Run time: 7:07In today's episode:Eight states sue OpenAI; the FTC sends demands to OpenAI and AnthropicChatbots miss GI emergencies far more often outside EnglishPenn patients text their EEG monitors in Nature MedicineDeepMind agent finds an antibiotic candidate; wet lab pendingZealand pays Generate $30M up front for AI-designed proteinsOpenEvidence goes live in Northwestern's Epic and signs HopkinsUnitedHealth says AI chart summaries reach 40% of its cliniciansFOLLOW-UP: UK accepts all 44 health AI regulation recommendationsMeta releases Muse Spark Open, 816B open weightsNEW: Anthropic embeds six METR evaluators with publication rightsTL;DR:Eight states (October 14) and the FTC (October 14, reported October 15) moved against frontier chatbot makers in the same 24 hours. The states' complaints name harmful medical guidance and a claimed overdose. All of it is allegation so far.An ACG abstract found four small chatbots failed to advise emergency care in 15% of English answers and about 40% of Chinese and Vietnamese ones. It's the most clinically useful number of the week, from the weakest kind of source.The week's AI biology "discovery" (DeepMind's bacteriocin) is computational only. The AI drug money (Zealand and Generate) is real, but $30M of it is up front and the rest is milestones.Sources cited:EngadgetThe New DailyGizmodo (FTC)HealioPenn MedicineGoogle DeepMindBioSpaceHealthcare IT NewsNorthwesternUnite.AI (Hopkins)Subscribe: YouTubemedAI 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.

Can a chatbot safely interview real patients before the doctor does? Google just tested it.Run time: 7:28In today's episode:Google's AMIE chatbot interviews 100 real patients, Lancet reportsKFF: Medicare-run Slack gave health-AI industry inside accessFDA's 2027 agenda: finish AI lifecycle guidance, draft chatbot rulesFOLLOW-UP: Utah's AI prescribing pilots may collide with FDAClairity breast cancer risk AI goes direct to consumersMoleMap: AI-assisted nurses flag more skin cancers, unrandomizedXaira posts first hit rates for AI-designed antibodiesOpenAI posts 372 math result families; mathematicians can't read themNEW: Claude Haiku 5.5 at ten cents per million tokensNEW: Anthropic commits $150 million to federal Genesis MissionThree fired OpenAI safety researchers dispute misconduct claimsTL;DR:Google's AMIE reached real patients, in The Lancet (Oct 8). 100 adults at a Beth Israel Deaconess urgent primary care clinic, a physician watching every chat, zero safety stops, final diagnosis in the differential 90% of the time. Single arm, single centre: feasibility, not benefit.OpenAI published 372 families of AI-generated math results (Oct 6), including a claimed proof of the Unique Games Conjecture. About 42% of top-line results are Lean-checked, and the people best placed to read the proofs say they can't yet.Health-AI policy is being shaped in rooms the public can't see. KFF Health News (Oct 9) documents a 1,700-member CMS-run Slack and an unlisted FDA listening session on patient chatbots, the same week the FDA put generative-AI mental health devices on its 2027 guidance agenda.Sources cited:The LancetGoogle blogpreprint, arXiv 2603.08448KFF Health NewsCBS NewsMedTech DiveRAPSSTATHealth Affairs Forefront backgroundEverlywell release via Yahoo FinanceSubscribe: YouTubemedAI 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.

California just made it law: AI can advise your nurse, not replace them.Run time: 6:35In today's episode:California law limits hospital AI to advice, not decisionsRFK Jr. says AI beats any doctor as second opinionSynthetic Hospital fools physicians, stumps ten AI modelsHarrison.ai head CT tool tested on 3,424 ED scansHeidi II moves scribe AI into referrals and follow-upsUltrasound-video AI lifts junior radiologists on lymph nodesAI "ECG age" refines stroke risk in borderline AFHHS launches SURPASS to rebuild clinical trials with AIGoogle ships Gemini 4 Argon, cyber defenders firstNEW: Claude Sonnet 5.5, same price, fasterLeaked Anthropic prospectus shows cloud-partner dependenceTL;DR:California signed AB 1979 (Sept 30): AI cannot independently perform clinical functions reserved for licensed professionals, effective in 2027. Newsom vetoed AB 2575, the bill that would have protected clinicians who overrule an algorithm.The federal government is sending the opposite signal. HHS Secretary Kennedy called AI a second opinion "better informed than any doctor" (Sept 29), and HHS/ARPA-H launched SURPASS to speed trials with AI (Sept 30).Evidence this week was mostly middling: a strong open benchmark (Synthetic Hospital, preprint), a retrospective head-CT validation, and several vendor launches with no outcome data.Sources cited:CalMattersGovernor's officeNational Nurses UnitedYahoo NewsNewsweekarXiv 2609.30027GitHubEmergency Medicine Australasia, DOI 10.1111/1742-6723.70353Fierce HealthcareRadiology: AI, DOI 10.1148/ryai.260131Subscribe: YouTubemedAI 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.

Hospital A I added nearly a billion dollars in charges. Nobody got sicker.Run time: 7:28In today's episode:Blue Cross: hospital AI added $942 million in chargesVA picks Abridge for ambient AI, $775.7M ceilingAllina doctors win first contract after AI-fueled strikeAnthropic, OpenEvidence give clinical AI to 100 countriesClaude flags a new CRISPR-adjacent enzyme systemPhone app separates Crohn's from intestinal TBUNC and Duke launch rural AI network for North CarolinaOnly a third of brain-imaging AI studies externally validatedNEW: Claude Opus 5.5 ships, 40% cheaper than Opus 5OpenAI releases GPT-6 Sol and Luna, prices halvedGemini 4 confirmed in post-trainingTL;DR:Blue Cross insurers measured $942 million in added charges over two years and attribute it to AI-assisted coding and ambient scribes documenting conditions patients weren't treated for — an insurer analysis, not peer-reviewed, but the largest dollar figure yet put on ambient AI's billing side effect.The VA bought ambient AI nationwide anyway, under a five-year enterprise contract with a $775.7M ceiling covering 9M+ veterans across both its record systems.Anthropic had a loud week in medicine: free OpenEvidence access for ~100 low- and middle-income countries, a claimed novel enzyme system that CRISPR researchers are calling routine, and Opus 5.5 with biology capability locked behind a vetting program.Sources cited:PYMNTSHIT ConsultantMPR NewsThe Next WebAnthropicroutine genome miningAliment Pharmacol TherUNC HealthJ Imaging Inform MedAnthropicSubscribe: YouTubemedAI 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.