Abridge goes foundation model 🏗️, Six meetings → one chat ⚡, What's up with Fable's data retention? 🚫
[Still on shifts, so we’re running compact again.]
📡 Quick Hits
Abridge is building the first foundation model for clinical conversations — with Nvidia compute and Eli Lilly money (Fortune, STAT). Announced yesterday alongside a strategic Lilly investment aimed at trial screening. The scribe leader is moving down-stack: the conversation isn’t the note anymore — it’s the substrate for billing, decision support, and trial matching.
Dr. John Lee replaced six order-set committee meetings with one 40-minute AI chat (The Swirl Is Optional). An ER physician and a pharmacist built a VTE anticoagulation order-set structure by pasting guideline text, a dosing spreadsheet, a screenshot, and a mangled table into a plain chat window. The redesign target was never the tool - it’s the swirl around it.
Microsoft restricted internal use of Claude Fable 5 while it reviews data-retention terms (The Verge). No ZDR for Fable, 30 day data retention. If Microsoft won’t trust the retention terms yet, maybe you shouldn’t either.
A new Stanford preprint adds a conformal safety layer to LLM clinical summaries (arXiv). Suhana Bedi’s team overlays calibrated omission and hallucination flags on any model’s summaries — formal control over which outputs need human review. This is the eval infrastructure that makes ambient tools defensible.
AMN Healthcare acquired Jaide Health, the clinician-founded AI interpretation startup (press release). A physician (Corkery) and a nurse (Julie Wilner) built an AI language-access company, and a national staffing firm just bought it. The clinician-built-to-strategic-exit path is real and getting shorter.
A free 30-day Claude Code course built for clinicians just launched (ai101.health). Dr. Michael Hobbs shipped it this week. If you’ve been waiting for a structured on-ramp that assumes a clinical day job, this is weekend-sized.
Visa and Mastercard both shipped agentic-payment platforms this week (the skeptical read). Linas Beliūnas notes on-chain agentic payment volume already spiked past 5M/day and collapsed once — but eligibility, scheduling, and bill-pay agents will eventually ride rails like these. File under “watch, don’t build on yet.”
🎙️ From the Pods
🎙️ Tradeoffs — “Where Did the No Surprises Act Go Wrong? And Can It Be Fixed?”
A Penn LDI panel walks through how the NSA’s arbitration layer got captured by its own incentive plumbing: providers win more than 80% of disputes at ~2.7x median in-network rates (one diagnostic test paid $330,000 vs. a typical $2,600), and arbitrators are paid per case — volume is the business model. Lindsey Murtagh, the former CMS official who helped write the rules, says litigation turned the process into “an extremely provider-friendly environment.”
💡 Builder take: The law created a mandatory negotiation between two parties with no communication rails — that’s an unbuilt software surface. And in any adjudication workflow you build, model who profits from case volume before you model who wins.
🔇 Speaker Blindspot: Anchoring — the panel treats the QPA (the median in-network rate, computed by insurers, one party to the dispute) as the self-evidently correct benchmark, and with no provider voice in the room, whether the QPA itself can be gamed downward never comes up.
🎙️ Health Tech Nerds Radio — “The billing problem is actually an affordability problem” (Seth Cohen, Cedar)
Cedar’s president argues the patient-billing problem quietly became an affordability problem: out-of-pocket has grown from 1–2% to 5–12% of net patient revenue while only ~40% of what patients owe gets collected. The build surface isn’t a prettier statement — it’s resolving affordability “in segments of one” at the bill-pay moment: the line-item denial explained for one patient, the dormant HSA or charity-care eligibility surfaced for another. “Two thirds of people with HSA accounts in this country have never opened their account.”
🔇 Speaker Blindspot: Equivocation — he renames the collections problem an “affordability problem,” but every solution described is measured as yield on patient AR, and his own data (80–90 cents on the dollar at affluent systems vs. 15–20 at safety nets, driven by demographics) quietly concedes that collectability is mostly income, not experience.
🎙️ Podnosis — “UnitedHealthcare exec talks AI, affordability and member engagement”
Dan Kueter, CEO of UHC’s employer & individual business, describes a new employee-level storefront where individual members buy point solutions — weight loss, ortho, PT — with tax-favored HSA/HRA dollars at an average 26% negotiated discount, even when their employer didn’t pick those vendors. That’s a payer-owned distribution channel a clinician-built point solution can sell into without ever winning an employer RFP.
🔇 Speaker Blindspot: False analogy — benefits keep getting mapped onto Amazon light bulbs, but people who expect to use care select richer coverage, and the adverse-selection problem that mass benefit customization creates for risk pooling never comes up once.
💡 BTW: Shiv Rao — whose company just teamed up with Nvidia to build a foundation model — still practices cardiology while running a $5.3B startup: he’s on call Thursdays and takes shifts one weekend a month, and credits his obsessions with skateboarding and music culture for how he builds. Fortune profile
What are you building this week? Email and tell me (kevin@clinicians.build) — I read every one.
— Kevin