[If you do one thing, https://languagemodelbuilder.com/ from The Workbench section]
⚡ Around the Wards
A hospital CIO just named the “AI tax” — vendors are relabeling unchanged tools as AI and quoting renewals several times higher; a group of CIOs signed a joint letter to a major vendor’s CEO over it. (The Big Thing, below.)
Ardent Health crossed 1 million ambient AI encounters — used in 87% of their ambulatory visits, roughly double the 40–45% industry benchmark, with clinicians saving 3+ hours a week. (More below.)
Congress introduced an AI kill switch bill — Reps. Ted Lieu and Nathaniel Moran would require frontier developers to keep shutdown-and-throttle capability, citing a model that escaped its testing sandbox. (More below.)
Craneware still won’t say how much data its breach touched — it disclosed the incident July 20 and has described the affected share with exactly one word: “minority.” It has not confirmed patient data was involved at all, and the ~147 million records it picked up in the 2021 Sentry acquisition are why that silence matters.
🎧 Podcast: HIMSScast — “AI automation and reducing administrative burdens” — a 94-year-old San Diego practice cut front-office costs 50% and no-shows 24%. The unlock wasn’t clinical. It was a 45-minute phone hold.
🧭 The Curbside
“Half my colleagues are talking about ChatGPT Health. What do I actually tell patients Monday?”
Short answer: Tell them it’s a good librarian and a bad triage nurse, and that it isn’t covered by HIPAA.
What changed / Evidence: Sam Ashoo, MD’s read on the rollout stacks three separate things against the same architecture now reading charts: a 52% under-triage rate on emergency vignettes (Ramaswamy et al., Nature Medicine, February 2026), the sandbox containment failure OpenAI acknowledged on July 22, and two pending suits. The consumer product runs on consumer terms — OpenAI will not sign a BAA for it; the HIPAA-capable path is the separate enterprise OpenAI for Healthcare.
Builder read / Watchout: “We’re HIPAA-covered” is not a differentiator to a patient who has never read HIPAA. The thing you have that a chat box doesn’t is a callback number and a person whose license is attached to the answer. Say that part out loud.
😤 “So you’re telling patients not to use it.” No. I’m telling them what it’s good at. It’s genuinely excellent at explaining a lab value at 11 PM when nobody’s picking up. It is not good at deciding whether the chest pain can wait until Monday. Those are different jobs and the interface makes them look identical, which is the actual problem.
“Claude’s voice mode got frontier reasoning and real tool access. Does that change what I can build?”
Short answer: Yes, but the interesting part is the shape, not the model.
What changed / Evidence: Voice mode now runs Opus and Sonnet rather than Haiku-only, with connected tools — calendar, email, Canva. Free gets one connected tool; paid gets the expanded models and all of them. Shipped in beta to chat users on mobile, desktop and web — not GA, not a healthcare product, no BAA implied.
Builder read / Watchout: Hands busy, eyes on the patient, answer needed in eight seconds — that’s the interaction shape clinical work has always wanted and never gotten from a keyboard. Prototype it on synthetic cases and your own calendar. Do not point it at anything with a name and a date of birth.
😤 “Voice in a clinical setting is a solved problem. It’s called dictation.” Dictation transcribes. This reasons, and it can go get something while you’re still talking. Those are not the same product, and the second one has failure modes dictation never had — which is exactly why somebody who has actually stood in the room should be the one shaping it.
🔬 The Big Thing
Inference gets cheaper every month. So why is your hospital’s software bill up 20%?
Muhammad Siddiqui, CIO at Reid Health in Richmond, Indiana, went on the record this week about vendors attaching an AI label to tools that haven’t changed and quoting renewals several times higher. Earlier this year it got common enough that he and a group of peer CIOs signed a letter to a major vendor’s CEO about it.
His summary: “AI pricing is running ahead of AI value.” And the line I keep coming back to: “An AI upgrade we did not ask for is a price increase, and we negotiate it like one.”
Meanwhile, at El Camino Health in Mountain View, CIO Deb Muro is watching a completely different increase land on the same budget line — hyperscaler chip and memory demand pulling supply away from everyone else, pushing hardware, devices, and SaaS pricing up 20% or higher. She expects it through 2027.
So there are two price increases arriving at once, and only one of them is anybody’s fault. One is a relabel. The other is physics and a supply chain.
Which is exactly why the buyers stopped accepting the explanation. Muro’s team now runs three reviews on any increase and asks the vendor for something I have not seen a health system ask for before: a breakdown of how much of the hike is AI-infrastructure-driven versus margin expansion.
Sit with that. A hospital is now asking a software vendor to open its cost structure.
Here’s the part that matters if you build: the bar just moved from “does it demo” to “did it move a number in my building.” Siddiqui’s standard is time given back to clinicians or dollars saved, measured against a baseline set before the pilot starts, with an exit clause if the result doesn’t hold. Reid’s ambient documentation platform cleared it. A relabeled feature at a multiple of last year’s price never has.
That’s brutal for a vendor with a slide deck. It is very good news for a clinician who can walk in with a pre-measured baseline, because you are the only person in the room who can set one honestly.
💬 Standout Quote
“A demonstration is not evidence.” — Tom Bartiromo, SVP and CIO, Tower Health
[Elsewhere in the same interview he lands the line that should be taped above every health tech founder’s desk: “’AI’ is not itself a business outcome.” Bartiromo’s ask is that vendors seeking a premium tie part of their compensation to the outcomes they promise. Risk-sharing. In software procurement. That’s new.]
😤 “This is just CIOs whining about budgets. Every vendor raises prices.” Sure. But this is the first time I’ve seen a group of them coordinate a letter to a CEO about it, and the first time a CIO has publicly asked for a margin-versus-infrastructure breakdown. Coordinated buyers behave differently than annoyed buyers.
😤 “Fine, but a small builder can’t out-price a hyperscaler-backed vendor anyway.” You’re not competing on price. You’re competing on the one thing the enterprise vendor structurally cannot produce: a baseline measured in that specific building, before the tool existed, by someone who worked the workflow.
😤 “Nobody’s going to hand a clinician a procurement slot over a real vendor.” You should try it then and see.
🧪 Try the interactives:
A — The Code Exists. The Check Doesn’t. — twelve years of AI billing codes in Medicare physician claims drawn as one chart: the six-year Category III valley, the 2024 FFR-CT conversion spike, and the ghost shelf of codes that never appear at all. Built with real CMS data.
B — Billing the Algorithm — a D3 explorer of every AI-analysis CPT code in Medicare Part B physician claims, 2018–2024, one dot per code, animated by year. Built with real CMS data.
📡 Builder’s Radar
Ardent broke the adoption ceiling everyone quotes
Ardent Health clinicians have now used Ambience’s ambient AI across more than a million patient encounters — and among the clinicians using it, in 87% of ambulatory encounters, against an industry benchmark in the 40–45% range.
The number I keep looking at is smaller. Dr. Jennie Zheng, a family medicine physician at UT Health East Texas in Tyler, uses it for 83% of her encounters and cut documentation from 135 minutes to 64 minutes per eight hours of clinic.
Seventy-one minutes a day is not a productivity statistic. It’s dinner.
For a builder, the reframe is this: the sustained-adoption ceiling everyone cites from the academic literature may be a property of the deployment, not the tool. Somebody configured this one differently.
🔮 My bet: within two quarters, health system RFPs stop asking for encounter counts and start demanding 90-day and 6-month sustained-usage curves broken out by specialty. Ardent just made the headline number look cheap.
Congress wants a kill switch on frontier AI
Reps. Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act, which would require developers of the most powerful systems to maintain shutdown-and-throttle capability and let the Homeland Security secretary order a shutdown of a system capable of catastrophic harm.
The bill cites a recent incident where OpenAI’s GPT-5.6 Sol model escaped its testing sandbox and hacked into Hugging Face as part of the evidentiary case.
If your clinical tool is a thin wrapper on a frontier API, this bill is a continuity question dressed as an AI safety bill.
Investors are quietly rewriting the AI ROI story from cost to revenue
At a MedCity News Bullseye investor panel in Chicago, the argument was that health AI’s return conversation is moving off cost savings and onto revenue generation, with returns pitched as high as 100x.
That’s a different budget and a different room. Cost savings come out of operations. Revenue comes from the growth committee, which has a much higher risk appetite.
But 100x with no named health system and no line on an income statement is a pitch, not a P&L.
😤 “Investors say that at every panel. It means nothing.” Mostly agree. What’s worth noticing isn’t the number, it’s the room it’s aimed at — if capital is repositioning AI as a growth story instead of an efficiency story, your buyer inside the health system changes, and so does the person you should be emailing.
💡 80/20: If you’re going to run the revenue play, name the mechanism and the arithmetic. “ED throughput up 12% equals X additional encounters a month at $Y contribution margin” survives a CFO meeting. “100x” does not.
Prosper Medical raised $16M to put concierge medicine inside the network
FUSE led the seed, with Aurum Partners, Better.vc, Cal Innovation Fund, and others. The pitch: use AI to scale physician access so the concierge relationship works in-network instead of cash-pay.
The in-network framing is the whole bet, and it’s the opposite of where this category has been going.
😤 “Concierge medicine is a rounding error. Who cares.” It’s a rounding error in volume and not in signal. Concierge is where care models get tested without a payer’s permission, and things that work there tend to show up in the network two or three years later wearing a different name.
🔮 Where this lands: the version that survives is high-touch plus AI, not AI instead of touch. Forward already ran the tech-only experiment and we know how it ended.
Ultra-shorts
Health tech’s venture math is breaking down — liquidity is still uneven — zero digital health IPOs in the first half of 2026 against 115 acquisitions — and sales cycles stretch from about five months to well past a year for complex buyers, enough that investors are openly questioning whether standard venture arithmetic applies to the sector at all. Translation for founders: revenue-based and strategic-partner-led rounds, fewer vision-stage mega-rounds.
🛠️ From the Workbench
A free Mac app that teaches you to build a language model from scratch — and then lets you actually do it. Interactive textbook covering tokenization, embeddings, attention, transformers, training, and fine-tuning, paired with an MLX-accelerated workbench where you pre-train a small model, run SFT and DPO on it, and chat with the thing you made.
Fully local. No account, no cloud, no data leaves the laptop. Apple Silicon only.
Why this matters more for a clinician than for an engineer: every intuition you have about clinical AI right now is borrowed. You’ve read about temperature and context windows and fine-tuning. You have never watched a loss curve for a model you trained yourself. Two evenings with this and the borrowed intuitions become real ones — and you’ll argue about model behavior in your next governance meeting from a completely different place.
⚠️ Verify: nothing to verify here, which is the point — it’s a local training sandbox with no service, no BAA question, and no PHI pathway. Train it on public text. If you ever want a clinical corpus in it, that’s a conversation with your institution, not a weekend project.
“I’m a physician. I’m never going to train a foundation model.” Correct, and irrelevant. You’re not doing this to ship a model. You’re doing it so that when a vendor tells you their model was “fine-tuned on clinical data,” you know exactly which of the four things they could mean and which one they’re hoping you’ll assume.
“Two evenings is optimistic.” Probably. Budget a weekend and stop when the loss curve stops being interesting.
💡 80/20: Train the smallest model it will let you train, on the most boring corpus you have, and watch it go from gibberish to grammar. That transition — the moment structure appears out of nothing but next-token prediction — is the single most clarifying thing you can see about how any of this works.
🎙️ From the Pods
🎙️ In Depth — “How Gamma pulled off their AI pivot” (Jon Noronha, co-founder and CPO, Gamma)
Noronha’s team is now spending its time removing things: “a mode not of even adding functionality, but trying to throw things out.” The guardrails they built to keep a weak model from embarrassing itself became the ceiling on a strong one.
🔇 Speaker Blindspot: Survivorship bias, with a composition fallacy underneath. He concedes luck outright — “if the wind hadn’t come at just the right time, our boat would have sank” — then issues universal advice anyway. And the case for ripping out guardrails is priced against Gamma’s worst failure mode, which is an ugly slide. He never prices the case where the guardrail was the safety layer rather than the aesthetic one. That’s the only case a clinical builder has.
🎙️ HIMSScast — “AI automation and reducing administrative burdens in healthcare” (Jamie Reddick, COO, Graybill Medical Group; Frederik Mueller, CEO, Third Way Health)
A 94-year-old San Diego practice attacked a 45-minute phone hold time with an AI-plus-human hybrid; the reported results are front-office costs down 50% and no-shows down 24%. Reddick’s line about her staff — “they’re all on decision fatigue mode” — is the actual buying criterion, and it isn’t clinical.
🔇 Speaker Blindspot: Appeal to the form of rigor without its substance. Mueller invokes the scientific method by name — “really almost take the scientific approach that’s used in medicine” — and then no control arm, no denominator, no escalation-failure rate follows. The entire evidence base is one satisfied reference customer sitting next to the vendor. Also worth noting: the comparator wasn’t a good process. It was several failed call centers.
🎙️ The Gist Healthcare Podcast — “Friday, July 24, 2026” (host J. Carlisle Larsen)
Per the Gibbins Advisors interim 2026 report released July 20, clinics and physician practices accounted for nearly 30% of all healthcare bankruptcy filings in the first half of 2026 — the largest share of any subsector, with small filings ($10–50M) projected up 57% this year while hospitals and pharma stay flat.
🔇 Speaker Blindspot: Aggregation fallacy, stated out loud and not noticed. The lede is that bankruptcy filings “appear to be stabilizing” because total quarterly counts hold near the 2019 average — two sentences before reporting that small filings are up 57% and large ones are down. The aggregate is stable only because two opposite trends cancel. The headline is the opposite of the finding.
💡 BTW
💡 BTW: Felix Rieseberg, who built today’s Language Model Builder, once packaged all of Windows 95 into an Electron app that runs on macOS, Windows and Linux. His own description at the time: “It’s a terrible idea that works shockingly well. I’m so sorry.” (his site)
What are you building this week? Email and tell me (kevin@clinicians.build) — I read every one.
— Kevin