Fable 5 got pulled by the feds 🚨, The task is the unit of work 🧱, 4 in 10 physicians now have a side gig ⁉️
[Still running light secondary to clinical shifts - couldn’t do a mindset this Sunday so some quickhits]
Fable 5 got suspended by a US Commerce Department export-control directive — and it broke every workflow that depended on it. Anthropic’s statement confirmed Fable 5 and Mythos 5 access was suspended for foreign nationals under a government directive. Separately, Wired reported that before the suspension, Fable 5 was already silently rerouting certain requests to Opus 4.8 without warning. First it was quietly nerfed. Then it was pulled entirely. If your clinical workflow depends on a specific model, what’s your fallback plan?
😤 “This is geopolitics, not a builder problem.” It is absolutely a builder problem. A government directive just made a production model unavailable overnight with no deprecation notice. If your ambient scribe or your CDS tool pins to a single model version, you’re one executive order away from a broken workflow and a silent failure nobody notices until a patient does. Model fallback isn’t an optimization — it’s a safety requirement.
😤 “Just use a different model. They’re all the same now.” They are not the same. If they were, Anthropic wouldn’t have charged a premium for Fable. The issue isn’t switching — it’s knowing you need to switch when the failure mode is a 403 your end user never sees.
Doug Fullington reframed how clinicians should use AI: the question is not the unit of work. In The Question Is Not the Unit of Work, the internist argues most clinicians still use AI like a search engine — ask a question, get an answer, then do all the work by hand. The real value is delegating bounded tasks with a four-part recipe: job, context, steps, check. The payoff isn’t time saved (a Stanford pilot showed mixed results) — it’s cognitive. You convert draining composition work into cheaper supervised review. Four escalating rungs: dictate substance → saved instructions → standing triggers → automated audited systems. [nice!]
OpenAI is acquiring Ona to give Codex agents persistent, secure cloud environments. The announcement targets the infrastructure gap for agents that need durable, isolated sandboxes for long-running enterprise sessions. This is the plumbing layer for reliable agentic workflows — the kind that matters when you want an agent to maintain state across a multi-hour clinical documentation review or a prior-auth appeal.
Arlen Meyers offered a four-role framework for physician AI side gigs. In Founder, Funder, Finder or Futurist, the SoPE founder maps four entry points for physicians engaging with healthcare AI. Medscape data: 4 in 10 physicians now have a side gig, ~75% report equal or greater fulfillment than clinical work. The three most common AI lanes: training AI with clinical judgment, AI-assisted content, or healthcare data consulting. “Always have Plan B since AI might soon replace you.”
Anthropic launched Claude Corps — $150M to place 1,000 AI fellows inside nonprofits. The program embeds trained fellows for year-long AI coaching engagements in the social sector. Relevant for community health, safety-net systems, and public health orgs that can’t hire an ML engineer but could host a fellow who builds capacity from the inside.
🎙️ From the Pods
🎙️ HIMSS26 — “Why Medicare Patients Don’t Trust AI in Healthcare“ (Dr. Mehmet Oz, CMS)
The CMS administrator said it plainly: “We query Medicare patients about what they think about AI. They do not trust it.” Nobody has gotten to them with a use case that makes AI feel like their tool rather than a hospital tool deployed against them. Oz’s ask to the HIMSS audience: make the case that AI saves lives, improves access, and manages a $1.8T program — or it stays an operations tool that patients fear.
💡 Builder take: If your patient-facing tool doesn’t explain itself in language a 72-year-old in rural Mississippi would trust, you’re building for the CIO’s slide deck, not for adoption.
🔇 Speaker Blindspot: Composition fallacy — “we as HIMSS … need to reach people” conflates the industry’s PR challenge with individual product trust. Patients don’t distrust “AI” abstractly; they distrust specific systems making decisions about their care. The trust problem is local and product-level, not a messaging campaign you solve with a better brochure.
🎙️ Lifers — “Why AI Is Primary Care’s Best Chance at Survival“ (Dr. David Carmouche, Lumeris)
Carmouche runs clinical and commercial at Lumeris (11,000+ physicians in value-based care). His thesis: primary care doesn’t have a technology problem — it has a time-and-complexity problem that only AI solves at scale. The agentic platform they’re building (”Tom”) isn’t replacing physicians; it’s handling the administrative friction that drives 50% of burnout.
🔇 Speaker Blindspot: Survivorship bias — Lumeris’s 11,000 physicians are already in value-based contracts, which is precisely the payment model that rewards administrative efficiency. In fee-for-service (still the majority of US primary care), “giving back time” might mean giving back revenue. The AI-for-primary-care thesis only works inside the right payment model — and Carmouche doesn’t name that dependency.
💡 BTW: Arlen Meyers, the SoPE founder who wrote today’s physician AI side-gig framework, is an ENT surgeon who invented a biophotonics device (Oncolight) that uses light to detect cancer — then pivoted to medical tourism (MedVoy) — then decided nobody teaches doctors how to be entrepreneurs and built the Society of Physician Entrepreneurs from scratch. Ear-nose-throat to cancer optics to tourism to education in one career. Collabs.io profile.
What are you building this week? Email and tell me (kevin@clinicians.build) — I read every one.
— Kevin