The AI Clinic · Issue 001Clinical AI
The generalists won.
General AI beat the specialist clinical tools
The big one this week is a blinded study where general AI beat the specialist clinical tools. Plus what a new CMS office means for your practice, and a faster way to get through the literature.
For someone who has screamed from the mountaintops that we must bring the humanity back to healthcare, some of you must be a bit surprised to see me jump on the AI wave so heavily. I see so much opportunity for AI to allow us to be more human and be able to lean into it even harder. This is issue 001, and hopefully one of many that bring you up to speed and keep you at the forefront of AI developments. Thanks for being here.
- A blinded test put general AI against OpenEvidence and UpToDate. The general models won.
- CMS opened a health-technology office. The rules that shape your clinic get written there.
- A faster way to get through the literature, plus three free AI courses worth sharing.
The Big Story
General AI beat the specialist clinical tools
A new study in Nature Medicine ran a blinded, randomized test. Twelve US clinicians scored answers from general frontier models (Google, OpenAI, Anthropic) against two tools built for medicine, OpenEvidence and UpToDate. They rated each on correctness, completeness, safety, and clarity. The general models came out ahead. Eric Topol, who flagged it, said plainly that this wasn't anticipated.
It changes the build-versus-buy question. If a general model answers a clinical question as well or better, the case for paying for a walled-garden tool gets weaker. It doesn't mean you trust it blindly. Scoring information quality isn't the same as bedside judgment, and the safety numbers still matter.
Mark's read: With how good AI agents and models are getting, I think you can train an agent to do what these specialized tools do just as well, if not better. These are the exact skills we're building in our community.
The Business of Care
CMS opened a health-technology office
CMS created the Office of Health Technology and Products, a signal that the agency is serious about AI, digital tools, and interoperability across Medicare and Medicaid.
For owners: this is the plumbing. The rules that decide what AI reaches your clinic, how it gets paid for, and how your data moves get written in rooms like this one. Worth watching for anyone running a practice.
Frontier AI
The US reportedly told Anthropic to suspend its two newest models
Via The Rundown: the US Commerce Department reportedly ordered Anthropic to suspend access to two of its newest models over national-security and export-control concerns. Anthropic is said to dispute the reasoning. This one is sourced from social media, so treat it as a developing claim until a primary source confirms it.
In the clinic: the tool you build a workflow around can be pulled by policy in a day. Don't build a clinic process on a single model you can't replace.
The Big Picture
AI should help us keep clinicians, not replace them
Topol shared a piece on workforce survival with a line worth sitting with: stop asking whether AI can replace clinicians and start asking how it can help us keep them.
For owners and clinicians: the honest question in your building isn't “will AI take my job,” it's “can it give me twenty minutes of my day back.” That is also a retention strategy.
AI for Researchers & Faculty
Send your papers straight to NotebookLM
Andy Stapleton, the go-to voice on AI for academics, walks through a workflow that pushes research papers straight into NotebookLM and ChatGPT instead of a reference manager like Zotero or Mendeley, to summarize, query, and synthesize the literature faster.
For faculty: a real time-saver if you read papers for a living, and it pairs with using NotebookLM as a reading aid. The caveat stands, AI summaries still need a check against the source.
Try This
OpenAI's free Academy courses
OpenAI released three free courses on building practical AI skills and repeatable workflows. They're free, they're credible, and they're a clean way to get a skeptical teammate, a student, or a front-desk staffer their first real win. Point one person at one course this week and ask them to show you what they built.
What else is happening
The last word
One blinded study doesn't settle anything. But it's a good week to notice that the tools you assumed were safest got quietly beaten by the ones you already carry in your pocket.
I would love to hear your thoughts on the newsletter. What was good? What wasn't? Was there anything missing? Reply to any issue and let me know. I read every one.
Get the next issue.
One email a week. The AI that matters for clinical work, in plain language.