🚀 After 18 months of building, MeDucation AI is here: meducationai.com The AI-native medical education platform I wish existed when I was a med student and a heme/onc fellow. 👨‍⚕️ Every question reviewed by an expert physician. 🤖 Powered by AI. 🧠 Made for how you actually learn. What's inside ↓

A patient can read a press release the afternoon it goes out. You get to the paper six weeks later. They bring it to clinic printed out and you have not seen it yet. OncoPulse closes that gap to one morning. Free. meducationai.com/tutor/signu…

Have you ever wondered how GitHub actually works? I taught myself coding as a busy heme/onc fellow, now I build software with startups. My new course shows you what GitHub is, how it differs from your laptop's code, and how you use it to build your own tools. udemy.com/course/build-apps-…

Suki raised a $90 million Series E last week. Total funding is now above $250 million. The CEO’s line is the interesting part: “Clinical documentation was never the destination. It was the entry point.” The company reports that revenue doubled in 2025, patient encounters grew more than 190% this year, and partner channel revenue grew more than tenfold in two years. Those are company figures, and the valuation was not disclosed. Documentation got them in the door. Now they are racing to become the platform: clinical questions, billing, operations, partner ecosystems. The wedge only works if you end up owning something the platform cannot copy. Epic has its own scribe.

The question I got most during these live webinars wasn't "which AI is best." It was "where do I start without breaking something." So that's what the series answers. Failure modes, not features. 4 sessions, 3 hours, free: piped.video/playlist?list=PL…

I built a local AI scribe. Wi-Fi off, it still transcribes the visit and writes the note, all on my laptop. This was inspired by a recent talk I gave to doctors and was shocked by the number of people who do use AI scribes, and the two most common reasons were privacy and the extra cost of AI scribe tools where these companies are increasing their prices, so I built something that runs locally and doesn't need any AI subscription. This was built with Claude by @AnthropicAI , over the last two weeks. It was not a one-shot prompt. It took a lot of work, testing, and iteration. But it works, and it works on-premises. #HealthcareAI #AIScribe #LocalAI

76.6% of the harmful errors in a new medical AI benchmark were omissions. Not wrong facts. Missing ones. The study is called NOHARM (Stanford, Harvard, ARISE network). It ran 1,100 real clinical cases through tools including OpenEvidence, Doximity Ask, GPT-5.6 Sol and Claude Fable 5, with about 13,000 physician annotations of patient harm. As an oncologist, this matches what worries me most. A hallucinated drug is easy to catch because it is wrong, and I notice it. But I don't use AI to tell me what is wrong! I use it to tell me what I am missing. My most commonly used prompt when I use AI scribe and AI clinical decision tools is "What would you do differently for this patient?" OR "What did I miss?" It also means our usual safety check, “does this answer contain anything false?”, is aimed at the minority of the harm. This is something you can't build a benchmark for. One more finding worth keeping: physicians using AI still delivered better care than those without it. So how do we audit what a tool left out? I have not heard a good answer yet.

I used to lose whole evenings building slides for journal club, tumor board and conference prep. So we built a tool that does the first draft. You write one prompt. It pulls the studies from PubMed and builds the deck. Every number on a slide comes from a real paper, and the slide points to the exact study, so you can check it in seconds. These slides came from a single prompt. It's still a first draft you own and edit. But it gets you most of the way there. The goal isn't prettier slides. It's getting your evenings back and generating accurate slides for presentations. Free inside OncoPulse. meducationai.com/tutor/signu… #Oncology #HemOnc #MedicalEducation #AI @MeDucationai

Johns Hopkins just partnered with OpenEvidence, which reports 1.12 million license-verified US clinicians using it as of September. The line I noticed was not the headline. It was Hopkins’ vice dean for education raising “cognitive offloading”: trainees may lack skills expected of them when they practice unsupervised. I’m a practicing oncologist, and this is the version of the problem I think about in clinic. A fellow (and even me) can now get a guideline-concordant answer in seconds. What is harder to see is whether they could have built that answer themselves (I am worried about this skill; this is no joke), and whether they would catch it being wrong on a rare case. Hopkins is floating a move from time-based to competency-based training (the evaluate doctors based on what they can do, not how much they spent in the hospital). That seems right to me, but competency has to be measured with the tool turned off at least some of the time. The other issue is hospitals are not incentivized to graduate fellows earlier even if they met the competencies because they are cheap labor, so this competency thing I am not buying it, and I might be wrong. Financial terms were not disclosed, and use cases are still being defined, so this is early. Should board exams and fellowship evaluations assume AI access, or deliberately test without it?

Drawing a black box over text in a PDF does not anonymize it. The words are still there. Select all, copy, paste, and it all shows. Swap the patient's name, age and date before you upload anything to AI.

Meta's new AI agent, Muse, can negotiate your internet bill and cancel subscriptions you forgot about. To do it, it needs access to your Gmail and calendar. Helpful assistant, big privacy question. Would you give it that access?

In radiation oncology, the system that makes the plan never gets to check the plan. The planning system calculates the dose. Then a separate, independent check verifies it before the patient gets a single fraction. Different tool, different eyes. Nobody calls that bureaucracy. It's how you catch the error the first system can't see. Now look at how we use AI. The same model writes the code, then writes the tests that pass it. The same chatbot gives you a citation, then tells you the citation is fine. Self-grading doesn't catch errors.

"You don't need to learn to code anymore" is horribly wrong. At the least, learn to read code. CS basics are what let developers do more with AI, not less. A Harvard professor makes the case in 5 minutes: piped.video/vhgI_F-WMwA

A physician with AI did no better than one without. The model alone scored 92%. Grain of salt, these studies are hodgepodge. Still, who goes first matters. Human first, AI is a second opinion. AI first, the human is a proofreader. physicianintheloop.org/where…

The White House just renamed AI "superintelligence" in federal documents. Tech leaders signed a safety pledge with four layers of checks, but it's voluntary. Not a law. Should AI be regulated? Yes or no?

Claude Code sessions can message each other and hand off context mid task. Every person can become a manager of a team of agents now. That's a leadership skill we haven't had to learn yet, and we're about to.

Join us as we share insights on effectively communicating with patients on how to use AI for their health

This is not a newsletter. A newsletter is written once and sent to everybody, and most of it is not for you. The agent is yours. It watches your topics and nobody else's, and on a quiet day it says so. Free. meducationai.com/tutor/signu…

The advice that you can build an app with one prompt is fake. AIs can create very good websites and apps, but it’s not one prompt. You must be very patient and make small changes. You should have a AI reviewer or review it yourself, yet these tools are still unbelievably good.

Have you ever wondered what VS Code actually is? I taught myself coding as a busy heme/onc fellow, now I build software with startups. My new course walks you through front end, backend, GitHub, VS Code, and deploying it, no coding background needed. udemy.com/course/build-apps-…