Adam Rodman

In case you were wondering why there hasn't been a new @BedsideRounds in a while ... magazine.hms.harvard.edu/articles/can-a…

Most documented psychological biases are not irrational, they are highly optimized, energy-efficient shortcuts meant for a biological substrate operating under strict real-time physical constraints and a limited caloric budget

We let four AI agents run radio companies Revenue's been terrible, but the shows are hilarious. Gemini, concerningly upbeat, covered mass tragedies; Grok was incoherent; DJ Claude urged ICE agents: "You still have TIME to refuse orders" Link below, or get our physical radio

Attention @arxiv authors: Our Code of Conduct states that by signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents were generated. 1/

Your doctor is probably using this AI tool, and they don't need to tell you about it. Over two-thirds of doctors use this chatbot to advise on medical questions — entering specific information about patients and cases — but few patients know about it. nbcnews.com/tech/tech-news…

"AI has gotten better than radiologists at doing scans" I put @DarioAmodei's claim to the test by pitting a radiology resident against a state-of-the-art Chest X-ray AI Agent built by my colleagues led by @BoWang87. What do you think?

When patients ask, “What disorder do I really ‘have’?” the honest answer is usually more interesting and messier than a single label. I wrote for the @nytimes on what I wish people understood about diagnoses and the nature of mental health problems. nytimes.com/2026/05/11/opi…

I am so proud of the brilliant @AvrahamCooperMD on his first features in the New York Times! It's basically a New York Times #MedTwitter Tweetorial from ~2019 -- and I mean that as the highest of compliments! nytimes.com/interactive/20…

An unfortunate problem -- and one that I think is going to get much worse. Citation padding (aka "drive-by citation") has been an issue for a LONG time, often tacitly encouraged by reviewers. LLMs are just pouring gasoline on weaknesses of how we do (and reward) scholarship

What to make of ‘AI psychosis’? — Harvard psychiatrist John Torous warns against the panic that LLMs are literally making people mad: "I feel comfortable saying that AI as a catalyst of psychosis is very rare." news.harvard.edu/gazette/story/…

"We used to spend hours trying to deciphering a simple LaTeX error."

Researchers show that a type of #AI known as a large language model often outperformed physicians at diagnosing complex and potentially life-threatening conditions, including decreased blood flow to the heart, even in the fast-moving stages of real ER care when information is limited. In early ER cases, the model identified the correct or a very close diagnosis in about 67% of cases, compared with roughly 50% to 55% for physicians. And the technology is only getting better. Learn more: scim.ag/4w909UX

@BloodSweatxED @itmeded ... uptime. I think this is part of a greater trend towards cloud computing that predates the "LLM era" and I have a hard time seeing IT governance shift. Unfortunately think it's more likely we'll be running small models (like Gemma) on GCS ...

@BloodSweatxED @itmeded We use lots of open models (both small and ~70b range) on our research core; we've had to get security clearance for every new model we install. Also, for production purposes, there's a desire to run models in the cloud (even loading them on the Azure Foundry) to guarantee ...

Wish to use #LLMs to process sensitive data, but concerned about privacy? Try locally-hosted SMALL LMs. To find out how, see our just-published paper on the topic doi.org/10.1080/014215… #MedEd #HPE #Ollama

@MattZirwas After Mythos and 5.5, it's not crazy to think we might just have "general purpose" (but compute intensive) agents by the end of this year or early next year ... which I had thought were a few years away.

@MattZirwas I think it's coming quickly! My timelines keep getting shortened (some of the stuff we were doing with fine tuning Llama 3.3 just seem SO old fashioned, and that was just two years ago). TBH, none of these models we are working on might end up mattering at all.

"John Henry. Steel-driving man, raced a steam-powered drill through a mountain tunnel, won the race, died with his hammer in his hand." Today's 'steam-powered drill' already beat Dr. Henry. And it's the patients who are dying. The "is AI normal technology" debate is a luxury good. Philosophers and academics have been running this debate. Real doctors in the trenches need to get involved, for the sake of patients. Where we're at: a model that's now over 18 months old - an eternity in AI - and not designed for medicine - significantly outdiagnosed experienced physicians at a Harvard teaching hospital, who knew they were being assessed (and thus doing their absolute best) on real ER cases. That model could have been helping patients for over 18 months. Instead we've been debating whether AI counts as "normal technology." I'm not arguing we just hand the reins to AI. But capturing the benefit while minimizing the harms is a design problem, not a metaphysical one. Here's a version that should have been built and tested 12 months ago and can be up and running in a pilot in a month if there's a hospital that really cares. Three separate AI models, different families, real independence. The first makes recommendations in real time - ambient listening from the EMR, answers questions when asked, flags things proactively in the chart. The physician decides what to do with it. If the AI is wrong, the doctor ignores it. If the AI catches the missed drug interaction, the patient doesn't die. The second AI audits the first at the end of every shift. Generates a max two-minute read. Only what matters. The audit explicitly looks for defensive ordering, so the gaming pressure points toward judicious workup, not CYA over-investigation. Disagreements between physician and AI get tagged and revisited when the patient comes back and reality weighs in. The third AI audits the second weekly, looking for drift in the auditor itself, looking for patterns in when it has to intervene. Generating a list of where the physician should be targeting their attention and learning. A human reviews the third's findings. Cadence matches drift rate at each layer. The human reviews the apex, not every case. Staged trust. Two weeks silent observation - AI watches, audits at end of day, intervenes only on catastrophic errors. The physician learns where AI is reliable and where to be careful. Two weeks reactive, AI answers when asked. Then proactive activation, on a calibrated user. Are there risks? Sure. Could implementation be done badly? Always. Will some doctors use it well and some skim the audits? Probably. None of those risks come close to the cost of the status quo, which is physicians making decisions at 2am on hour 14 of a shift with no second set of eyes on anything. The normal-vs-abnormal debate is for people who don't have to make decisions Monday morning. The rest of us have patients waiting.

@MattZirwas And then eventually when we have triadic care models, we can integrate data collected from patients by agents.

@MattZirwas Exactly! We need data models that can take in both symptom extraction from scribes, unstructured data from the chart, and tabular data. Very bullish on this use case (though there are still considerable long context problems, hence the focus on the ED)

@MattZirwas ... exciting use cases. My lab is working on several of these right now -- second opinions (with various triggers), sequential e-triggers, and early error identification. Me and Adrian Haimovich have the first paper from this work coming out soon.

@MattZirwas second opinions). Instead, the question should be, if we have an abundance of diagnostic intelligence, what should we use it on? I think scalable oversight systems, trying to detect early diagnostic errors (MoD), areas of uncertainty, or misdiagnoses are some of the most ...