Drowning in bureaucracy & EHR hell? Step into the OR. Subtle and still surprising varations in renal hilar anatomy. The ureter peristalsing like clockwork. The course of the obturator nerve... Unchanged since the dawn of surgery. Untouched by spreadsheets. Untouched by admin. It's breathtaking. Lucky us.

Agree with @Cacciamani_MD a tweetorial probably wasn't the smoothest option for a reply. Contamination is a strong argument (that the frontier models might have already "known" the questions from their training set). But the authors already conceded this. The RCQ not being public is a legitimate knock. You can't audit queries you can't see. The "RCQ was only added after peer review" line is framed as a scandal, but I read it the other way. Reviewers said the contaminated benchmarks weren't enough, the authors built a clean blinded one in response. That's peer review working the way it was inteded to work. The conflict-of-interest accusation is interesting, and if documented it's serious. But it's coming from OE, the party that "lost" in this paper so it sounds like sour grapes... who knows... The most uncomfortable result is that free Google AI Overview scored as well as or better than OpenEvidence on real physician queries judged by blinded clinicians. They don't seem to have an answer to that.

A twittorial? Ironically, you at @EvidenceOpen do not seem very “open” to “evidence” after all. Why not submit a letter to the editor addressing these concerns directly? Present the methodological issues, challenge the study’s conclusions, and support your arguments with rigorous, reproducible evidence That is how science advances: through open debate and critical appraisal!

OpenEvidence responds…. 🍿

What happens when patients better understand their conditions? The relationship with the care team becomes a partnership. Outcomes improve. Fixing patient experience fixes everything.

This is one of the more important AI-in-medicine papers I've read in a while (and I read the whole paper), and the result is uncomfortable for anyone betting on specialized clinical tools. NYU Langone group put OpenEvidence and UpToDate Expert AI head to head against GPT-5.2, Gemini 3.1, and Claude Opus 4.6. The frontier general-purpose models won. Every benchmark, every dimension. (Disclosure: I wrote this post, then used AI to clean it up and make it read better. The ideas are mine) The part that should get your attention: they included free Google AI Overview as a control, and it scored as well as or better than both clinical tools on real physician queries. The specialized answer engine performed about the same as the summary that pops up when you Google a question. Now, credit where due. Building a HIPAA-compliant evaluation off 100 real clinical queries with 12 blinded physician raters is a serious undertaking, and they were honest about their own limits, which I appreciate. They flag that HealthBench is an OpenAI benchmark that GPT then won while also sitting on the grading panel, so they tell you to discount it and treat it as supplementary. That's the right call. But here's where I'd put the asterisk. The frontier models were queried through APIs (computer to computer interfaces). The clinical tools were queried manually through the browser interfaces (like you use it), with all the hidden prompts and formatting that comes with that. OpenEvidence's weakest score was clarity, not correctness. So is the underlying model actually worse, or is the wrapper just presenting answers in a way clinicians liked less? The study can't separate those, and that distinction matters a lot for what we conclude. One thing worth noting: OpenEvidence isn't something clinicians pay for. It's free for verified docs, same as Doximity's tools, though some health systems pay for BAAs. So this isn't really a "you're wasting subscription dollars" story. It's a "the specialized layer doesn't appear to add accuracy over a frontier model" story, which is the more interesting claim anyway. What I'm comfortable saying: the specialized clinical wrapper doesn't appear to produce better answers than a frontier model, and might cost you on clarity. What I'm not ready to say: that domain-specific tuning is inferior as an approach. The authors don't claim that either. This is a snapshot of a field that's moving fast, and the frontier models have already been updated since the study was done. But if you're a health system making decisions here, it's worth a close read.

😂🤣

Please repost for visibility.

Command-shift-G on a mac Paste the file or folder path 🎉brings you right to it

Has anyone else had similar problems dealing with @AnthemBCBS ?

Mark Cuban for president

This has changed everything about the way I work

Most second opinions end the same way: I almost always recommend less than the patient was told elsewhere. Other than removing cancer or clearing a blockage, doing less tends to serve patients better. Took me years to trust that.

Thoughtful review of the PROTEUS trial by @DrSpratticus First, real credit to the investigators. Pulling off a phase 3 perioperative RCT in localized prostate cancer is a massive undertaking, and easy to forget when you're critiquing from the armchair (mea culpa, I don't run trials like this). That said, Daniel is essentially pointing out (I think 🤪) that PSMA-PET catches BCR so early that "metastasis-free survival" is really just biochemical recurrence in disguise, and EFS is a lousy surrogate for OS (see his tweetorial for the reasons). The part that confuses me most: this isn't compared to standard of care. We don't give ADT with prostatectomy. So what are we supposed to do with this information? nejm.org/doi/full/10.1056/NE… @DrSpratticus @adamkibel_uro @uretericbud @daviesbj

I propose annual mandatory ECG recertification for physicians. And I immediately veto my own proposal.

We have to get BLS certifications, laser certifications, fluoro certifications. And recertify... and recertify... And yet my inbasket fills up with preop ECG reports, which I am now wholly unqualified to interpret at this point.

I just learned something new today. Either I'm late to the game (likely), or perhaps other Urologists also don't know this. Angiotensin II receptor blockers (ARBs) as a class, modestly suppress PSA levels. Worth knowing. (thought to downregulate androgen receptor protein expression)

Show up to a robot console to start my day to find the velco straps on the masters secured in tight loops. My immediate thought. I wonder who used this yesterday...

Best basic description of how AI works

sudo pmset disablesleep 1 #IYKYK

This is kind of a big deal