The brain drain continues – the U.S. has lost arguably its most important scientific awards ceremony:
arstechnica.com/science/2026...
11.03.2026 05:35
👍 102
🔁 36
💬 1
📌 4
The brain drain continues – the U.S. has lost arguably its most important scientific awards ceremony:
arstechnica.com/science/2026...
Time to pass a law that any party spokesperson of media outlet that opines with hindsight reckons on pandemic management has to show this chart in the background at all times #nzpol
Here is a full implementation of the Chinese Room using a printed copy of GPT-1 in case you have a few spare years and want to actually run the thought experiment for real, for some reason. You can also download 80 volumes of GPT-1 weights for free to do it: weights-press.netlify.app/pdfs/The-Chi...
I mean … that’s basically a defining feature of networks full of weights. Nodes can widely promote and repress other nodes based on the relationships seen in training data.
Not sure I would humanize it by saying “unconscious processing”.
When this came round last night, I said to my wife "I have an idea, but I am reluctant to inflict it on people", but I have thought of a was of showing the actor that is less confrontational. So, worst you say...
------
Your job is to pick the worst actor to play James Bond.
Prasad also made collages of doctors wearing masks to shame them and spread doubt and mistrust.
That’s the professionalism and maturity he brought to the table.
Well, it’s NEJM AI, what did you expect. 🙃
Everyone wants a symptom checker, a navigator, a chatbot – you name it – to magically direct patients to the "correct" next step in care.
Well, this is what Ada Health is claiming their symptom digester will do:
#medsky #mlsky
www.evidencetriage.com/p/how-do-sym...
Makary claims he’s headed back to UCSF, however. 😑
Makary claims he’s headed back to UCSF, however. 😑
No self-respecting university ought ever hire him again.
Unmistakable call when you have one in your yard! Always a lovely day with a visit.
when our hospital lost power two years ago, the emergency handbook said “call this number”.
… but the phone system was affected by the same power outage. 🙃
An ineffective intervention can be cost-effective if it displaces another, more expensive, ineffective intervention. 🙃
Everyone loves a good poop joke – or poop study.
Here's a look at potentially using emergency department-specific wastewater monitoring to detect and characterize viruses circulating in the community:
#medsky
www.evidencetriage.com/p/whats-in-y...
A scheduling web app, but “Canoodle” instead of “Doodle”.
A scheduling web app, but “Canoodle” instead of “Doodle”.
That is indeed unhelpful.
I also worry about the fallibility and automation bias of the so-called "human in the loop" as a mitigation strategy – it's not enough to just "hope" these errors are caught in an endless flood of AI slop, and these real issues requiring ongoing oversight and governance.
It's really a fascinating ongoing debate – it's possible patient harms by a thousands cuts where ambient summaries are inaccurate, but clinical pressures otherwise mean clinical care delayed/denied due without some kind of productivity augment .... which is more harmful?
(no, we can't hire more)
I would probably argue the ambient summaries have more leeway to be imprecise than a transcription — the human process it mimics is also known to be error-prone.
Yes - we’re on very different sides of the same coin!
I suppose it all depends on the definitions of “frequent”, the determination of “clinically important”, and the errors made by the best available clinician operator.
Some settings and documentation frameworks will favor humans, others are a scramble against competing pressures in a zero-sum game.
*never let a doctor use AI that isn't proofreading the notes
You're not wrong – 1) the transcript engine makes errors, and 2) the summarization engine makes errors. The errors are infrequent, but absolutely still at the point where mitigation strategies are essential.
They've been working on incorporating AI into healthcare for at least 50 years – so, you're a little late to object! 🙃
Can we offload some simple, routine radiology to AI? Not quite yet – and this is why uptake is so hard, because it requires workflows to be precisely designed as to augment a human-in-the-loop, not disrupt.
#medsky #MLsky
www.evidencetriage.com/p/a-radiolog...
What, your ED isn't a bloodbath of serious harms and deaths?
And this is just the harms and deaths from diagnostic error, not from run-of-the-mill procedural complications, delays to care, staff shortages, medication errors, other adverse effects ....
You'd think such a high-profile clown act would be career-ending, but, alas ....
