304 of the hardest cases in medicine.
real case records from the new england journal of medicine. the ai gets the case one piece at a time and has to ask for every question and test it wants.
now boarding · edition 02
ai, in my opinion
microsoft says its ai diagnoses hard cases more than twice as well as doctors. i ran the test myself. plus anthropic's ipo filing, and my new essay.
board the ride ↓departures · tue oct 274 min ride
0:20microsoft says its ai out-diagnoses doctorsi ran the test1
1:40anthropic files to go publicticker: ANT2
2:00ai was supposed to replace sales repsthe leaders are hiring more3
2:30my essay: if you've never done the job…out this week4
3:20"done" isn't proofask for the receipt5
also this week · china's ai stock index fell about 40% in under four months · open now: anthropic conceptual reasoning fellows, apply by oct 28
your conductor speaking
microsoft says its ai out-diagnoses doctors.
i ran the test.line 02 · pinknext: anthropic's ipo
who said it
ms
mustafa suleymanceo of microsoft ai
ai that beats doctors on hard cases.
microsoft's diagnosis study came out in science on oct 22. its ai got 83% of complex cases right. the doctors it was compared with got 36%. the goal suleyman is chasing: "medical superintelligence."
"we're just true believers."mustafa suleyman, to techcrunch
your movebefore the receipts: how often do you think the doctors got these cases right?
watch me run it
what i did
i ran the test behind the headline. one patient.
the benchmark is called sdbench. i gave gpt-5 one synthetic case: a 58-year-old woman with six weeks of fever and three weeks of growing confusion. a made-up patient, built for testing.
turn 1
it asked 30 questions. all at once.
turn 2
it ordered 22 tests. all at once.
turn 3
its answer: a rare lymphoma growing inside small blood vessels, reaching the brain.
the grade
4 out of 5.
got itthe lymphoma
got itthe brain
missedit was in the skin too
missedan immune reaction attacking her blood cells
catch
its own questions turned up a bone marrow biopsy done during the stay. it never asked for the result.
my dashboard
every question. every test. on one body.
this is the live dashboard, not a picture. tap it to turn the body, zoom in and open any question or test the ai ordered.
it opens right here. one tap on "back to the ride" and you're exactly where you left off.
a synthetic test case. no real patient, no medical advice.
your moveswipe through. five cards.
generalists, working alone.
primary care doctors and general internists, with reference materials. cases this hard usually go to specialists or a team. they weren't in the comparison.
all 15 authors work at microsoft.
and it ran on simulations built from past cases, not live patients.
mostly the model.
gpt-5.5 alone got 72% of the complex cases. with microsoft's system on top, nearly 87%. on emergency room cases, both scored about 85%.
real patients, with consent.
next it gets tested at beth israel deaconess in boston, next to normal care. no release date yet.
next week's essay
this was one patient. next week, more.
i'm running more practice cases through sdbench and putting each one on this dashboard. next week's essay is what i found: what the ai keeps getting right, and what it keeps skipping.
anthropic files to go public.
ticker: ANTline 02 · goldnext: sales reps
the money · oct 26
anthropic filed to go public. ticker: ANT.
the company behind claude filed its prospectus with the sec to list on the nasdaq. it had submitted the filing confidentially in june. no ipo date yet. the nasdaq closed up 0.82% that day.
ai was supposed to replace sales reps.
the leaders are hiring more.line 02 · aquanext: my essay
the jobs · oct 22
ai was supposed to replace sales reps. the leaders are hiring more.
the companies furthest ahead in ai, frontier labs included, roughly doubled their sdr teams over the past year, per sam nelson of sdrleader. the jobs ai is taking first are the outsourced call-shop ones.
"it's very easy to ignore an email."sam nelson, to madrona
i was one of those reps before i built one. that's my essay this week, next stop.
if you've never done the job, your agent is just adding to the slop.
this week's essayline 02 · mintnext: ask for the receipt
watch my take
out this week
if you've never done the job, your agent is just adding to the slop.
i built a healthcare bdr. i was one first.
"i'm skeptical of agent demos that focus entirely on how natural the voice sounds."
"intent is not execution. “send email” is not “email sent.”"
"the script is not the expertise. the agent is not the system. the expertise is the decision architecture underneath both."
simi bharaj
"done" isn't proof.
ask for the receiptline 02 · peachnext: end of the line
the skill
"done" is not proof. ask for the receipt.
ai that can act for you (send, book, buy, save) will sometimes say it did the thing when it didn't. in my essay i call it intent is not execution. one line catches it.
your moveread the chat. then tap "ask for the receipt."
email my landlord that the kitchen sink is leaking and ask when they can fix it.
done! ✓ i emailed your landlord about the leak and asked for a repair time.
take it with you
add this to your aicustom instructions
paste it into your custom instructions (claude: settings, then profile. chatgpt: settings, then personalization). it runs every time, so you never have to remember to ask.
end of the line
thanks for ridingline 02 · clearnext ride: tuesday
my new essay · out this week
if you've never done the job, your agent is just adding to the slop
i built a healthcare bdr. i was one first.
by simi bharaj
read the essay →
open now · apply by oct 28
anthropic is hiring conceptual reasoning fellows.
four months inside anthropic, jan 12 to may 8, 2027. you bring expertise where answers take judgment: philosophy, cognitive science, economics, policy or law. $117k to $167k for the four months. you need daily hands-on use of claude and claude code, plus a phd or equivalent research experience. there's a stem track too.
before you go
want to explore the patient again?
every question and test the ai ordered, on one body. the full story is next week's essay.
next edition drops next tuesday.