Cars and Trucks and Things That Go

Both Things Are True

Stuck at the gate this weekend, I watched the runway for an hour and thought about Richard Scarry. Busytown: the marshallers waving their wands, the fuel bowser, the bag train, the tower stacking arrivals, every little animal doing one job so the whole thing moves. I'm fifty-something and I still think airplanes are a miracle — not because I'm naive but because I actually run through everything that has to go right. That habit, looking at the dance under the surface, is the only thing worth bringing to a week like this one.

Because David Holz, the Midjourney founder, just walked a body scanner onstage and the internet split into its two useless halves. Step into a shallow pool, sink five centimeters a second past a ring of forty ultrasound-on-chip modules (Butterfly's chips, the real thing), and a diffusion model paints your insides in about a minute. No radiation. A few dollars a session. Holz called it the first new whole-body imaging method in fifty years and said it beats an MRI.

Half of that is real and half is a sales pitch, and you have to hold both at full strength. The real half: putting an image-generation model on top of a sensor someone else already shipped is a true act of invention, and for muscle, soft tissue, and a torn meniscus it could make painless imaging cheap and everywhere. My wife, a radiologist who specialized in ultrasound, will tell you that's not nothing. The sales-pitch half: sound doesn't pass cleanly through bone or air, and the brain sits behind the worst bone there is. Eric Topol, an evangelist for AI in medicine and no one's idea of a hater, says the brain image Holz demoed is physically impossible. He's right. Holz led with the one organ the physics won't allow and handed every cynic on X a free shot.

Here's the tell. The model half gets better on the AI curve; eighteen months from now that reconstruction will be startling. The physics half doesn't. No model talks sound through a skull. So bet the house on the soft-tissue scans, stay sober on the rest, and ignore anyone who can only do one.

That's the whole game now, and almost nobody plays it. The cynic sees a rocket caught by a metal claw and says scam. The hype man sees the same rocket and says magic. Neither one looked at Busytown. Neither did the work. And the same week Holz got the cameras, OpenAI's o3 quietly diagnosed eighteen kids with diseases no doctor had cracked, in a peer-reviewed journal, and almost nobody hit forward.

Be amazed and stay skeptical. It's the same muscle, and whoever only has one is telling you more about themselves than about the machine.

At COAI today: the full Signal/Noise — the three kinds of skeptic, why the Butterfly partnership is the real story, and why o3 mattered more than the spa — is live at getcoai.com.

Outsider Labs Advisory →

18. The number of children OpenAI's o3 diagnosed this week with diseases no physician had solved — published in NEJM AI, with Boston Children's, peer-reviewed and quiet. The same week, a body-scanner-in-a-spa got the headlines and a brain image its own physics can't produce. One result was graded by years of review; the other by a press release. The attention economy paid the spectacle and buried the science. Look where the cameras aren't.

Upgrade your skeptics before you trust them. When you hear "this won't work," sort the source. The reflexive hater is unfalsifiable and carries zero information — he'll hate the next thing too. The disinterested expert, Topol on the skull, is gold. The threatened incumbent, every MRI maker watching a few-dollar scan, might be right or might be guarding a franchise. Check the artifact, never the motive.

Split every hardware claim into software and physics. Before you fund or buy any AI-device pitch, draw the line. The software half — the model, the image reconstruction — compounds fast, so eighteen months changes everything. The physics half does not. Bet aggressively on the first, sit on your hands for the second, and price the demo to match which half it's selling.

Go read the thing the cameras skipped. o3's eighteen diagnoses got a fraction of Midjourney's coverage because rigor is quiet and spectacle is loud. Make it a standing rule this week: the result with the slowest, most boring grading process is usually the one that matters. Put the buried paper on the team agenda, not the viral demo.

Thursday, June 18

Midjourney built the anti-MRI, then oversold it. Holz's scanner pairs Butterfly's ultrasound chips with a diffusion model for a 60-second, few-dollar full-body image. Real for soft tissue, science fiction for the brain. The technology is a genuine advance; the "superior to MRI" line is the problem. (Full analysis above.)

o3 cracked eighteen cases nobody could. OpenAI's model diagnosed eighteen children with rare diseases that had stumped their doctors, published in NEJM AI with Boston Children's. The most important AI-in-medicine story of the week, and it drew a tenth of the scanner's attention. (Full analysis above.)

OpenAI raided Google's brain trust in a single day. Noam Shazeer, a transformer co-inventor, left Google for OpenAI despite a $2.7B retention package, and ex-White House AI adviser Dean Ball joined as Head of Strategic Futures. OpenAI is stacking the lab bench and the policy desk at the same time.

Day six, and the best American model is still dark. Anthropic's Fable 5 ban hit its sixth day with no deal; 76 cybersecurity experts call Washington's "prove it can't be jailbroken" standard impossible. You can't prove a negative, so the model may be un-shippable on those terms. Don't build a critical workflow on a single sovereign-gated frontier.

Trump put Intel in Apple's supply chain. The President said Intel will fabricate chips for Apple in America, and INTC jumped 10%. Tim Cook, meanwhile, calls price hikes "unavoidable" on a hundred-year flood in memory costs. Your device price is now set in Washington as much as in Cupertino.

— Harry and Anthony

Sources: