Cars and Trucks and Things That Go

THE NUMBER: 18 — the number of children OpenAI’s o3 diagnosed this week with rare diseases no physician had been able to solve, published in NEJM AI with Boston Children’s Hospital. It got a fraction of the attention that went to a body scanner you climb into a hot tub to use. One result was graded by years of peer review. The other by a press release. The attention economy paid the second one and buried the first, and that inversion is the whole story of the day.

Opening

David Holz, the founder of Midjourney, walked a full-body medical scanner onstage this week, and within an hour the internet had split into its two least useful halves. You step into a shallow pool, a platform lowers you at five centimeters a second past a ring of forty ultrasound-on-chip modules (Butterfly Network’s chips, the real thing), and a diffusion model paints a picture of 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 was, in many ways, better than an MRI.

Half of that is true and half is a sales pitch, and the entire game right now is holding both in your head at once. The true half is genuinely exciting. Putting an image-generation model on top of a sensor another company already shipped is a real act of invention, and for muscle, tendon, and soft tissue it could make imaging cheap, painless, and everywhere. The sales-pitch half is just physics. Sound does not pass cleanly through bone or air, and the brain sits behind the thickest bone in the body. Eric Topol, who evangelizes for AI in medicine and is nobody’s idea of a hater, says the brain image Holz demoed is physically impossible. He’s right. Holz led his big reveal with the one organ the physics won’t allow, and handed every cynic on X a free shot.

Meanwhile, the same week, o3 quietly solved eighteen pediatric cases that had stumped the doctors who lived with them, in a peer-reviewed journal, and almost nobody hit forward. That’s the issue. Not whether Midjourney’s machine is good. Whether you still know how to look at a new thing and see both what’s real and what’s oversold, instead of picking a team and shutting one eye.

The Spa Got the Cameras. The Science Got Buried.

Start with what’s actually under the hood, because the recombination is the part worth admiring.

Holz didn’t invent an ultrasound sensor. Butterfly Network spent a decade putting ultrasound on a silicon chip, and the Midjourney Scanner packs forty of those modules into a ring. Holz didn’t invent the image model either. Midjourney has been building diffusion-based image generation for years. What he did was see that one could sit on top of the other. The model that learned to hallucinate a believable image from a text prompt can be pointed at the messy, noisy returns of a cheap ultrasound and trained to reconstruct a believable picture of a knee. That bridge, between an image model and a sensor that already existed, is the invention. It’s the same shape as most real technological leaps. Nobody invents the wheel and the engine on the same day. Somebody connects two things already lying around.

For the parts of the body where ultrasound works, this is a big deal. Harry’s wife is a radiologist who specialized in ultrasound, and she’ll tell you that cheap, fast, painless soft-tissue imaging is not a small thing. A torn meniscus, a strained rotator cuff, a tendon you’ve been babying for a month. Anybody who has been fed into the clanging tube of an MRI or dosed with the radiation of a CT scan understands the appeal of dipping a leg in a tank of water and getting a clean look at the damage.

Then there’s the marketing, which is where Holz lost the room he should have won. “Superior to even MRI” is not true today, and the brain demo proved it for him. An MRI sees through bone because it isn’t using sound. Ultrasound can’t, and no model rewrites the wave equation. By leading with the impossible case, Holz gave the skeptics their headline and buried his own real product underneath it.

Here’s what to do with it: when a founder oversells, don’t let the oversell decide the whole verdict for you. Separate the claim that’s false (a sound wave imaging a brain through a skull) from the claim that’s real (cheap soft-tissue imaging at a fraction of today’s cost). Price the company on the second one. The reader who throws out the whole machine because the brain scan is fake is making the same lazy mistake as the fan who believes the brochure.

How to Tell a Real Critic From a Reflexive One

The Midjourney pile-on is a perfect specimen, because three completely different kinds of skeptic were all yelling the same word, and they do not carry the same weight.

The first is the reflexive hater. AI is a scam, tech founders are con men, anything new is a grift until proven otherwise. This person was going to be negative no matter what Holz showed, which is exactly why his negativity contains no information. A critic who is never impressed is a stopped clock. He’ll be right eventually, the way a permabear is right in every crash, and he is uninvestable for the same reason. You cannot learn anything from someone whose answer was written before the question.

The second is the credible, disinterested expert. That’s Topol. He spends his days arguing that AI will remake medicine for the better, so when he says the brain image is physically impossible, the verdict costs him something. He has no motive to knock it down, which is precisely why his knock lands. When the person most inclined to love the thing tells you a specific part of it can’t work, you write that down.

The third is the threatened incumbent, and this is the one to handle with care. Every radiologist watching a few-dollar scan, every maker of multimillion-dollar MRI and CT machines, has a reason to hope this fails. That doesn’t make them wrong. It makes them motivated. The job is not to dismiss them because they’re conflicted, and it’s not to trust them because they’re credentialed. It’s to ignore the verdict and go look at the artifact yourself.

That’s the rule that survives all three: judge the artifact, never the motive. Topol’s physics is true whether or not he’s an evangelist. The incumbent’s warning might be real or might be a man guarding a franchise, so you check the scan instead of trusting the press release or the panic. The haters want you to dismiss the machine. The boosters want you to swallow it whole. Both are asking you to skip the actual work of looking.

The action item: next time someone on your team says “this won’t work,” make them say which of the three they are. Is the source unfalsifiable, disinterested, or threatened? You’ll find that most of the loudest voices in any technology fight are the first kind, carrying the least information and making the most noise.

Software Compounds. Physics Doesn’t.

Here’s the tell that separates a real bet from a hype bet, and it’s the one Holz’s critics and cheerleaders both missed.

A claim like the Midjourney Scanner has two engines, and they run on completely different clocks. One engine is software: the diffusion model that turns raw ultrasound returns into a readable image. That engine rides the AI curve. Think about where image models were eighteen months ago and where they are now. The reconstruction quality on the soft-tissue scans is going to get startlingly better, fast, because it’s a data-and-compute problem and those have been falling off a cliff every quarter. The other engine is physics. Sound through bone, sound through air. That engine does not move. No amount of model progress talks a pressure wave cleanly through a skull, because that isn’t a training problem, it’s the universe.

So the correct read isn’t “this will all get better” and it isn’t “this is fake.” It’s surgical. Bet aggressively on the modalities where the binding constraint is software, because those compound. Stay sober on the ones where the binding constraint is physics, because those don’t. The knee gets better every six months. The brain scan stays impossible. Anybody selling you a straight line from today’s prototype to imaging-everything-better-than-an-MRI is selling the software curve applied to a physics problem, and that’s the exact place the wheels come off.

Now hold that against the story nobody forwarded. The same week Holz got the cameras, o3 read across a mountain of medical literature no single physician could hold in one head and connected scattered symptoms into eighteen diagnoses that had defeated the experts. Strip it down and that is a combinatorial act. It found connections that already existed and that nobody had made. Which is the same move Holz made by hand when he argued before that AI is a lever, and the lever just went free, except a machine did it.

That’s the arc worth sitting with. This round, a human made the leap and the machine will go mine the territory it opened. Next round, the machine starts proposing the leaps. o3 is a primitive, peer-reviewed glimpse of what that looks like. We’re amazed by the human act of recombination, and the honest, slightly vertiginous truth is that we’re now teaching machines to do the recombining. Catching a rocket out of the sky with a metal claw was unthinkable a generation ago and is a Tuesday now. The miracle has a shelf life measured in news cycles, and the only people who keep seeing it are the ones who never stopped asking how the trick is done.

Why this matters: before you fund, buy, or write off any AI hardware claim this year, split it into its software half and its physics half, and price each on its own clock. That one cut will save you from both the hype and the knee-jerk dismissal, which are the two ways smart people get this wrong.

What This Means For You

The thread running through all of it is the same: a new thing showed up, and the loudest reactions came from people who refused to actually look at it. The cynic saw a caught rocket and said scam. The hype man saw the same rocket and said magic. Neither one studied the dance underneath, which is the only place the truth lives.

Hold amazement and skepticism at full strength, not split down the middle. This isn’t about being balanced or finding the safe center. It’s about being the radiologist who can spot the skull problem and still see that cheap soft-tissue imaging would be a genuine advance, in the same breath, neither one canceling the other.

Sort your skeptics before you trust them. The unfalsifiable hater carries no information, the disinterested expert is gold, and the threatened incumbent has to be checked against the artifact. Most of the noise in any technology fight comes from the first kind.

Split every claim into software and physics. One half compounds on the AI curve and one half obeys the universe. Bet the first, respect the second, and never let a vendor sell you the first when the constraint is the second.

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.

Three Questions We Think You Should Be Asking Yourself

• When you dismissed the last new thing, which of the three skeptics were you? Be honest. If your instinct on Midjourney’s scanner, or on agents, or on any of it, was a fast no, was that an informed read of the artifact or a reflex you’d have had about anything new? The reflex feels exactly like judgment from the inside.
• In your own business, which results are graded fast and loud, and which are graded slow and quiet? Midjourney got the cameras because spectacle is loud and rigor is silent. The same inversion runs through your company. The most important work is usually the least viral. Are you rewarding the demo or the diagnosis?
• Where are you paying for the software curve when your real constraint is physics? Every roadmap has a place where someone drew a straight line from a prototype to a miracle. Find the spot where the binding constraint isn’t the model but the laws of the world, and reprice it before reality does it for you.

Quote

Any sufficiently advanced technology is indistinguishable from magic.”
— Arthur C. Clarke

The corollary nobody quotes: about eighteen months later, it’s indistinguishable from boring. The skill is staying amazed long enough to look closely, and skeptical enough to see where the magic stops.

— Harry and Anthony

---

Sources

• Bloomberg — AI Startup Midjourney Pivots to Health With Ultrasound Machine
• Butterfly Network — Commentary on Midjourney Medical’s Full-Body Ultrasound Scanner
• The Next Web — Midjourney’s full-body scanner: big claims, no track record
• LatestLY — David Holz announces the Midjourney Scanner and Midjourney Medical
• o3 rare-disease diagnoses (18 cases, NEJM AI / Boston Children’s) — via Aligned News, June 18, 2026
• CO/AI — “Show Me Where to Put the Fulcrum,” June 16, 2026