Midjourney’s New Scanner Could Change Everything

ELI5/TLDR

Midjourney, the company famous for making AI images, just announced a full-body medical scanner. You stand in a warm tub of water, get scanned by sound waves for 60 seconds, and out comes a 3D map of your insides — no radiation, no giant magnet, no claustrophobic tube. The trick: ultrasound is cheap but blurry, so they plan to use AI to sharpen the blur into something as good as an expensive MRI. If it works, you could get scanned cheaply every few months and catch diseases before you ever feel sick.

The Full Story

Why an image company is building a body scanner

On the face of it, this makes no sense. If you’d asked which AI firm would jump into healthcare, you’d guess OpenAI or Google, not the people who make pretty pictures from text prompts. But the video’s central argument is that Midjourney has never really been in the picture business. It’s in the reconstruction business — rebuilding reality from incomplete information.

Before Midjourney, founder David Holz created Leap Motion, which reconstructed the position of your hands from sensor data. Midjourney reconstructs images from noise, and now this new company is attempting to reconstruct the interior of the human body from ultrasound signals.

Three projects, one underlying problem: take messy, partial data and infer the clean, full picture behind it.

How the two technologies actually work

An MRI works by magnetism. It drops you into an enormous magnetic field that lines up the protons in your body like compass needles, pulses them, and listens to how they wobble back. The wobble carries information. It’s beautiful physics — and it needs superconducting magnets, baths of liquid helium, and a machine that costs millions and lives in one or two hospitals per city.

The Midjourney device uses sound instead. Picture a ring of more than 40 tiny speakers that buzz back and forth, sending sound into the water around you; the water carries it into your body, and the echoes that bounce back get read as slices. It’s the same principle as the ultrasound used on pregnant women — just wrapped around your whole torso instead of aimed at one spot. (It stops at the neck; nobody wants to hold their breath underwater while their head gets scanned.) The scanner reportedly packs two petaflops of compute and produces 3D body-composition maps in about a minute.

The catch is physics. Sound is fussy. Air blocks it (hence the water for good contact), bone casts acoustic shadows you have to correct for, and the resolution is simply worse than MRI out of the box.

The bet: let AI make up the difference

This is where the analogy lands. Think of Tesla’s approach to self-driving versus everyone else’s. Rivals bolt on lidar — lasers that physically measure the world. Tesla uses cheap cameras plus heavy AI compute to reconstruct the 3D world. Midjourney is making the same wager in medicine: cheap, noisy sensors plus enormous data and AI reconstruction, instead of expensive, precise hardware.

If you start with a noisy physical image of an actual human body, the goal here is to recreate the reality of what that would look like if it was denoised.

That’s literally what their image models already do — start with noise, end with something believable. The novelty isn’t AI-enhanced ultrasound (the FDA has already cleared systems trained on millions of scans). The novelty is doing a whole-body scan cheaply and harvesting a dataset so large the reconstruction keeps getting better — and better at flagging anomalies worth a second look.

The three things that could kill it

Physics. Ultrasound starts behind MRI on resolution and penetration. The presenter thinks Midjourney’s AI chops likely already have this on a good trajectory — they’ve apparently worked on it in secret for over a year.

False positives. Scan 100 million healthy people and you’ll find countless harmless cysts, benign lumps, and anatomical oddities that look alarming. Flag too many and you create a flood of needless follow-up MRIs and surgeries — and the boy-who-cried-wolf problem, where someone ignores the one warning that mattered.

Medicine really isn’t defined by its technology, it’s defined by its outcomes.

Regulation. Tellingly, they’re launching as a wellness / body-composition product, not a diagnostic one. Two reasons: they need the data, and diagnostic approval means years of FDA trials, validation, and brutal liability exposure — a false negative (“you’re fine”) that turns out wrong could bankrupt the company.

The upside, and the odds

If it works, medicine shifts from reactive (wait for symptoms, compare yourself to population averages) to longitudinal and proactive (scan yourself every few months at the corner pharmacy, compare yourself to your own past, catch trouble before it has a name).

The presenter’s probability estimates:

First SF scanner by 2027: 80%+
Wellness version succeeds: 60–70%
Reaches MRI-comparable imaging in 5–10 years: ~30%
Becomes FDA-cleared diagnostic platform: 20–25%
Delivers civilization-level impact: 10–15%

His framing: even a 10% shot is worth taking because the payoff is so outsized. Failure changes almost nothing; success changes how humanity understands its own body.

Key Takeaways

The device is a water tub ringed with 40+ on-chip ultrasound sensors; you’re submerged to the neck and scanned in ~60 seconds, producing a 3D body-composition map. No radiation, no magnets.
MRI images by magnetic resonance — aligning protons in a huge magnetic field and reading their wobble. It needs superconducting magnets and helium cooling, which is why it’s expensive and rare.
Ultrasound is cheap and safe but physically limited: air blocks sound (water solves contact), bone casts shadows, penetration varies, and resolution is lower than MRI.
The core thesis: replace expensive precise hardware + proprietary software with cheap sensors + massive datasets + AI reconstruction — the same noise-to-image trick Midjourney’s generative models already do.
Direct parallel to Tesla’s camera-plus-AI self-driving versus lidar: substitute compute and data for costly precision hardware.
AI-enhanced ultrasound already exists and has FDA clearances; what’s new is doing it whole-body, cheaply, at population scale to build an ever-improving dataset.
The “false positive trap”: scanning millions of healthy people surfaces benign anomalies; over-flagging causes unnecessary procedures and erodes trust in real warnings.
Strategic launch as a wellness/body-composition tool (not a diagnostic) sidesteps FDA hurdles and liability while collecting the training data needed for later diagnostic claims.
A false negative is the existential legal risk — telling someone they’re healthy when they aren’t could bankrupt the company.
The promised shift is from reactive medicine (population averages, post-symptom) to longitudinal medicine (you vs. your own baseline over years).

Claude’s Take

The underlying insight — that Midjourney is fundamentally a reconstruction company, not an image company — is genuinely the smartest thing in this video, and it reframes the announcement from “bizarre” to “coherent.” The noise-to-signal continuity from Leap Motion to image generation to ultrasound denoising is a clean, persuasive throughline.

That said, this is a reaction video built on a product announcement, two tweets (Robert Scoble, Mark Kuchman), and a prototype held together with C-clamps. There’s no working diagnostic system, no published validation, no peer review — just a demo and a vision. The presenter is admirably honest about this; his own odds put MRI-comparable imaging at 30% and real impact at 10–15%, which is appropriately humble. He doesn’t oversell, and he names the right killers (physics, false positives, regulation).

A couple of caveats he glosses over: AI “reconstruction” of medical images carries a specific danger that generative image models don’t — a model trained to produce plausible anatomy can hallucinate plausible-but-wrong tissue, which in diagnostics is worse than blur. And the data-flywheel story assumes ground truth; you only know the AI was right if you can compare against MRI or biopsy, which is exactly the expensive step they’re trying to avoid. The wellness-first framing is also a tell — it’s the move you make when you can’t yet back diagnostic claims.

Scoring it a 6: clear, well-structured, the central analogy is sharp and the probability calibration is refreshingly honest. It loses points for being speculation on top of a press event, with the heavy lifting (does it actually work) entirely unresolved. Good explainer, no verdict possible yet.

Midjourney's New Scanner Could Change Everything