What I did not see coming: Midjourney, the team behind the famous AI image generator, is making a retro-futuristic ultrasound device and, just like that, also became a medical company.

What I did see coming: the backlash from the biotech and medical community. You could set your watch to it.

Sure, it’s important not to get carried away. Midjourney is pitching 60-second MRI-like imaging, a San Francisco spa and up to a billion scans a month. That deserves scrutiny.

But some of the objections also sound a little too familiar. One is the old fear that indiscriminate screening will create a flood of false positives and extra work for doctors. It’s ironic isn’t it: these industries often complain about the lack of data. And if devices like this ever work, the point will not be humans staring at every scan. It will be AI systems parsing repeated measurements over time.

Another objection is the physics of ultrasound. Sound struggles with bone and air; it does not pass through the body like MRI or CT. But the story is not quite as simple as “ultrasound cannot do this.” Low-frequency 3D ultrasound and 360-degree tomography are some of the attempts to work around those limits. Midjourney’s scanner also has a direct scientific lineage to human ultrasound-tomography work published this spring in Nature Biomedical Engineering.

Maybe the skeptics are right. Often they are. Maybe this will never work at body scale.

Still, I think it is pretty neat that a company best known for AI images - and for all the angry AI-slop criticism that came with them - is now trying to make us healthier by reimagining an old technology.

In this issue we also have OpenAI helping rare-disease teams revisit unsolved cases, PICASSO dissecting pathology models, Benchling giving AI hypotheses private lab context, AI-ranked drugs extending worm lifespan, and AlphaFold’s co-creator heading to Anthropic.

Let’s dive in.

Midjourney, the AI image company, has announced Midjourney Medical, a new division around the Midjourney Scanner: water-immersion ultrasound that lowers a person through a sensor ring and reconstructs the results into body slices and a 3D map.

The pitch was bound to trigger a cycle of hype and backlash on social media: 60-second MRI-like imaging, saunas, cold plunges, a San Francisco spa in 2027, then 50,000 scanners by 2031, up to a billion scans a month. Easy peasy?

Strip away the golden glow of the marketing material and the idea is less magical. Ultrasound is cheap, safe and fast. Water immersion is old too: early systems put patients in water baths, but slow mechanical scanning lost out as electronics improved and contact ultrasound with gel became standard.

Midjourney believes that chips, sensor arrays and a whole lot of compute can make the old tech shine again - at body scale.

A prototype uses 40 Ultrasound-on-Chip modules from Butterfly Network, the Burlington, Massachusetts company behind handheld ultrasound probes, under a co-development deal. Together they form a 70-centimeter ring with about 358,000 ultrasonic elements, each roughly 200 microns wide, firing up to 1,000 times per second and captures 17 gigabytes of data per second. Software reconstructs and labels the slices.

But then there’s the physics to contend with. Ultrasound struggles with bone and air. Many of the things people expect from full-body imaging sit behind ribs, spine, skull, lungs or bowel gas. Sound can still reach many useful regions, especially soft tissue, and a 360-degree view can reduce shadowing by looking from many angles. But it does not pass through those barriers the way MRI or CT can.

The timing is interesting: in April, a Caltech group published whole cross-sectional ultrasound tomography in living people.

This is not a coincidence. Jinhua Xu, co-first author of the Caltech paper, is now a research scientist at Midjourney and says he has spent the past year and a half continuing his graduate work on ultrasound tomography in the company’s scanner. Co-first author David Garrett also spent several months with the team, while another paper author, Yousuf Aborahama, now works at Midjourney. The Caltech system and Midjourney prototype use different hardware, but the scientific lineage is direct.

The paper points to earlier piglet work showing that low-frequency 3D ultrasound can image despite bone and air. But that older system imaged 10-12 day-old piglets, not living adult humans. It also took 15-25 minutes to acquire data and another 20-25 minutes to reconstruct images, a problem for living subjects who breathe and move.

Caltech’s contribution was to make the approach faster and more human-scale. Its system used a 60-centimeter water-immersion ring, 512 receiver elements and a rotating transmitter to capture complete abdominal and leg cross-sections. In healthy volunteers, it produced 10-second abdominal slices and visualized organs including the liver, stomach, spleen, aorta, kidneys and vertebral body. The authors say the 360-degree geometry reduced shadowing from bone and possible air pockets.

Midjourney is starting with body-composition maps, not diagnosis, and diagnostic uses would need FDA clearance. Midjourney founder David Holz told Business Insider it is not currently using AI, though it hopes to later “to improve everything across the board.”

Why it matters: The realistic promise, for now, is repeatable imaging: compare you to you, week by week or year by year. Screening can save lives when the target, population and follow-up are well chosen. Historically, the problem with indiscriminate whole-body screening has been that it can uncover many harmless abnormalities and trigger anxiety, biopsies and more scans. But that debate often assumes humans will be staring at every image and chasing every odd finding. If cheap, frequent imaging is paired with AI systems that track change over time, the value may come from having more data points: not just “what looks unusual today,” but what is growing, shrinking, stable or newly appearing.

Did you know? Midjourney Medical is hiring.

OpenAI’s o3 Deep Research helped rare-disease specialists at Boston Children’s Hospital and Harvard revisit 376 previously unsolved cases - and physicians established 18 diagnoses after expert review, additional testing and clinical confirmation. The findings are published in NEJM AI.

The key word is “revisit.” Diagnosis does not end when a genome comes back negative. The patient’s DNA is mostly the same, but the evidence around it changes: new gene-disease links appear, variants get reclassified, and old cases sit in databases while the literature moves on.

The model received de-identified clinical and genomic information and produced evidence-linked hypotheses, not diagnoses. Human experts checked the reasoning, applied standard clinical genetics rules, and ordered more testing where needed.

The yield was 4.8%. That may sound small, but these were cases that had already resisted standard genetic analysis and expert review. The reanalysis produced 18 confirmed diagnoses. Seven were rediscoveries: diagnoses made elsewhere, but missing from the records the team reviewed.

Why it matters: Rare-disease diagnosis is often a search problem spread across genomes, symptoms, papers, variant databases and years of changing medical knowledge. Human specialists remain essential, but no team can continuously reprocess every old unsolved case against everything the field has learned since.

Did you know? OpenAI Foundation is funding the Manton Center’s next step: a platform-agnostic genetics AI copilot for rare-disease teams.

A new preprint introduces PICASSO, a framework that turns the hidden representations inside pathology foundation models into thousands of visual concepts that pathologists can inspect and manipulate.

Pathology foundation models read ordinary H&E slides - the pink-and-purple tissue images hospitals already make - and convert image patches into embeddings, numerical summaries used to predict tumors, mutations or prognosis. We’ve covered slide-to-signal models many times before, from GigaTIME to Noetik’s TARIO-2.

PICASSO asks: what did the model see?

The University of Washington and Cold Spring Harbor team trained PICASSO on more than 120 million tissue patches across 32 cancer types. It breaks each embedding into concepts - lymphocytes, spindle cells, tumor glands, tissue folds - then uses a diffusion model to show what happens when one concept is turned up or down.

Two examples stand out.

☑️ Pathology slides are thin tissue samples on glass. Sometimes the tissue wrinkles or folds during preparation. That mark is visible to the AI, but it is not biology - and a model can mistake it for a cancer clue. PICASSO could find that fake clue and switch it off. In one test, that made the model catch far more real tumors.

☑️ In lung cancer, PICASSO pointed to a specific tumor-cell morphology that may explain what the model is picking up when it predicts EGFR mutation status from an ordinary pathology slide. That finding needs validation.

Why it matters: Pathology models are moving toward clinical infrastructure. PICASSO tries to make their internal visual vocabulary legible, editable and biologically grounded before those systems are trusted with harder decisions.

Did you know? Code is promised for academic use after journal publication.

Scientific agents have read the papers. Now they need the lab notebook.

Benchling, the San Francisco company that makes software tools for biotech labs, has launched Hypothesis Generation. The feature generates experiment ideas from both public literature and a customer’s own experiment history - including failed assays and program decisions that never make it into papers.

Benchling says the feature can search the web, look across internal data and run multiple model providers in parallel. Scientists can ask for ideas around a target, a selectivity problem or an in vivo result, and get hypotheses grounded in both public and private context.

“A hypothesis worth running is novel, specific, testable, and relevant to your exact scientific goal, not just to biology in general. What you need is an AI Scientist that can generate high quality hypotheses based on your science,” writes Nicholas Larus-Stone, Head of AI at Benchling.

Why it matters: The public literature can suggest what might work. Internal data can show what a team already tried, what failed, and which hypotheses fit the exact program in front of them.

Did you know? Hypothesis Generation with web search is available now for Benchling customers.

Six drugs selected by machine-learning models trained on mouse lifespan data significantly extended median lifespan in C. elegans, the tiny roundworm used in many aging studies, according to an updated preprint from João Pedro de Magalhães and collaborators.

The first version, posted in 2024, was a computational screen. The team trained models on DrugAge, a curated database of compounds tested for lifespan effects, using its mouse studies as the starting point. They then ranked drugs from DrugBank as possible lifespan-extension candidates.

Why start with mice? Because mouse data is closer to human biology than worm data - but much harder to get. Mouse lifespan studies are slow, expensive and scarce. Worms, by contrast, are fast enough for a first experimental check.

In the updated version, the team tested 22 AI-prioritized compounds in worms. Six significantly extended median lifespan, with effects ranging from 7% to 28%.

Why it matters: The worm results do not prove the drugs will work in mammals. But they show that a model trained on mouse lifespan studies can find candidates that survive a quick organism test. For a field where mouse lifespan experiments are expensive and years-long, that could become a useful triage layer.

Did you know? The team has launched a public web server with its top classifier ensembles.

John Jumper, the Google DeepMind scientist who co-created AlphaFold and shared the 2024 Nobel Prize in Chemistry, is leaving for Anthropic.

His role has not been disclosed. Jumper says he will take some time to recharge before joining. But the signal is hard to miss: one of the clearest proof points that AI can transform biology is moving from the lab that built AlphaFold to the lab behind Claude.

This comes during a broader AI talent scramble. The same week, Google’s Gemini co-lead Noam Shazeer - one of the authors of Attention Is All You Need, the 2017 paper that introduced the transformer architecture - left for OpenAI. Frontier labs are competing for the few people who have already built field-defining systems.

The Anthropic timing is especially interesting. It comes amid all the Fable 5 furore, where biologists were frustrated that ordinary biology questions were routed away from Anthropic’s strongest public model.

Not long after Anthropic suspended access to Fable 5 and the more restricted Mythos 5 after a US export-control directive. The controversy has since widened: The Economist reported that a senior US senator said NSA and Cyber Command chief Gen. Joshua Rudd had told him Mythos broke into “almost all” of the NSA’s classified systems in hours during a red-team exercise.

Why it matters: Anthropic is becoming both gatekeeper and magnet. Its biology safeguards have angered researchers who want access to frontier models. Its hiring suggests the company still wants serious science talent inside the building. That tension may define Anthropic’s role in AI x bio: who gets to use the models, and who gets to build them.

Did you know? Want to be a Nobel laureate’s colleague? Anthropic is hiring.

Michaela Lie’s AI x bio indexes are worth keeping open in another tab. The Biological Foundation Models index tracks 323 models and resources across sequences, molecules, structures, omics, cells and tissue images. The Autonomous Science Agents index tracks 344 systems and benchmarks across literature, hypothesis generation, experiment planning, lab automation, analysis and reporting.

☑️ “What would a Figma for science look like?”

☑️ Nature asks whether AI is already deskilling doctors and coders.

☑️ Takeda will deploy BoltzMol-1 and BoltzProt-1 across R&D.

☑️ Towards Conversational AI for Disease Management.

☑️ Eric Xing argues many “AI agents” are just engineered workflows in disguise.