Newsletter · June 21, 2026
Weekly Digest 25
Midjourney launches a medical division built around a full-body ultrasound scanner, betting consumer-health imaging can shift from episodic diagnosis to continuous measurement; Z.ai's open-weight GLM 5.2 shows how fast China is closing the frontier coding gap as US labs absorb the Fable export-control fallout; and a wave of star-researcher departures from Google DeepMind raises the question of whether being the only publicly traded lab is a liability or an underrated edge.
Topics we are tracking
Midjourney wants to make medical imaging casual
Source: Midjourney — Introducing Midjourney Medical
Midjourney, the company best known for image generation, is launching Midjourney Medical, a new health division built around a full-body ultrasound scanner. The device is called Ultrasonic CT and uses sound, water, and a 60-second scan to create internal body imagery. Midjourney says the first location will open in San Francisco at the end of 2027.
Ultrasound is not MRI. It is operator-sensitive, tissue-dependent, and not equally useful across the whole body. Midjourney has shown a prototype, not a clinically validated imaging platform. The first use case also looks closer to body-composition tracking than regulated diagnosis.
A hospital scan is scarce, expensive, and usually triggered by suspected disease. A spa scan is routine, fast, and wrapped in wellness. The product is a bet that imaging can move from episodic diagnosis to continuous measurement.
This is part of a broader shift in consumer health. Wearables made heart rate, sleep, and activity visible. CGMs made glucose visible. GLP-1 telehealth turned weight loss into a managed consumer workflow. At-home labs, DEXA scans, longevity clinics, and biomarker dashboards all point in the same direction, where the body is optimised through continuous measurement.
If Midjourney can make full-body scans cheap, repeatable, and socially normal, this would be a very valuable dataset: repeated internal body maps linked to biomarkers, symptoms, drugs, exercise, diet, and outcomes. AI companies are looking for proprietary data surfaces outside the already-scraped internet. The human body is one of the largest remaining ones.
The risk is that abundant imaging can become a machine for overdiagnosis. The more often you look, the more you find. Some findings matter. Many do not. Without clinical context, consumerised full-body scanning can create false positives, unnecessary follow-up procedures, and a new market for health anxiety.
It is unclear how this impacts Midjourney's wider product offering, but it may be a sign that the company is under increasing pressure from the big labs and Chinese labs in image and video generation, and is looking for other areas to expand into.
Image and video generation have not seen the same level of durable traction as other AI categories. OpenAI has shut down Sora, Meta AI's Vibes app has not broken out, and Google's Nano Banana 2, while technically excellent, appears to have lost momentum in the online conversation.
Part of the issue may be compute budgets and ROI. Image and video generation are expensive workloads, and the return is less clear than in code generation or agents, where the value is easier to tie to saved time, workflow automation, and enterprise budgets.
GLM 5.2 shows how fast China is closing the model gap
Source: Latent Space — GLM 5.2 passes the vibe check
Z.ai launched GLM 5.2 this week, its new open-weight model for long-horizon coding and agentic tasks. The model comes with a 1 million token context window and is designed for work that runs over hours rather than single prompt-response interactions.
The release quickly turned into a public benchmark for the US-China AI race. After Elon Musk suggested that China could reach Fable-class capabilities by Q1 next year, Z.ai founder Tang Jie replied that it "won't take that long."
GLM 5.2 is being positioned directly against frontier coding and agent models from OpenAI, Anthropic, and Google. Z.ai claims the model is already close to recent Claude Opus levels on long-horizon coding tasks, while remaining open-weight and cheaper to deploy. Those benchmark claims still need independent validation, but there is a clear direction.
Artificial Analysis Intelligence Index
Claude Fable 5with fallbackClaude Opus 4.8
GPT-5.5
GLM 5.2top open-weight model
Gemini 3.5 FlashClaude Sonnet 4.6Gemini 3.1 Pro
Qwen3.7 Max
MiniMax-M3
DeepSeek V4 Pro
Kimi K2.6That puts pressure on the US labs. OpenAI, Anthropic, and Google are already dealing with the aftermath of Fable 5 and the prospect that future frontier releases may require more government review before they reach the market. In the absence of a global framework for AI safety and release standards, that creates an obvious asymmetry. US labs face more friction just as Chinese labs are pushing capable open-weight models into the market.
That does not mean the US should have no controls. But unilateral controls become less effective when comparable models are available elsewhere, especially if they are open-weight and can be downloaded, modified, and deployed by anyone. If the Chinese open-model frontier keeps improving, the US may slow its own labs without meaningfully slowing global capability diffusion.
Google's AI talent scare misses the bigger picture
Source: The Information — Nobel laureate John Jumper departs Google DeepMind for Anthropic
John Jumper, the Google DeepMind researcher who helped build AlphaFold and shared the 2024 Nobel Prize in Chemistry, is leaving for Anthropic. The move came shortly after Noam Shazeer, one of the most important figures behind Google's Gemini work and the co-founder of Character.AI, left for OpenAI.
Alphabet's share price fell sharply on Monday, with investors reading the departures as a signal that Google might be losing the talent wars to Anthropic.
Anthropic has become one of the strongest talent magnets in AI, helped by its culture, focus, and position close to the frontier. Google DeepMind, under Demis Hassabis, and Google Research, under Jeff Dean, have largely kept pace with the private labs on talent and research quality. But Google is still a public company with many large businesses outside AI. Search, YouTube, Cloud, Android, and ads still drive Alphabet's quarterly earnings. OpenAI and Anthropic are pure frontier AI companies.
What is more surprising is how much freedom Alphabet has given its AI teams to pursue long-horizon research with limited immediate revenue impact. AlphaFold, AlphaProof, AlphaGo, world models, robotics, and scientific AI have all been major technical or scientific breakthroughs, even if most have not yet translated into large direct revenue lines. DeepMind itself was long viewed as an expensive research bet. Waymo was viewed the same way until recently.
Our view is that these excursions may start to converge. Scientific AI, reasoning systems, robotics, world models, autonomous driving, and foundation models are no longer separate research curiosities. They are moving toward the same market: physical AI. In that world, Alphabet may be better positioned than the market currently gives it credit for.
The recent departures show the tension of being the only publicly traded lab. Google has to balance shareholder expectations with long-duration research. OpenAI and Anthropic have had more freedom to prioritise frontier progress over near-term profitability. That may not last. Both companies are expected to go public this year, and when they do, they will face the same problem Google already has: how to fund long-term frontier research while convincing shareholders that the returns will eventually show up in earnings.
