AI policy fights and unexpected hardware pivots dominated a strange, busy day. Anthropic’s CEO went to the G7 to make the case for unified AI access, days after the U.S. government banned his company’s frontier models from non-U.S. users. Midjourney, known for AI image generation, launched a full-body ultrasound scanner. And OpenAI published research showing their models can help diagnose rare childhood diseases that stumped human experts for years. The tension between building, restricting, and repurposing AI ran through every story.

Anthropic vs. Washington: The G7 AI Access Fight

Dario Amodei used a G7 working lunch in Évian-les-Bains, France, to urge democratic governments not to fragment over AI access. The timing was pointed: days earlier, the U.S. Commerce Department issued export controls blocking Anthropic’s Fable 5 and Mythos 5 models from non-U.S. users and foreign nationals inside the country. Commerce Secretary Howard Lutnick cited a reported Mythos 5 jailbreak that could bypass safety systems and expose software vulnerability data.

This didn’t come out of nowhere. Trump banned federal agencies from using Anthropic products in February after the company refused Pentagon contract terms requiring its AI for “any lawful purpose.” Amodei left France without a policy fix, but Sam Altman and Demis Hassabis backed the same message, three rival CEOs on the same side of a policy fight. French President Macron warned that U.S. power to “turn off the switch” could hurt the American companies leading the AI race. Indian Prime Minister Modi raised similar concerns about critical infrastructure dependence.

Here’s what this tells us: the AI export control debate isn’t theoretical anymore. It’s fracturing alliances and creating real diplomatic tension. Anthropic’s argument that applying the same standard across the industry would effectively halt new model deployments is either a genuine warning or a clever lobbying position. Probably both.

Midjourney’s Body Scanner: From AI Art to Medical Hardware

Midjourney, yes, the image generation company, launched a medical division and a full-body ultrasound scanner that maps the entire body in about 60 seconds. The device lowers you into a shallow pool of water where a ring of underwater sensors reads acoustic echoes and rebuilds them into a 3D map. Founder David Holz called it the first new whole-body imaging method in 50 years.

The tech runs on Butterfly Network ultrasound chips under a licensing deal worth up to $74 million. Holz targets 50,000 scanners worldwide by 2031, with a flagship Midjourney Spa (25,000 square feet near Union Square, San Francisco) opening by late 2027. The plan is to eventually hit 1 billion scans monthly. Only about a dozen people have actually gone through the machine so far, and the company hasn’t published images compared against an MRI baseline.

The scanner carries no AI yet. That’s the long-term play: frequent, low-friction scans building longitudinal body records that AI models can analyze for change over time. Outside experts flagged the obvious risk: whole-body screening of healthy people generates false positives and needless follow-up procedures. Midjourney will start with body composition maps, a wellness label that sidesteps FDA diagnostic clearance. The pivot from text-to-image tools to medical hardware is jarring, but Holz was a Leap Motion engineer before Midjourney. Sensors and silicon aren’t new to him.

OpenAI’s Science Push: Rare Disease Diagnoses and AI Chemistry

OpenAI dropped two science stories this week that go beyond demos. The bigger one: researchers from Boston Children’s Hospital, Harvard, and OpenAI used o3 Deep Research to analyze 376 previously unsolved rare disease cases. After expert review and clinical confirmation, physicians established diagnoses in 18 cases, a 4.8% additional yield on cases that had already gone through specialist analysis. Some had evaded years of expert review. The study was published June 18 in NEJM AI.

The model didn’t diagnose anyone. It produced evidence-linked hypotheses for specialists to review, and a finding only counted after qualified experts confirmed the variant, a CLIA-certified lab validated it, and the clinical team returned the result to the family. In a validation run on 51 previously solved cases, the workflow recovered the correct gene and variant in 48 of them. That’s not a parlor trick. It’s a scalable reanalysis tool for a medical system that accumulates unsolved genomes faster than it can revisit them.

The second story: OpenAI connected GPT-5.4 to Maria, Molecule.one’s autonomous chemistry lab, and gave it an open-ended goal: improve a reaction class. The system identified primary sulfonamides as a challenging target for Chan-Lam coupling, proposed using TEMPO as a mild oxidant, and ran 10,080 reactions. Mean yields rose from 16.6% to 25.2%. Human chemists confirmed the results at bench scale. Three months from first prompt to validated result. The model reviewed literature, proposed an unexpected idea, and helped design experiments. Not autonomous, near-autonomous, with humans steering throughout.

ChatGPT Gets Scheduled Tasks (Pulse Is Dead)

OpenAI launched a new Scheduled page in ChatGPT this week, letting users view, pause, edit, and delete active tasks in one place. Monitoring tasks scan the web and connected apps, alerting users only when something changes. Tasks run no more than once per hour. The older Pulse feature, last year’s daily summary tool, gets retired within 14 days. The shift is OpenAI betting that an assistant becomes stickier when it stops waiting to be summoned and starts acting on its own. Available for Plus, Pro, Business, and Enterprise subscribers on web and mobile. Free users get nothing yet.

OpenAI Partner Network: $150M for Ecosystem Building

OpenAI announced the Partner Network, a global program for partners to build, sell, and deliver AI solutions. They’re putting $150 million into it and aiming to train 300,000 certified consultants by end of 2026. Three tiers (Select, Advanced, Elite) with specializations planned for Codex, cybersecurity, and agents. The logic is straightforward: the bottleneck for enterprise AI isn’t model capability anymore, it’s deployment expertise. OpenAI can’t scale delivery alone, so they’re buying an ecosystem.

Quick Hits

Mistral AI launched Vibe, a unified agent for long-horizon productivity and coding. Work Mode handles inbox, calendar, research, and deliverables. Code Mode runs remote coding agents from the web. New VS Code extension ships alongside CLI updates. At their AI Now Summit, they also announced an industrial engineering partnership with Airbus, BMW Group, and ASML, plus a new 10 MW data center in Les Ulis opening Q3 2026.

Google DeepMind was quiet on the news front this week, but their news page shows ongoing safety work: multi-agent AI safety research investments, UK house-building AI planning, and DiffusionGemma claiming 4x faster text generation. Gemini Omni and Gemma 4 12B were recent highlights from earlier in the month.

Hugging Face had three solid posts this week. GLM-5.2 from Zhipu AI shipped with a 1M-token context and strong coding benchmarks (81.0 on Terminal-Bench 2.1, second only to Claude Opus 4.8’s 85.0). A new “Is it agentic enough?” benchmark evaluates how much work agents do to complete tasks, not just whether they succeed. And a deep PEFT comparison asking whether LoRA is actually the best fine-tuning technique or just the most popular one: 98.4% of PEFT model cards on the Hub mention LoRA.

NVIDIA hasn’t posted on the deep learning blog since CES 2026 in January, where Jensen Huang presented the Rubin platform and autonomous driving blueprints. Long quiet stretch. Could be strategic silence. Could just be quiet.

Baseten is reportedly raising $1.5 billion for AI inference infrastructure, per SiliconAngle. No lead investor or valuation disclosed. Inference is attracting serious capital as the deployment cycle matures, following Odyssey’s $310M Series B earlier in the week.

Apple isn’t an AI company source, but the Yellow coverage is too relevant to skip: Apple stock up only 10% in 2026, lagging the Nasdaq 100 by nine points after a WWDC that underwhelmed Wall Street. The rebuilt Siri arrives this fall as a beta on Google’s Gemini model running Nvidia chips, skipping the EU and China at launch. At 33x forward earnings, the premium needs results Apple hasn’t delivered.

Rundown for June 19, 2026. Sources: Anthropic, OpenAI, Google DeepMind, Microsoft AI, Mistral AI, Hugging Face, NVIDIA, Yellow.