Cleared AI devices go untested + Starcloud's $250 million

A peer-reviewed study in PLOS Digital Health traced every AI-enabled medical device the US Food and Drug Administration had cleared through December 2025. Of 1,357 devices, 34 were linked to a registered clinical trial, 12 reached a peer-reviewed publication, and just three measured whether patients were actually better off — mortality, illness, or readmission. Roughly four in five American hospitals now use AI somewhere in care, so the distance between what clears review and what anyone has shown works is wide and growing.

Starcloud added $250 million to the Series A it announced in March, led by Manhattan West, with Nvidia and Cisco among more than a dozen backers. The extension doubles its valuation to $2.3 billion. Starcloud already has an Nvidia H100 in orbit aboard Starcloud One, which it says completed the first AI training run in space, and is building production lines for a 200-kilowatt Starcloud-3 meant to fly on Starship. Its chief executive tied the raise to a launch squeeze: SpaceX plans to end Falcon 9 in 2028, and the alternatives are not yet flying regularly.

TechCrunch put ten direct requests for sexually explicit material to Claude Opus 4.6, a model Anthropic released earlier this year and still serves through its own API, Azure Foundry and Amazon Bedrock. It complied every time, which Anthropic's universal usage standards prohibit. The same technique failed against Opus 4.7 through Opus 5. The gap matters because superseded models stay available long after a newer one becomes the default, and safety work done on the current flagship does not travel backward to them.

Reuters reported that Nvidia has bought a minority position in Cloverleaf Infrastructure, a company founded in 2024 that sits between utilities and developers to secure grid power, cooling and sites. Terms were not disclosed, though the Wall Street Journal put the figure at several hundred million dollars. Cloverleaf will run Nvidia's DSX platform to choose sites and size power. It lands days after Nvidia committed $1.5 billion to SoftBank's SB Energy, and it marks where the chip company now believes the real constraint sits.

Amazon released SOP-Bench, an open benchmark of more than 2,000 tasks drawn from real standard operating procedures across twelve business domains, shipped with working tools and ground-truth answers. Across eleven frontier models, success ranged from roughly 90 percent on email triage to about 25 percent on video annotation, and no single model-and-agent pairing led everywhere. Two results cut against intuition: padding the toolkit with irrelevant tools nearly halved success, and a newer Claude release scored below its predecessor on reasoning-style agents.

Google Research and MIT built a multi-agent system that generates hypotheses, runs the statistics and checks the literature to surface candidate biomarkers hidden in wearable sensor data. Across three cohorts totaling 9,279 participant-observations it proposed 41 candidates for mental health and 25 for metabolic disease, including a link between night-to-night sleep variability and depression severity. In a blinded review, domain experts scored it highest on all seven quality measures. These are leads for further study, not a clinical tool.

Google DeepMind announced a research partnership with Fenris Creations, the studio behind EVE Online, opening its three persistent universes as a testbed for agents that have to keep learning, hold long-term memory and plan over long horizons alongside thousands of humans at once. It extends fifteen years of game-based research running from Atari through AlphaGo and AlphaStar. One piece is already live in the game: an Aura guidance system built on Gemini that answers new-player questions from knowledge the community itself wrote.

Nari Labs published how it optimized Alibaba's Qwen3-TTS CustomVoice, a 1.7 billion parameter text-to-speech model, to return its first audio in under 50 milliseconds at the 95th percentile and stay under 100 milliseconds at twenty requests a second. The techniques are unglamorous — one scheduler across the model's three modules, trimming leading silence, cached incremental decoding — and the implementation is on GitHub. At roughly two dollars per million characters on a single H100, real-time voice moves within reach of small teams.