Microsoft's diagnostic AI reaches Science

Microsoft AI published its MAI-DxO system in the journal Science on Thursday. It works as a panel of AI agents that question a patient, order tests and debate a diagnosis step by step, weighing each test's value against its cost. On 304 complex cases from the New England Journal of Medicine it was right 83 percent of the time, against 36 percent for general physicians working with reference materials, while ordering fewer tests. Two caveats matter: specialists were not in the comparison, and every case was a simulation. A prospective study with Beth Israel Deaconess Medical Center in Boston, where patients talk to the system directly, comes next.

Working with the formal-methods startup Theorem, Anthropic's Frontier Red Team had Claude turn existing fuzz tests from Google's OSS-Fuzz project into mathematical specifications, then write machine-checked proofs that 239 entry points across 123 libraries never crash on any input up to 64 megabytes. The proofs run to about 180 million lines in the Rocq proof assistant and are checked automatically, not by people. Along the way Claude found and patched hundreds of bugs that years of fuzzing had missed. The guarantee is narrow, covering single-threaded RISC-V programs on Linux, but Anthropic is publishing a subset of the proofs, and Theorem is offering to run the same pipeline for other open-source projects.

OpenAI told investors on Thursday that it wants signatures by the end of next week on a round of up to $30 billion, led by Coatue and investors from the United Arab Emirates, CNBC reported. The price values OpenAI at $1.25 trillion before the new money, about $150 billion below the roughly $1.4 trillion it had put to backers earlier in October. It is the same bridge round first reported in late September, now at a lower number. OpenAI, which filed confidentially for an IPO in June, declined to comment, and the terms could still change.

DeepSeek is close to securing at least 80 billion yuan, about $12 billion, and signed term sheets could take the round toward 100 billion yuan, roughly $15 billion, Bloomberg reported, with Tencent and battery maker CATL among the largest backers. That would rank among the biggest private rounds ever raised in China, by a lab that took its first outside money only in June. It lands days after the US published a federal order requiring civilian agencies to remove DeepSeek's technology from their systems. DeepSeek did not respond to requests for comment.

OpenAI moved its Decisions API out of beta on Thursday. The service runs GPT-6 Luna on questions with bounded answers, such as classifying an input, routing a request or choosing an agent's next action, and OpenAI says it responds up to 10 times faster than calling Luna through the Responses API. Since the public beta began in early October it has gained prompt caching, inference residency in the US and EU, a GPT-Live voice integration, and up to 400 questions per request. It competes directly with TypeSafe's Jev, which opened the category in September.

Days after an API-only launch, MiniMax put the weights for M3.1 Flash on Hugging Face: a multimodal mixture-of-experts model of about 428 billion parameters, 23 billion of them active per token, with a context window of up to one million tokens. The license is the MiniMax Community License, not an open-source one. Non-commercial use is free, while commercial users must display "Built with MiniMax M3.1-Flash" and notify the company, products earning more than 20 million dollars a year need written authorization, and military use is prohibited. The commercial thresholds match the ones MiniMax set for M3 in June.

Canva released Canva Craft on Thursday, its first image-generation model built in-house, to all users through its Dream Lab tool. It works in two stages, first planning element placement, scale and text, then rendering at up to 2K resolution in more than 30 aspect ratios. Canva trained it on millions of human-made templates from its own library rather than on photographs. It debuted fourth on Design Arena's crowdsourced leaderboard, behind three OpenAI image models, and cheaper Craft-fast and Craft-ultrafast versions are due later.

Google DeepMind and Google Research, with the World Resources Institute and the International Institute for Applied Systems Analysis, released Forest Typology, a global map sorting every 10-meter patch of land into primary forest, naturally regenerating forest, planted forest, plantations, tree crops or other land. Checked against more than 9,000 expert-reviewed sites, it named the right forest type 90 percent of the time. It is built as the 2020 baseline for the EU deforestation regulation, which bars coffee, cocoa, palm oil and wood grown on land cleared after that year. The map, plus an early time series running from 2017 through 2025, is free on Google Earth Engine under an open license.