Desk
Broad artificial intelligence coverage.
Microsoft president Brad Smith told a BBC podcast that the time has come for global AI safety rules and that the UK is better placed than the U.S. or EU to lead, citing its AI Security Institute’s pre-deployment testing. He called for breach alerts, outside evaluator approval and kill switches.
At the OCP Global Summit, NVIDIA contributed Vera Rubin rack and tray designs through its MGX architecture, released DSX software as open source, and joined Google, Meta and Microsoft on a specification for how giant AI loads should behave during grid disturbances.
Microsoft announced Maia 300, its third-generation in-house AI chip, saying it will become the primary inference system for its products from 2027. Microsoft claims 40% more tokens per megawatt than leading GPU systems, a figure not yet independently measured.
A roundup of twelve live translation products shows the category is real and fragmented: consumer earbuds, meeting add-ons, conference platforms, and APIs, with language counts that still aren't comparable.
TII's 1.6B Falcon-ASR leads Arabic and Emirati speech benchmarks in the institute's own tests — and ships without weights, a license, or an API, only a sixty-second demo.
Cloudflare's new open-weight Clef-omni scores text, images, audio, and video in one decision call — a cleaner agent routing path, with weaker hard-reasoning scores than its text-focused siblings.
At HumanX, PolyAI CTO Shawn Wen and Otter CMO Alex Gay told TechCrunch voice AI has not had its breakout moment. Full-duplex speech exists; the hard parts are fast reasoning, accurate recognition and disclosure that callers are talking to a machine.
.png)
Google Labs’ Playground lets anyone build a working browser game by describing it in plain language. A WIRED reviewer with little coding experience made three playable games in a day, but found imprecise iteration, usage limits and growing sameness across public games.
OrcaRouter's new cyber model posts near-ceiling scores on vendor benchmarks and ships with a million-token context. Useful for red teams if the numbers hold; a problem if they do not.
In a long X post, Microsoft CEO Satya Nadella argued that closed and open-weight frontier models should be handled like insider risks: contained first, trusted last, and always subject to a human who can pause or shut them down mid-task.
Sakana AI stopped grading AI reviewers on how human they sound and started grading them on whether they catch planted mistakes. That is the right test, and the results are encouraging and humbling at once.
Microsoft’s CEO wants every powerful model to run inside a separate harness, with external safeguards, tamper-proof action logs and a guaranteed human ability to pause or stop it mid-task. Where that brake lives, who controls it and how fast it acts will decide whether it matters.