From the archive
AI research tests new methods for training, evaluating and understanding artificial intelligence. Find news on machine learning papers, reasoning, model behaviour and experiments that challenge existing assumptions. Read what researchers measured, how they tested it and where the limitations lie, with links to the original work and discussion of what the findings could mean beyond a benchmark.
Lead story
California is drawing up plans to switch off AI. Governor Gavin Newsom has issued an executive order asking state agencies to recommend stronger independent reviews of the technology and to consider a shut-off function should it get out of control, while a convened panel of experts has two…
Read the report & discussion ↗Watch Desk
A model that cannot chat, cannot code and refuses to write a sentence has become the week's most argued-about release. TypeSafe, a San Francisco startup fresh out of stealth, has shipped Jev, a model with text generation stripped out entirely: hand it information and a predefined set of…
Zhipu AI / Z.ai GLM Watch
Zhipu AI has launched GLM-5.3-FlashX on its BigModel platform, with the API live and a headline claim of up to 200 tokens per second of output. The pitch is refreshingly narrow: the new model keeps the intelligence and pricing of GLM-5.3-Flash and simply generates faster, which is precisely…
Anthropic Watch
Anthropic appears to be preparing to let Claude read your bank account. Leaked screenshots of the Claude app for iOS show an unannounced feature, reportedly called Claude Money, that invites people to link their bank accounts and ask the assistant about spending and financial plans. Nothing is confirmed…
AWS AI Watch
Amazon has told the industry's loudest running argument where it stands: powerful AI models should reach the public only when they are ready and safe, and governments belong in the room when the rules get written. That places one of AI's biggest investors, model builders and compute suppliers…
Watch Desk
Interactive practice is the part of digital learning everyone praises and nobody has time to build. Google Research has decided the supply problem is the thing to attack. On Thursday it published work on a system that turns a teacher's request into a custom, game-style learning simulation, and…
Watch Desk
A French-speaking criminal crew wired a safety-stripped AI agent into a credential harvesting line, then left the entire control room exposed on the open internet. Researchers at SOCRadar found the server: 4.9 GB across 9,299 files, holding the crew's DXSCAN scanning platform, phishing tools, extortion material and a…
Watch Desk
The argument over whether to slow down frontier AI has produced its first institutional sketch. Canada's prime minister wants a global AI stability board modelled on the body that watches the world's banks, the European Commission is offering Canada and the UK a partnership on model evaluation and…
Moonshot AI/Kimi Watch
Kimi Code has shipped version 2.0.0, jumping from the 0.43 line it occupied days ago, and the changelog's single major change is a doorway: a /desktop command that opens the Kimi Code desktop app page in the browser. Under the milestone number sits a release about maturity, with…
Cohere Watch
Nick Frosst has picked a loud week to make a quiet argument. As Cohere signed its definitive merger with Germany's Aleph Alpha, a combination valued at roughly $20bn and pitched as the first transatlantic sovereign AI provider, the Toronto company's co-founder sat for a long interview and argued…
Meta AI Watch
Meta built the smart-glasses market on a camera strapped to your face. Now it is reportedly working on a version without one, aimed squarely at the people who have spent five years saying no thanks. The stakes are subtler than one new product. Meta is the clear market…
Google DeepMind Watch
Ohio State University is about to have Google as a physical neighbour. The university has announced a partnership with Google Cloud that plants a dedicated Google space in its Innovation District, one of the first outposts of its kind on an American campus, and gives researchers access to…
Watch Desk
Dreamforce was meant to be a marketing spectacle. Instead, within hours of each other, Dario Amodei, Jensen Huang and Sam Altman took the same San Francisco stage and gave three different answers to the only question that matters this week: who decides how fast frontier AI is allowed…
Hugging Face Watch
LAION and TTS Arena have launched Voice Acting Arena, a public platform where listeners judge AI voices on whether they can act rather than merely sound real. Same scene, two anonymous performances, your verdict.
xAI Watch
A science writer asked two consumer chatbots whether AI is going to end the world, and got the week's argument in miniature. Pressed to argue against its own reassurances, xAI's Grok improvised a gleeful three-phase plan to wipe everyone out. Asked the same question straight, Anthropic's Claude gave…
OpenAI Watch
Anthropic and OpenAI want independent safety evaluators working inside their labs, and researchers are welcoming the access as unprecedented. The catch comes from the researchers too: access alone proves nothing, and meaningful oversight needs transparency, real independence and, eventually, regulation.
AMD Watch
AMD has filed its broadest MLPerf Inference submission ever, and it has arrived with fighting words attached. The company says its Instinct MI355X GPUs beat selected NVIDIA B200 and B300 submissions on the GPT-OSS-120B language model at eight GPUs, led the 72-GPU GB200 result too, and the brand…
