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
OpenAI has claimed that one of its internal models found a breakdown scenario in the Navier-Stokes equations, a mathematical result with obvious drama and surprisingly little immediate effect on engineers. The equations still do the day job in aircraft design, weather modelling and biomedical devices.
Read the report & discussion ↗Watch Desk
An AI system has reportedly designed, verified and deployed a small inference-chip design from a written specification, with Architect Labs saying the flow took under two weeks. The practical significance is not that a chatbot has magically become a semiconductor foundry, but that hardware and software optimisation may…
AMD Watch
AMD has published a fuller set of benchmark claims for its upcoming EPYC Venice server CPUs, including a result that puts its 256-core flagship well ahead of NVIDIA Vera in one SPEC CPU 2026 throughput test. The useful story is both the scale of AMD’s claim and the…
Watch Desk
TypeSafe AI's Jev is being pitched as a different kind of AI model: not a chatbot, not a reasoning engine and not a replacement for your main LLM. Theo Browne describes it as a fast classifier for structured decisions such as routing, ranking, tagging and escalation, where the…
Microsoft AI Watch
Microsoft has a new playbook for what it calls the Frontier Firm, and its central wager is blunt: generic AI output is a convincing average of everything a model has seen, and averages do not make anyone distinctive. The edge, it argues, now lies in teaching AI the…
Watch Desk
Ramin Hasani, co-founder and chief executive of Liquid AI, has a number for the corporate AI graveyard: 80% of production-grade efforts never reach production quality. He draws it from roughly 200 Fortune 500 engagements over two and a half years, recounted in a Latent Space interview. His diagnosis…
Anthropic Watch
Anthropic's threat intelligence team has laid out its fullest account yet of attempts to turn Claude into a weapon, and the case studies run from bioweapons grant drafts to a kamikaze drone swarm. By the company's own account, it disrupted every operation described.
Watch Desk
Ramin Hasani has put a number on the intelligence Mercedes-Benz plans to build into its cars: about 600 megabytes, small enough to reach the whole fleet as an ordinary over-the-air update. The figure came up on the Latent Space podcast, and its significance is less benchmark than business…
Cohere Watch
The story that Chinese AI progress is mostly borrowed homework has an official version in Washington and, as of Friday, a prominent sceptic inside the US industry itself. Cohere chief executive Aidan Gomez called Chinese AI models "world-class", said the advantage held by the leading American labs is…
AWS AI Watch
Amazon Connect Talent, AWS's agentic AI hiring solution, is now generally available, and its agents do far more than sift applications. They analyse the role, draw up an interview plan, invite candidates to book a slot at any hour, run voice interviews around the clock and hand the…
AWS AI Watch
Amazon has published a year-to-date accounting of its SageMaker inference work, and it reads less like a launch post than a decision guide. Thirteen capabilities have landed in 2026 across two deployment paths, and the review's real contribution is putting the choice between them on a single page.
Anthropic Watch
Anthropic is letting outsiders through the lab door. The company announced on 18 September a partnership with Accenture to embed independent evaluators inside Anthropic with access comparable to an employee's, and each side expects to invest at least $1 billion in building that capacity over the next five…
Hugging Face Watch
Write code, run it, read the feedback, revise. A community article published on Hugging Face today argues that the loop which made coding agents useful is now being pointed at robots, world models and editing tools, with code serving as the interface between a model and the physical…
Google DeepMind Watch
AI has made scientific ideas cheap, and the queue behind them keeps growing. A report by Google, Google DeepMind and MIT FutureTech finds that 44% of surveyed scientists say their main research bottleneck has shifted over the past two years toward physical experimentation and data collection, while 41%…
AWS AI Watch
Moonshot AI's Kimi K3 is now generally available on Amazon Bedrock, and it turns up with a first to its name: per AWS, it is the first open weight model on the platform to support explicit prompt caching. It arrives with a 1-million-token context window, native vision and,…
Watch Desk
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…
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…
