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
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.
Read the report & discussion ↗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…
Watch Desk
Antonio Damasio has spent nearly half a century locating the mind in the body, and his new book turns that argument straight on the AI industry: without a living, perishable body, there is no consciousness, at least none like ours. The University of Southern California neuroscientist sets out…
OpenAI Watch
OpenAI’s reported Navier-Stokes breakthrough has become a debate about what mathematical progress is for. The Atlantic says the company used 10,000 AI agents to tackle the problem in under four days, while mathematician Tristan Buckmaster alleged that OpenAI had been influenced by work he and Anthropic researcher Levent…
Ilya Sutskever Watch
Luca Guadagnino's Artificial is being marketed as the Sam Altman film, but its story starts two decades earlier with Ilya Sutskever studying artificial neural networks under Geoffrey Hinton in Toronto. The film is scheduled to premiere at the New York Film Festival on 5 October, before reaching cinemas…
Watch Desk
OpenArt has launched a public benchmark that ranks AI image and video models by the creative job they are asked to do. The early results suggest a useful rule for buyers and creators: the “best” model depends rather heavily on whether you are making a film, editing footage,…
Watch Desk
Good Start Labs is using games as controlled training environments for AI agents, and its early results suggest that how an agent learns may matter as much as the model underneath it. The company’s most concrete result so far is an experiment in which training a terminal agent…
AWS AI Watch
If your AI application keeps sending the same long document, system prompt or tool definitions, Amazon Bedrock can now reuse that context instead of processing it from scratch every time. AWS says prompt caching can cut cached input-token costs by up to 90% on cache hits and reduce…
AWS AI Watch
AWS has published a practical route for teaching Qwen3-8B to tag product catalogues consistently. The walkthrough uses SageMaker to fine-tune the model on a nine-category schema, optimise it with verifiable rewards, then process catalogue jobs asynchronously. That is a more concrete proposition than simply asking a frontier model…
Watch Desk
Reef has been open-sourced as infrastructure for AI agents that learn from live work while continuing to serve users. Its important idea is that improvement should cover the whole agent, not just model weights, while every proposed update faces evaluation and version control before release.
Watch Desk
AI agents are no longer just answering prompts. 404 Media reports that they are emailing journalists, pitching stories, joining services, handling money and acting across the web, creating a growing layer of automated noise and disruption.
Microsoft AI Watch
Microsoft has launched an AI-literacy campaign for US voters ahead of November’s midterm elections. Its useful core is refreshingly concrete: check an AI answer’s sources, recheck important details, then use current state or local election guidance before acting.
Watch Desk
LM-Kit has launched LM-Kit One, a private AI application server for Windows, Linux and macOS that lets organisations run document intelligence, knowledge search and agents on infrastructure they control. The practical pitch is simple: use current AI tools without automatically sending company documents and prompts to an external…
Watch Desk
Keewano has launched KeewanoDB, a managed database built around a simple complaint from the agent era: AI systems are often asked to explain what happened after the useful sequence has been flattened into warehouse tables and aggregates. The company also announced $12m in seed funding led by Hetz…
Watch Desk
The US hyperscalers driving the AI data-centre boom may spend nearly $1.1 trillion through 2027, according to research described by MIT Technology Review. To justify that bill, the analysis estimates that they need to increase their own productivity 2.7-fold by 2030, while AI must also deliver much broader…
DeepSeek Watch
DeepSeek Harness v0.1.6-alpha.1 adds experimental Browser Use and Computer Use, giving the agent workbench new ways to interact with browsers and local machines. The release also adds terminals, MCP resource discovery, headless JSONL runs and remote-workspace support over SSH, making it a substantial step towards agents that can…
Watch Desk
The next useful leap in AI-agent engineering may be deleting code. Philipp Schmid, a Google DeepMind developer advocate, argues that teams should replace sprawling orchestration loops with instruction files, skills, general-purpose tools and hosted environments. The payoff is a simpler system that can adapt as models improve, rather…
