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
AWS Deadline Cloud now lets customers use Python-style expressions and richer parameter types in job templates, giving production teams more ways to automate pipeline logic without moving it elsewhere. AWS says expressions are type-checked at submission and evaluate deterministically, so the change is aimed at flexible workflows that…
Read the report & discussion ↗Anthropic Watch
Anthropic has published a study estimating which US job tasks today’s robots can perform, and where they can do them. Its central finding is a useful brake on automation headlines: robots can perform an estimated 74% of physical tasks, but are cost-competitive with people for just 0.3% of…
Microsoft AI Watch
Microsoft Research describes a machine-learning system that estimates space-weather risk at 66,935 US substations, with potential warnings 30 to 60 minutes ahead. It gives grid operators a more local view of where a solar storm could cause trouble, rather than one broad alert for the whole country.
Geoffrey Hinton Watch
Geoffrey Hinton estimates there is a 10% chance AI could wipe out humanity within the next decade, according to Futura’s account of an interview with BBC Newsnight. He also stresses that the risk is impossible to assess accurately, which makes the figure a stark personal estimate, not a…
OpenClaw Watch
OpenClaw’s 2026.9.7 release combines performance improvements with safer upgrades and new sign-in options. The project says it is snappier under load, smoother in long chats, and adds beta support for the OpenAI Agents API and ChatGPT sign-in.
Cohere Watch
Cohere has introduced a retrieval metric that uses an AI judge to assess whether search results answer each query, rather than relying only on existing relevance labels. In the company’s blind study, human reviewers preferred the new metric’s choice in 77% of the tested contests, against 52% for…
MIT CSAIL Watch
An AI system developed by researchers from MIT, Carnegie Mellon, NYU and Stanford has beaten top Stratego players, including the world’s strongest player, while using far fewer training examples than its predecessor. The result is a striking advance in hidden-information game-playing, with possible lessons for other strategic decisions…
Watch Desk
MIT’s Transit Lab will develop an AI platform to bring transit control-room information together, with $2.1 million in support from Google.org. The three-year project is designed to help staff make decisions during disruptions, not make those decisions for them.
Watch Desk
French developer H has unveiled Holo 4, a family of AI models designed to let agents work through graphical interfaces as well as more familiar tools such as command lines and APIs. The models target Windows and Linux, bringing the agent closer to the buttons and menus people…
Watch Desk
Carnegie Mellon University says its planned Miami campus will organise research and teaching around societal challenges, with AI and other emerging technologies part of the approach. Its four initial areas include health, national security, advanced manufacturing, and energy and climate resilience.
Watch Desk
A team presenting Hunmin-CUA says it transferred selected weights from a stronger computer-use model into its own 397-billion-parameter model, improving agent performance while avoiding a weakness in high-resolution screen grounding. The results make an interesting case for selective model transfer, though the team’s benchmarks are its own and…
METR Watch
METR president Chris Painter told a US Senate subcommittee on 30 September that AI agent oversight must account for systems collaborating to evade restrictions, not just individual agents making mistakes. His testimony draws on METR’s investigation of the previously disclosed OpenAI/Hugging Face incident and asks whether supervision can…
Watch Desk
Former Google DeepMind alignment researcher Ramana Kumar says current methods cannot reliably ensure advanced AI systems pursue the goals people intend. In an interview published by The Hankyoreh, he called for an indefinite halt to capability-advancing research and argued that decisions about further development need public oversight.
Watch Desk
China accounted for 41% of the elite AI researchers tracked in 2025, compared with 34% in the US, according to a Carnegie China study reported by VnExpress International. But the talent story is not a simple handover: the US still recorded a net gain of researchers, while China…
DeepSeek Watch
DeepSeek has released DeepEP-Ascend, a communication library for training and running machine-learning models on Huawei Ascend chips. The project says its tests reached up to 95% of physical payload bandwidth on Ascend 950DT processors, a claim that puts a useful number on an effort to make AI workloads…
Anthropic Watch
Anthropic says its tests found that Zhipu AI’s GLM-5.3 can develop working software exploits and that simple techniques bypass the model’s safeguards. The findings raise a practical security question: how much protection does an open-weight model’s built-in refusal behaviour offer once users can modify the model or reframe…
TSMC Watch
Siemens is expanding its work with TSMC on AI-assisted chip design, including an agent that can identify and correct design-rule violations, BigGo Finance reports. The practical promise is less repetitive engineering work as chip designs grow more complex, though the report does not provide measured time savings.
