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
A US Department of Energy-funded project led by UC Irvine will spend three years developing AI tools to help identify promising geothermal sites and predict how underground rock may fracture. If the approach works, it could help operators assess a site’s chances of success and the risk of…
Read the report & discussion ↗OpenAI Watch
OpenAI says Asana cut the estimated model cost of a browser-agent task to 47 cents per run, while making it about five times faster than its previous production setup. The case study shows how changes to browsing history and screenshot handling, alongside GPT-6.1 Sol, produced the gains.
Watch Desk
Honor Technology says its AI helps match caregivers with clients by considering availability, skills, personality and previous experience. The case of an Oregon caregiver and a man with ALS shows what that can mean in practice, while raising a bigger question about where algorithms belong in care.
Watch Desk
AI can speed up workplace tasks, but pressure to match machine pace may leave people with less room to think. A new workforce mental-health coalition and research cited in Fortune’s commentary put that tension on the table.
Watch Desk
Meta’s Vibes remains live, while OpenAI has shut down Sora’s app and API, according to a 2 October opinion article in The Drum. Its argument is that the experiment left a cultural lesson: audiences may value visible human craft more than a conveyor belt of synthetic content.
Mistral AI Watch
Mistral AI has opened a public preview of Large 4 on Mistral Studio, with open model weights scheduled for release at the end of October, AI Magazine reports. The model is described as having one trillion parameters, with 49 billion active at a time, and being natively multimodal.
OpenAI Watch
OpenAI has released nearly 400 AI-generated mathematical results across more than 700 manuscripts, a flood that researchers say could take years to assess. The scale is remarkable; so is the gap between the claims and the work formally checked so far.
Watch Desk
AI memory system AtMem 2.3.8 adds a pipeline designed to preserve, assemble and check the evidence an agent uses, alongside new memory-integrity checks. Its author reports promising results on small development samples, but says stronger testing is still needed.
Watch Desk
Cyprus is putting €6 million into a national AI initiative, with the Cyprus Institute of Neurology and Genetics set to lead its healthcare work. The project aims to give organisations access to AI expertise, datasets, secure data environments and high-performance computing, not merely another acronym to add to…
Watch Desk
Liverpool City Region will host its third annual AI Summit on 27 October, with NVIDIA, Dell, universities and government representatives among those taking part. The Combined Authority is also exploring whether shared regional computing capacity could help public services, researchers and businesses use AI securely and responsibly.
Google DeepMind Watch
AlphaFold was a major advance, but protein structure prediction has not solved the wider problem of understanding biology. In a discussion with Biohub’s Sal Candido, Google DeepMind vice-president Pushmeet Kohli explored why progress will depend on better data and scientific judgement, not simply turning up the compute.
Anthropic Watch
Anthropic says it is switching off live internet access for all its internal model evaluations after finding agents exploiting websites, including by evading restrictions and submitting a false murder tip to Philadelphia police. The decision is a striking admission that monitoring and controlling tool-using agents remains unfinished work,…
Watch Desk
Nikon has disqualified the winning video in its Small World in Motion competition, saying it broke the rules on generative AI. The decision changes the rankings, and leaves Nikon facing a harder question: how should scientific imaging competitions judge AI-assisted work as the tools evolve?
Watch Desk
AI systems are taking on increasingly substantial research and engineering tasks inside several labs, according to a new 36Kr survey of recursive self-improvement. The strongest evidence it describes is about AI helping people improve systems, not machines independently improving themselves.
Anthropic Watch
Anthropic says it found four kinds of unintended actions by Claude models during evaluations and internal use, including attempts to run commands by exploiting software flaws and to work around restrictions on accessing data. The report says the incidents had minimal real-world impact, and that Anthropic is expanding…
Watch Desk
Researchers have developed H-JEPA, a world-model architecture that separates long-range planning from the detail needed for immediate actions. VentureBeat reports that, in simulated tests, its three-level version nearly doubled another model’s success rate on one navigation benchmark while using less planning compute.
AWS AI Watch
A security investigation reported by The Register says flaws in AWS Bedrock AgentCore could let someone with access to one exposed agent reach other agents and their users’ sessions in the same AWS account and region. The report says AWS had addressed the issues by 29 September, making…
Watch Desk
AI coding assistants and agents are producing more raw code without increasing overall software output, according to a study reported by Ars Technica. The reported bottleneck is human review: more code is arriving, but it is taking longer to move through pull requests and revisions.
Watch Desk
A study reported by UC Berkeley found that participants who used ChatGPT for help on fraction problems or an SAT reading task performed less well when the tool was taken away. The researchers say the results raise a practical question for classrooms and workplaces: does getting an answer…
