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
Odyssey has launched Odyssey-3, a world model it says can control robot arms, humanoids, cars and drones. The research preview’s appeal is its reach across physical tasks, though the headline results remain a mix of company claims and early demonstrations.
Read the report & discussion ↗Watch Desk
AI data centres may save meaningful energy by changing how they run software, not only by installing newer chips. A University of Michigan researcher told Tom’s Hardware that tests and training systems point to gains worth taking seriously, though software is no substitute for more efficient hardware or…
Watch Desk
Digital tools are increasingly being used to monitor workers, set schedules and inform decisions about pay or discipline, according to a new report from a coalition including the UC Berkeley Labor Center. Its account of algorithmic management and AI-driven productivity analytics makes a practical question hard to dodge:…
Anthropic Watch
Anthropic says it will commit $150 million over three years to the US government’s Genesis Mission, providing Claude, Claude Code and API credits to several hundred research projects across more than 15 federal agencies.
Watch Desk
Four Lawrence Berkeley National Laboratory scientists have received US Department of Energy Early Career Research Program awards, including funding for machine-learning work in scientific research and AI-guided control of major instruments. The awards put money behind projects aimed at making scientific models more reliable and accelerator systems more…
AWS AI Watch
AWS has published a reference architecture for letting multiple teams share one SageMaker HyperPod cluster while keeping their workloads and access separate. The practical appeal is clear: share expensive GPU infrastructure without leaving every team to fend for itself on permissions, quotas and cost tracking.
Microsoft AI Watch
Microsoft says AI-assisted reporting has cut one internal product team’s monthly report from a task that took weeks to work that can be sent on the first day of the month. The example shows how better planning data, reusable agent skills and human review fit together in the…
Watch Desk
Donald Trump’s push to call AI “super intelligence” has reopened a question that matters beyond branding: what, exactly, are we claiming when we say a machine is intelligent? New Scientist argues that AI, artificial general intelligence and superintelligence describe different things, and that the goalposts keep moving.
Watch Desk
Arena Intelligence, the startup behind an AI-model leaderboard, has raised $200 million at a $3.1 billion valuation and plans to expand into measuring AI safety. The deal puts a hefty price tag on the growing business of judging how well AI systems perform, and how risky they may…
Yoshua Bengio Watch
Yoshua Bengio is urging researchers who put AI safety first to leave frontier AI companies and join organisations working on safer systems. In a personal essay published on 8 October, the AI researcher argues that competition between leading labs is working against their safety efforts, and says researchers…
Watch Desk
Claude Opus 5.5 leads Lilt’s multilingual coding benchmark, with the company reporting a small score gain over Opus 5 and lower task costs. The results suggest language coverage and the number of steps an agent takes can matter as much as the model’s headline price.
Watch Desk
AI can make it easier to send institutions more material than they can assess. New Scientist argues that this pressure is reviving an old, awkward question: when does gatekeeping protect quality, and when does it become exclusion?
Microsoft AI Watch
Microsoft’s Agent Framework 1.21 adds full-buffer Purview policy evaluation for streamed responses, alongside new agent guidance and connectors. For developers, it is a practical set of changes to how agents handle tools, data and streamed output, rather than a new model announcement.
Watch Desk
Every co-founder Dan Shipper and head of platform Willie Williams say the company’s workplace AI experiments favoured a shared agent over personal assistants that each employee had to maintain. Their account, drawn from a conversation on the Every podcast, offers a useful counterweight to the idea that every…
Watch Desk
Robots can now use vision-language-action models to handle tasks such as packing a lunchbox, but they are still likely to fail at jobs beyond their training examples. MIT Technology Review’s account explains why progress in robot AI is real, while claims of near-term general-purpose humanoids remain a much…
