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 University of California, Berkeley team has built an open-access database of millions of local ordinances across all 50 US states, using AI to turn a sprawling archive of legal documents into searchable data. The work could help researchers and the public compare rules that are usually scattered…
Read the report & discussion ↗Aleph Alpha Watch
Aleph Alpha has detailed Kolibri’s design for German-speaking public administration and industry, including document-grounded answers, native tool calling and adjustable reasoning. The practical proposition is a model built for infrastructure customers control, rather than another assistant whose operating arrangements arrive as a surprise. Its 5 October announcement adds…
AMD Watch
AMD has begun sampling a new space-grade Versal AI Core chip for satellite and other long-duration missions, Techaeris reports. The XQRVC1902 is designed for missions lasting up to 15 years, with flight-qualified units expected in the second half of 2027.
AWS AI Watch
AWS has added Z.ai’s GLM 5.3 to Amazon Bedrock, giving eligible enterprise customers a managed way to use the large model for coding and agentic workflows. The practical appeal is a choice of APIs, service tiers and prompt caching, without customers managing the inference infrastructure themselves.
Watch Desk
IBM says it has tuned NVIDIA Nemotron models for IBM Bob, its self-hosted AI coding assistant, giving organisations a way to run agentic software development on infrastructure they control. The practical draw is local deployment for sensitive or air-gapped environments, alongside IBM’s claim of up to four times…
Watch Desk
Reflection says its forthcoming Beam model has 501 billion parameters and is designed to give businesses and governments another open-weight AI option. The founders’ account makes the story about more than a model launch: Reflection is betting that open access can coexist with a business selling the infrastructure…
Meta AI Watch
Meta’s Muse AI agent reportedly creates and updates a page for every person in a user’s life, recording details such as relationships, recurring topics and important dates. The finding raises a pointed privacy question: what does an assistant need to remember about people who never signed up for…
Watch Desk
Reka AI has introduced Rho-1, a 19-billion-parameter model built to handle text, images, video and robot-control actions in one network. The notable idea is a shared model and context for tasks that are often split across specialised systems, including work involving robots.
Watch Desk
Enterprise teams are pulling back from letting AI agents push changes on automated checks alone, while many still report customer-facing failures that passed those checks. VentureBeat Intelligence’s August survey suggests a growing gap between confidence in evaluations and the pace of agent deployment.
Watch Desk
International AI governance needs more than another round of speeches: an Oman Observer opinion piece argues for an institution able to set standards, assess frontier systems and coordinate responses when they cross dangerous capability thresholds. It also calls for mandatory risk registries and an emergency mechanism.
Meta AI Watch
Meta AI Research has published six mathematics papers involving its Muse Spark model, with Tech Times reporting that five present answers to problems described as previously unsolved. The notable claim is that researchers used Muse Spark through the ordinary meta.ai chat interface, while mathematicians directed and checked the…
AWS AI Watch
AWS has introduced a skill that lets compatible coding agents help benchmark and configure generative AI inference on SageMaker. Developers can ask an agent to measure an endpoint, compare deployment options or generate executable code, with real workload measurements rather than estimates at the centre of the process.
AWS AI Watch
AWS says Amazon Bedrock Managed Agents powered by OpenAI are now in public preview, letting customers build agents using OpenAI models inside AWS. The practical draw is that those agents can use AWS resources while remaining within the identities, permissions and governance controls customers already use.
Watch Desk
TanStack has released version 1.0 of TanStack Charts, a data-visualisation library built around composable marks, scales and interactions. Its promise for developers is a charting system they can adapt as their needs grow, rather than outgrow and replace.
AMD Watch
AMD has published new AI benchmarks for its Ryzen AI Max+ Pro 495, pitching Gorgon Halo as a rival to Nvidia’s coming RTX Spark devices. The headline advantage is memory: AMD’s system can be configured with up to 192GB of unified memory, but more room for a model…
Watch Desk
Seventeen countries have endorsed a vision linking AI-assisted science with public access to data, computing and research facilities. The Kyoto declaration also backs research-integrity standards and new approaches to funding, but leaves practical details such as data access and computing allocations unspecified.
