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
Reflection AI has introduced Beam, its first open-weight model, and is pitching reasoning efficiency rather than a clean sweep of the model rankings. The reported results put it near an earlier GLM model on one coding test, while newer Chinese models score higher on several comparisons.
Read the report & discussion ↗Watch Desk
Argonne National Laboratory has connected an AI agent to an X-ray microscope, letting researchers ask it in natural language to inspect areas of interest. The experimental system can collect and analyse data, reconstruct images as they arrive and flag places to investigate next.
MIT CSAIL Watch
MIT researcher Christina Delimitrou is applying machine learning to make data centres more efficient, reliable and secure. Her work targets a practical problem: getting more useful computing from existing equipment could reduce wasted energy and pressure to build more capacity.
Watch Desk
Jamie Dimon says AI will eliminate some jobs, but argues businesses and governments should phase in the transition and help affected workers. His example of automated lorries is hypothetical, but the policy choice behind it is real: who absorbs the cost when technology changes work faster than people…
Watch Desk
Singapore’s financial regulator wants firms to have every AI use case independently reviewed before it goes live. The guidance also calls for ongoing monitoring, including of third-party AI, and says firms remain accountable for AI used in the services they provide.
Meta AI Watch
A phishing campaign is using a fake Meta Muse Ads website to trick advertising workers into handing over account credentials and multi-factor authentication codes, The Register reports. The site is one of several AI-branded lures built on the same operation, a reminder that a convincing login window can…
Watch Desk
TypeSafe AI has emerged from two years in stealth with Jev, a model designed to give software typed decisions and confidence estimates rather than conversational answers. Its pitch is a different kind of AI tool for automation, with the company claiming striking speed and cost advantages on its…
Google DeepMind Watch
Google DeepMind’s Philipp Schmid says its managed agent environments let developers give an AI agent a persistent place to run code, install tools and continue work across requests. The practical pitch is fewer bits of infrastructure for developers to assemble themselves, and a shift from chatting with a…
Watch Desk
The UK AI Security Institute has released Transect, an open-source tool for exploring long AI-agent evaluation transcripts. It turns a sprawling run into an interactive timeline, helping reviewers connect an agent’s activity, tool use and delegation to the evidence behind the analysis.
Watch Desk
Unity has announced Spark, a browser-based tool for making and sharing games with AI, and says it will not require prior game-development experience. Kotaku’s account of the promotional trailer shows a workflow built around prompts, ready-made Unity assets and tools for directing gameplay. The harder question is whether…
Watch Desk
The White House is set to announce more than $1 billion in industry commitments for its Genesis Mission at an AI science summit on Thursday. The plan also includes a multi-state computing hub and more than $100 million in federal funding for AI science tools, putting research capacity…
Watch Desk
Generative AI helped patent attorneys draft, but a three-month study found no average improvement in junior lawyers’ professional judgement. The results raise a practical question for firms: does AI help newer staff learn the work, or mainly help them produce it faster?
OpenAI Watch
OpenAI says it has launched GPT-6 Sol and GPT-6 Luna globally for free and paid ChatGPT users, replacing GPT-5.6 models. Its safety hub classifies both as high capability in cybersecurity and biological and chemical domains, a consequential designation alongside the company’s claimed safety improvements.
Stanford HAI Watch
Stanford HAI researchers found that paralegals used an AI tool 58% more, and experimented with it 70% more, when managers framed it as a way to make work more meaningful rather than simply faster. Their study suggests that training, time to explore and measures that recognise higher-value work…
Anthropic Watch
Anthropic has launched Claude Haiku 5.5, a small model it says is designed for high-volume, cost-sensitive work such as summaries and classification. Its published API pricing starts at $0.10 per million input tokens and $0.50 per million output tokens, with higher rates for prompts over 100,000 tokens.
OpenAI Watch
OpenAI publicly released a Lean-verified proof of a variant of the 2-to-1 games conjecture, after theoretical computer scientists hurried to publish their own work amid rumours of an AI proof. The episode puts a concrete question about AI’s place in mathematical research on the table: what changes when…
