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
Heurist Finance is letting its AI agent buy premium market data one query at a time, with a spending cap attached to every run. Built on Amazon Bedrock AgentCore, the system shows how agentic payments can move beyond the demo stage without handing a bot an open wallet…
Read the report & discussion ↗Watch Desk
Pathway has developed an AI architecture that reasons through recurrent internal states rather than generating a long trail of intermediate tokens. Its reported ARC-AGI-1 result suggests that capable reasoning might become dramatically cheaper if the architecture, rather than sheer scale, does more of the work.
OpenAI Watch
OpenAI has defined two levels of trusted access for cybersecurity teams: Daybreak Blue for defensive work and the more tightly controlled Daybreak Red for advanced authorised testing. The practical point is wonderfully un-mysterious: buying access is not enough, Red approval does not arrive by osmosis, and both tiers…
OpenAI Watch
OpenAI has detailed GPT-Rosalind, a specialist AI model for early-stage life-sciences research across ChatGPT Enterprise, Codex and the API. The important bit is not merely a biology-flavoured chatbot: OpenAI is pitching tool-driven work across scientific data, databases and laboratory planning, while keeping access tightly controlled.
AWS AI Watch
AWS has shipped version 3.1129.0 of its JavaScript SDK, bundling a surprisingly broad set of AI and compute updates into one release. The useful thread is not a new model, but more control around the systems that feed, evaluate and run AI workloads.
Watch Desk
ToolGrad flips the usual recipe for training AI agents. Instead of starting with a user question and making an agent hunt for a solution, Google Research’s framework builds a successful tool-use chain first, then writes the prompt and response around it. That matters because agents are only as…
Jacob Coxon Watch
The public argument over AI risk has widened beyond one researcher's departure. The National Post brings together warnings from several people working on frontier-model safety, alongside a call for international coordination over automated AI development. That makes this a report about a growing debate, not a prediction that…
AWS AI Watch
AWS has published a deployment guide for Alibaba’s Qwen3.8-2.4T-A95B, showing how to serve the open-weight model on Amazon SageMaker HyperPod with vLLM. The practical headline is that a 2.4-trillion-parameter model can be made usable on one ml.p6-b300.48xlarge instance with eight NVIDIA B300 GPUs, using NVFP4 quantisation and an…
Watch Desk
Researchers have used AI alongside geological techniques to reconstruct Gondwana, the ancient supercontinent that may have covered roughly 80 percent of Earth’s landmass between 550 and 500 million years ago. The result offers a striking link between deep-Earth processes and the sudden expansion of complex life during the…
NVIDIA Watch
NVIDIA has unveiled a broad expansion of its AI for Media stack at IBC 2026 in Amsterdam, putting synthetic-video checks, live sports analysis, localisation and video enhancement into broadcast and streaming workflows.
Watch Desk
AI is not a magic tutor, according to a new analysis of OECD education data. The reported finding is more interesting than the usual “students use chatbot, civilisation ends” routine: outcomes depend heavily on what students ask AI to do and whether they are taught to question its…
Microsoft AI Watch
Microsoft reportedly fixed 972 vulnerabilities in its September security release, including 112 rated critical and two zero-days. The practical message for Windows and Microsoft server administrators is wonderfully unglamorous: patching has just become the day’s main event.
Watch Desk
Meta has introduced Muse, a US personal AI agent that connects to everyday services and completes tasks on a user’s behalf. This is Meta’s biggest move yet from chatbots that answer questions towards software trusted to send, book, buy and keep working after you close the app.
OpenAI Watch
NYU mathematician Tristan Buckmaster has accused OpenAI researchers of racing towards the same Navier-Stokes result after learning about his collaboration’s progress. OpenAI mathematics lead Sébastien Bubeck denies the claims, leaving a consequential dispute about academic priority, AI-assisted research and access to researchers’ model interactions.
AWS AI Watch
AWS has extended its managed MLflow and Amazon SageMaker AI Model Registry synchronisation to support cross-account model governance. The practical point is simple: teams can centralise oversight while keeping development accounts separate, using either a hub-and-spoke design or a hybrid model.
Google DeepMind Watch
Google DeepMind is launching AlphaGenome Atlas, an AI-powered searchable database covering the predicted impact of all 9 billion possible single-letter changes in human DNA. The point is to give researchers a much faster way to explore which variants might matter for biology and disease.
Arm Watch
Arm has announced Arm CSS for Mobile 2, an AI-native compute platform designed to bring more capable AI processing into next-generation mobile experiences. The important bit is the integration: Arm is presenting CPU, graphics, neural acceleration and system-level optimisation as one platform rather than a collection of separate…
