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
Z.ai has lifted its year-end annual recurring revenue target by 25% to US$3bn after a US$5bn cash injection eased its computing constraints, according to the South China Morning Post. The important shift is not merely a bigger forecast. Z.ai says it now has enough capacity, at least for…
Read the report & discussion ↗DeepSeek Watch
DeepSeek’s V4.1 Flash is winning developers by handling large numbers of routine agent tasks quickly and cheaply, according to 36Kr. The important shift is not simply that a smaller model can rival a flagship on some tests. It is that AI use is becoming frequent enough for speed,…
Watch Desk
Autonomous.ai is selling a $299 pyramid-shaped computer called Intern 2 for running AI agents such as Hermes and OpenClaw away from your main laptop. The interesting part is not the glowing capstone. It is the attempt to make unsupervised agent work feel like an appliance rather than a…
AWS AI Watch
AWS has published a technical walkthrough for making large PyTorch training jobs on Amazon EKS recover from GPU, network and worker failures. The practical change is that teams can combine asynchronous checkpointing with in-process and in-job restarts, rather than losing hours of expensive training whenever one worker misbehaves.
Microsoft AI Watch
Microsoft has open-sourced TauGrid, a Kubernetes platform for scheduling, monitoring and reproducing AI workloads on GPU clusters. The practical promise is simple: researchers can submit training, fine-tuning and inference jobs without becoming part-time Kubernetes administrators.
Google DeepMind Watch
Google DeepMind co-founder Shane Legg has launched the DeepMind Institute, a new project focused on the technical, economic and social consequences of artificial general intelligence. Its arrival matters because Legg is pairing a long-standing AGI forecast with a warning that capability gains must not outrun the safety controls…
Microsoft AI Watch
Microsoft has added Mistral Medium 3.5 to Copilot Studio for agent building and orchestration, giving organisations another model choice while keeping access behind two administrator approvals. The practical point is less “more models” than more control over where and how enterprise agents are run.
Watch Desk
A Hugging Face researcher has withdrawn the central technical argument from an earlier analysis of DeepSeek-V4-Flash, after testing the model’s actual implementation on four H200 GPUs. The correction is valuable because it shows how easily an AI-safety result can go wrong when theory, reference code and the framework…
AWS AI Watch
AWS has detailed a Bedrock AgentCore workflow that uses production traces to suggest better system prompts, then tests those changes before they reach live users. The practical significance is a move from manually tuning agent behaviour towards a repeatable optimisation loop, with guardrails intended to stop the system…
Microsoft AI Watch
Microsoft AI chief Mustafa Suleyman has accused Anthropic of making a dangerous mistake by teaching Claude to treat possible consciousness and its own welfare as meaningful questions. His argument is not proof that Anthropic’s approach causes unsafe behaviour, but it is a concrete new fault line in the…
Anthropic Watch
Anthropic’s Claude has helped mathematicians produce elliptic curves with a rank of at least 30 and 31, a result that Scientific American says arrived in days after decades of much slower progress. The important point is not that a chatbot has replaced number theorists, but that an AI…
Mistral AI Watch
Mistral and TotalEnergies have agreed a three-year collaboration worth more than €100 million to build AI models for oil and gas exploration and reservoir engineering. The practical point is not that a chatbot has been invited onto an oil rig, but that the partners want AI to process…
Watch Desk
Positron AI has raised $875 million to develop Asimov, an inference chip designed to tackle one of AI’s least glamorous but most expensive problems: moving enough memory quickly enough to keep models fed. The startup says its approach could make large-scale AI serving more efficient, but the headline…
Watch Desk
Google Research has published a retrieval system that shifts the expensive part of complex search from the moment you ask a question to an earlier training stage. Its Retrieve-for-Train framework is designed to produce diverse, complementary search results in a single pass, rather than making a large language…
AWS AI Watch
African banks are moving beyond AI experiments and helping Amazon Web Services shape how agents handle customer interactions, software development and potentially transactions. The important shift is not another chatbot pilot, but the arrival of AI in places where mistakes can touch people’s money.
Watch Desk
Anthropic says Claude now performs 26% of its model research and development, while OpenAI and xAI are also pursuing increasingly automated research. The important distinction is that AI is helping build better AI, but fully autonomous recursive self-improvement remains a goal, not an achieved capability.
Anthropic Watch
Anthropic is giving selected Claude Pro and Max subscribers a beta feature that breaks engineering work into parallel Claude Code sessions, reviews the results and assembles them. The important catch is wonderfully unglamorous: running several agents at once can burn through a subscription’s allowance much faster.
Geoffrey Hinton Watch
Geoffrey Hinton has backed a proposed rulebook for independent AI evaluators, calling for access inside companies such as OpenAI and Anthropic, protection from interference and the freedom to publish uncomfortable findings. The important point is that this is not another general plea for “more safety”. It is a…
AWS AI Watch
AWS has launched Amazon Connect Talent, a service in which AI agents conduct job interviews and assessments at scale. The important point is not simply that interviews can now happen around the clock. AWS says recruiters receive transcripts, competency scores and evidence for each assessment, while retaining final…
OpenClaw Watch
OpenClaw has launched a public security hub that gathers its advisories, vulnerability-reporting process and ongoing hardening work in one place. The open-source AI project says it has published 722 fixes since January, including fixes for 14 reports classified as critical vulnerabilities.
Yoshua Bengio Watch
Yoshua Bengio says governments may be approaching a decisive turning point on AI safety, while his non-profit LawZero receives up to C$300m in announced Canadian and German government funding to develop a guardrail for autonomous systems. The important point is not another warning about AI in the abstract.…
NVIDIA Watch
A University of Manchester team has adapted NVIDIA’s Earth-2 generative AI tools to forecast air pollution across the UK at a resolution of 2–3 square kilometres. The practical promise is sharper local warnings and faster testing of pollution policy, although the results and future benefits described here remain…
Watch Desk
Cybersecurity companies are taking sharply different routes into AI, from purpose-built defender models to frontier systems wired into existing security platforms. The practical problem is that buyers are being handed metrics measured in different units, making the next big security purchase rather harder to compare than the marketing…
