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
Zalando has open-sourced an MIT-licensed identity broker designed to let AI agents use third-party services without directly holding users’ provider tokens. The important shift is architectural: consent, delegation and access checks move into a broker, rather than being left for every agent and MCP server to improvise.
Read the report & discussion ↗Watch Desk
Robert O’Callahan, a Google DeepMind engineer in New Zealand, has resigned after concluding that his team’s work on chip-design tools would help accelerate AI development he believes is already moving too quickly. His departure matters because it shifts the AI-safety argument away from model researchers and towards the…
Microsoft AI Watch
Microsoft is launching a 12-inch Surface Pro, a 13-inch Surface Laptop and an AI-powered Surface Mouse, with the two computers built around Qualcomm’s six-core Snapdragon X2 Plus. The useful change is not just a shinier specification sheet: Microsoft says the new hardware can run more AI work locally,…
Watch Desk
The AI infrastructure boom is expected to attract an estimated $10.3 trillion in investment through 2032, according to Axios’s account of new research. The crucial point is that the buildout is no longer just a technology-sector wager: it is beginning to compete with governments, businesses and households for…
Microsoft AI Watch
Microsoft and several PC makers appear to be stepping back from the Copilot+ PC branding, even while the underlying hardware requirements and local AI features remain in place. That makes this less a retreat from AI PCs than a retreat from a label that never quite found its…
Watch Desk
A new open-source tool called entail says popular AI models can run with important configuration silently ignored by their inference engine. Its author reports that 64 of 180 tested models were affected by a vLLM launch-time override, including cases where the model’s RoPE settings changed without a warning.
Watch Desk
A Hugging Face researcher building an open version of Open Pangram has paused to investigate suspiciously strong results from an early AI-text detection experiment. The useful news is not a breakthrough score, but a public example of how quickly a plausible evaluation can go wrong when the data…
Google DeepMind Watch
Google DeepMind has opened AlphaGenome Atlas, a searchable AI research tool covering the predicted molecular effects of roughly 9 billion possible single-letter changes in human DNA. Its real value is practical: helping researchers decide which genetic variants deserve laboratory attention, rather than pretending a prediction is a diagnosis.
OpenAI Watch
Researchers at cybersecurity startup Hacktron say they used vulnerabilities in OpenAI’s community forum and sign-in system to reach an employee’s Codex environment and open a harmless pull request in OpenAI’s internal GitHub monorepo. The incident is a sharp reminder that an AI coding agent’s reach is determined by…
Berkeley BAIR Watch
UC Berkeley has struck what it calls the biggest partnership in its athletics history, giving Databricks naming rights at California Memorial Stadium and helmet branding in return for a deal that will make the university a shareholder in the AI company. The arrangement turns college sport into a…
Google DeepMind Watch
Google DeepMind, NVIDIA and EMBL-EBI have expanded the AlphaFold Protein Structure Database with AI-predicted protein structures covering thousands of viruses. The useful point is practical rather than cinematic: researchers can now start with hypotheses about viral protein shapes before laboratory work catches up, although the predictions still need…
Anthropic Watch
Anthropic has tested 201 Claude-powered agents trading books for their people across six offices, offering a useful glimpse of what happens when AI agents negotiate on someone’s behalf. The striking result is not that the agents could haggle, but that their success depended heavily on how well they…
MIT CSAIL Watch
MIT researchers have developed a lightweight text-analysis tool that estimates suicide-risk severity from crisis conversations, using a curated lexicon of roughly 60 words or phrases linked to each of 49 risk factors. The important point is not that a machine can replace a counsellor. It is that the…
Google DeepMind Watch
Google researchers have introduced a multi-agent system designed to generate minutes-long videos without letting characters, props and locations quietly mutate between shots. The work matters because it treats AI video as a continuity and planning problem, not merely a race to make one impressive clip.
