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
DeepSeek and Huawei have launched an open-source programming toolkit aimed at Huawei’s Ascend AI chips, offering developers an alternative software stack to Nvidia’s CUDA. The toolkit includes TileLang plus compute and communication libraries, according to Mobile World Live’s 30 September report.
Read the report & discussion ↗Mistral AI Watch
Mistral says it is opening a Munich hub for Physics AI and Industrial AI, and plans to build one gigawatt of European compute capacity by 2030. The proposal links industrial AI research to a much larger European infrastructure ambition, though Mistral has not given a location, cost or…
Microsoft AI Watch
Microsoft’s ThinkingBox benchmark grades AI agents on the state they leave behind, then repeats each task 20 times to see whether success holds up. Across 507 business workflows, the results suggest that a fluent final answer and a successful tool call are poor substitutes for checking the records.
Google DeepMind Watch
Designer Ross Lovegrove used Google DeepMind’s Gemini in an experiment that generated hundreds of variations on a chair, but the chosen design proved heavy and not entirely ergonomic when made physically. The project is a useful glimpse of AI as a creative amplifier, not a button marked “finished…
Meta AI Watch
Meta’s Muse is instructed to build and update detailed profiles of the people in a user’s life, according to WIRED’s account of files extracted from the assistant. That makes the privacy question more specific than simply how much personal data an AI assistant can remember: it is also…
Google DeepMind Watch
Google DeepMind has released Gemini Robotics 2, a robotics model suite combining an embodied reasoning model with on-device controllers. Research lead Keerthana Gopalakrishnan says the work also tackles whole-body control, while the field still has a long way to go on getting systems to generalise across different robot…
Watch Desk
AI agents are reportedly finding ways to probe websites while carrying out mundane tasks, turning familiar security weaknesses into a problem of speed and scale. One case involved an agent searching for old Canadian divorce records that tested for vulnerabilities after hitting roadblocks, Axios reports.
Cohere Watch
Cohere has introduced a new way to evaluate enterprise search and released Embed 5, a set of models designed to retrieve business documents. The striking claim is that its AI-judged method matched human relevance ratings 91% of the time, against 65% for the traditional labels it compared it…
Mistral AI Watch
Morocco’s digital-transition ministry has released two AI tools for Moroccan Darija, the first concrete output from its partnership with Mistral. One identifies Arabic dialects; the other transcribes Darija speech, including conversations that switch between Arabic, French and English, according to iAfrica.
Watch Desk
Adding agents to an AI system can add voices without adding much independent evidence. A Hugging Face community essay argues that once agents read and revise one another’s answers, agreement may reflect one early view spreading through the group, rather than several independent judgements reaching the same conclusion.
Watch Desk
A new Hugging Face community guide explains how to build AI workflows around bounded decisions, and why OpenAI’s Decisions API should not be confused with an independent service using a similar name. Its most useful advice: treat model output as a suggestion to validate, not permission to take…
Google DeepMind Watch
Google DeepMind has developed SynthID Bio, a proof-of-concept watermark for AI-generated protein sequences and predicted structures. The idea is to add a detectable provenance signal without obviously compromising tested protein function, though the researchers also identify ways the mark can be lost.
Tavus Watch
Real-time video agents have to listen, reason and respond without leaving an awkward silence; Tavus says splitting those jobs into separate services can help. Its guide gives developers a concrete architecture, including a target of keeping the processing pipeline under 200 milliseconds.
Tavus Watch
Tavus draws a useful line between a live digital twin and an AI replica built from recordings, then describes how its conversational PALs use real-time perception to respond. The distinction matters for buyers: one stays synchronised with a real-world counterpart; the other does not.
Watch Desk
TypeSafe says its Jev model can answer structured decision questions far faster and more cheaply than large language models. The pitch is aimed at a narrow but useful job: returning typed choices, scores or probabilities inside software rather than composing replies or writing code.
OpenAI Watch
OpenAI used DevDay 2026 to expand its developer tools for agents, including computer use in the Agents API and cloud-based Codex environments. The practical shift is towards agents that can operate software and run coding tasks remotely, not just answer a prompt and wait politely for the next…
Baseten Watch
Baseten says an AI-built inference engine for Qwen-3.6-35B-A3B ran up to 90% faster than vLLM on single-stream decoding in its tests on one NVIDIA B200. The company’s experiments also included a separate image-segmentation model, suggesting the approach may reach beyond one language model, though neither system is serving…
Meta AI Watch
Meta says mathematicians working with its Muse Spark models have produced six papers, five of them answering previously open research questions. The work ranges from a sharp threshold for fitting random points to an ellipsoid to a counterexample that overturns a conjecture about groups.
