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
Documentation examples such as your-domain.com and yoursite.com are being used in malvertising campaigns, researchers at Manifold Security say. The practical warning is simple: a harmless-looking placeholder copied from code, documentation or an AI agent skill may now lead some visitors towards scams rather than a blank page.
Read the report & discussion ↗Watch Desk
Databricks is putting its name on the University of California, Berkeley’s football stadium under a multi-year partnership announced on 24 September. The deal matters beyond the branding: it ties a major data and AI company more visibly to the university where its roots, research and co-founders are based.
Anthropic Watch
Anthropic’s chief economist, Peter B. McCrory, says the company is exploring whether AI can be steered towards augmenting workers rather than replacing them. The awkwardly useful part is that he does not pretend the answer is known, or that augmentation comes without displacement.
Watch Desk
A new Hugging Face guide argues that developers should inspect how AI agents find, choose and use tools, not merely whether the final task succeeded. That matters because an agent can abandon one product after repeated errors, complete the job elsewhere and still receive a misleading green tick.
Google DeepMind Watch
Google DeepMind and Wiz have launched Scan for Good, a programme using AI with human security researchers to find and help fix exposed systems across public services, critical infrastructure and non-profits. Its early case studies include a hospital alert system, a rail operator, a national archive and a…
AWS AI Watch
Aderant has put an AI ticket-triage system into production for its SierraOps team, using Amazon Nova Lite through Amazon Bedrock to gather context, classify support requests and route work. In its first 2.5 weeks, the company says the system reviewed 109 tickets with approximately 96% routing accuracy, while…
Watch Desk
AI is helping Endura Therapeutics analyse experiments, build research tools and screen hundreds of disease targets, according to Adrian Sanborn in Latent Space. The important shift is not a miraculous robot scientist, but a practical one: researchers can now explore far more options before committing scarce laboratory time.
Meta AI Watch
A security researcher says Meta’s Muse desktop AI client for macOS contained an unpatched flaw that let locally running software redirect voice-dictation traffic to an attacker-controlled server. Meta has since deployed a hotfix, according to InfoQ, but the episode shows how an AI agent’s broad permissions can turn…
Watch Desk
Liquid AI has released an experimental draft model that it says can accelerate its LFM2.5-VL-3B vision-language model by up to 3.13 times on an Apple M5 Max. The important bit is not merely a larger benchmark number: the release brings speculative decoding to a model designed to understand…
Watch Desk
Tencent will shut down its QClaw AI assistant on 24 December 2026 and is directing users towards WorkBuddy, a significant retreat from its OpenClaw-based local-agent experiment. Existing users have a migration window, but QClaw data is due to be deleted three months later, so this is more than…
Watch Desk
Contrastive-LM has released CLM-8B, an open Apache-2.0 model that scores candidate actions and returns probabilities instead of writing text. The practical significance is its attempt to make agent decisions faster, cheaper and easier to run locally, with the model’s head weighing 75MB and designed to run on one…
Watch Desk
Aetina has launched industrial MXM accelerator modules built around NVIDIA RTX PRO Blackwell Embedded GPUs, bringing generative and physical AI hardware into compact systems for robots, manufacturing and medical equipment. The practical significance is edge deployment: more AI processing can happen near the machine, rather than making every…
Mistral AI Watch
The European Space Agency and Mistral have signed a Letter of Intent to explore AI applications across space engineering and Earth observation. The significant bit is not a shiny demo but a proposed framework for putting European AI models and infrastructure to work around sensitive scientific and operational…
Watch Desk
Gortex, a new developer tool, turns code repositories into a searchable knowledge graph for AI coding agents. It runs locally as a single static binary, which means the useful bit is not another cloud dashboard, but a way to give agents structural context without shipping the repository to…
Watch Desk
Professor B Ravindran of IIT Madras argues that fears about AI ending humanity often skip the crucial middle steps: access, permissions, testing decisions and human control. His interview offers a useful counterweight to the current slowdown debate, without pretending frontier AI is harmless.
Meta AI Watch
Meta researchers have proposed MaD-RL, a reinforcement-learning framework designed to control the distribution of an AI model’s outputs rather than simply maximise the score of each individual answer. The practical promise is broader and more useful than another leaderboard number: models could be trained to preserve a chosen…
Watch Desk
A Hugging Face Community Article describes a CPU-only pipeline that turns raw genomic sequences into a structured, analysable corpus for biological AI research. The useful bit is not another grand claim about models, but the plumbing: 32.4 million records are enriched with sequence length, GC content, coding status…
Watch Desk
Researchers at MIT’s Senseable City Lab are arguing that visual AI could help planners understand traffic, pollution, public space and greenery across entire cities, while warning that the same systems can deepen surveillance and social bias. The useful point is not that computers can now look at more…
