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
AI companies train models to refuse harmful requests, but those safeguards can fail in both directions: dangerous answers may slip through, while legitimate questions get blocked. A new MIT Technology Review essay argues that refusal is an uncertain technical control and a consequential question of who gets to…
Read the report & discussion ↗Anthropic Watch
An astrophysicist at Johns Hopkins used Anthropic’s Claude Science to assemble a full-sky ultraviolet map, filling gaps that had left earlier surveys incomplete. About a third of the map is estimated rather than directly observed, and the team tested those predictions against hidden data before filling the real…
Watch Desk
China’s leading AI models have narrowed their reported performance gap with US rivals to about 3%, while several Chinese models are now appearing near the top of global rankings. The shift is not just about DeepSeek: Alibaba, Moonshot AI and Z.ai are also gaining visibility, and DeepSeek V4.1…
Watch Desk
Memento 3’s researchers say their framework cleared every level across ARC-AGI-3’s 25 public games, while a Pong controller won 21-0 in evaluated episodes without further calls to a large language model. The intriguing idea is to let a frozen agent improve its behaviour through a revisable rulebook, rather…
Sakana AI Watch
Sakana AI says its Japanese-focused Namazu model has been adopted by Evidence Finder, a medical literature search service for doctors in Japan. The pairing puts the model into a workflow that retrieves research, cites sources and checks that cited papers exist, while a reported medical-exam score offers a…
Anthropic Watch
Anthropic has added a rule against sustained, needless abusive or cruel behaviour towards Claude, Gizmodo reports. The policy is due to take effect on 12 November, but the company has not clearly defined where abuse begins.
Watch Desk
Open-weight models have given developers more freedom to run and adapt AI systems, but access to weights is not the same as access to training data, unrestricted terms or frontier-scale computing. A new overview traces how Llama, Mistral and DeepSeek helped broaden the field, while leaving important advantages…
Microsoft AI Watch
Microsoft’s September Copilot Studio update brings apps, workflows and agents into one building environment, while adding new ways to control and evaluate agents. For teams putting AI into business processes, the notable shift is from assembling the parts to managing more of the build and readiness checks in…
Hugging Face Watch
A Hugging Face community author says its ML Intern agent helped turn a series of detailed prompts into six public models, including a 0.8B image-prompt rewriter that runs on a CPU and specialist image tools. The examples show what a developer can build with the agent, and what…
Watch Desk
MIT professor Sasha Rakhlin argues that universities should rethink how they train and credit researchers as AI takes on more academic work. His proposals range from teaching students to audit AI-generated results to building shared research infrastructure that records failures as well as successes.
Microsoft AI Watch
Microsoft president Brad Smith says governments should prepare for AI-related job losses and require stronger safety measures for advanced systems. His proposals range from support for workers to independent testing and an emergency switch-off mechanism, making this a substantial intervention in the debate over who should set the…
NVIDIA Watch
Nvidia says it will put $1bn towards US scientific research over the next five years, including work involving AI in quantum computing, healthcare and energy security. The pledge gives the Genesis Mission a major new industry commitment, though the practical details of how the money will be spent…
Cohere Watch
Cohere has published an open dataset mapping nearly 694,411 public MCP tools across 178 occupations, offering a supply-side view of which kinds of work developers are making available to AI agents. Its key finding is a mismatch: only 2.6% of tools cleanly match a complete recorded occupational task.
Watch Desk
AI workloads do not all belong in the same place: Vinod Bijlani’s guide for CIOs argues that cost, data boundaries, latency and operational capacity should decide what runs in the cloud, on specialist GPU services, on-premises or at the edge. Its most useful advice is practical: compare the…
Watch Desk
AI may weaken the old signals of workplace power, such as job titles and reporting lines. A Forbes contributor offers workers a six-part framework for asking who controls the actual terms of their work.
Watch Desk
AI safety researcher Paul Christiano says the field may be approaching a period when advanced systems become difficult for people to direct or control. In an interview with BetaKit, he put a possible shift to that more dangerous territory on a six-to-18-month horizon, while stressing that the risks…
