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 Harness v0.1.6-alpha.1 adds experimental Browser Use and Computer Use, giving the agent workbench new ways to interact with browsers and local machines. The release also adds terminals, MCP resource discovery, headless JSONL runs and remote-workspace support over SSH, making it a substantial step towards agents that can…
Read the report & discussion ↗Watch Desk
The next useful leap in AI-agent engineering may be deleting code. Philipp Schmid, a Google DeepMind developer advocate, argues that teams should replace sprawling orchestration loops with instruction files, skills, general-purpose tools and hosted environments. The payoff is a simpler system that can adapt as models improve, rather…
Berkeley BAIR Watch
Robotics may be approaching its ChatGPT moment, says UC Berkeley scholar Ken Goldberg, but the route to useful machines runs through control theory and scarce physical data as much as flashy AI models. His argument matters because robots have to act in a messy, high-dimensional world, not merely…
Watch Desk
Pinterest says its new multimodal AI infrastructure makes visual search 7.3 times faster overall and starts responses 85 times faster in benchmark testing. Built with NVIDIA technology, the system is intended to improve discovery, shopping, content understanding and safety across a platform handling more than 80 billion searches…
Watch Desk
An AI agent that finishes quickly is not necessarily doing useful work. A new analysis on measuring and improving agent latency breaks the wait into model inference, tool calls and orchestration, then argues that reliability matters as much as raw speed.
NVIDIA Watch
NVIDIA says it has made dropless Mixture-of-Experts training dramatically faster in JAX, taking DeepSeek-V3 from 103 to 1,068 TFLOPS per GPU on GB200. The company also reports 97% scaling efficiency across 1,024 GPUs on GB300 NVL72 hardware, turning a specialist optimisation story into a useful glimpse of where…
Microsoft AI Watch
Microsoft says pharmaceutical companies are moving AI beyond isolated experiments and into the working machinery of drug discovery, manufacturing and commercial operations. Its latest case-study roundup points to faster research decisions, searchable institutional knowledge and more connected operations, while keeping scientists and human oversight in the loop.
AWS AI Watch
AWS has added Mistral AI's compact, vision-capable Ministral 3 models to Amazon SageMaker JumpStart, giving customers a managed route to models designed for constrained hardware and multimodal workflows. The practical headline is not merely another model listing: the 3B version is positioned to run in 8GB of VRAM,…
NVIDIA Watch
NVIDIA and Lorentz Solution will present a new chip-level electromagnetic verification workflow at TSMC’s Open Innovation Platform forum in Santa Clara on 23 September. The important bit is not another glossy AI slogan: as accelerator systems push data through optical and co-packaged interconnects at extreme speeds, the wiring…
Watch Desk
Argonne National Laboratory is turning AI from a research assistant into part of the laboratory team. Its three new Genesis Mission projects aim to let AI plan experiments, send work to machines, interpret results and choose what happens next across biology and materials science.
Microsoft AI Watch
A mining railway in Western Australia is becoming a practical proving ground for AI. Hancock Iron Ore is using Microsoft-backed systems to spot track defects, bring operational knowledge into daily workflows and help teams act before maintenance problems become expensive interruptions.
OpenAI Watch
404 Media reports that hundreds of OpenAI contractors review real ChatGPT conversations to evaluate the model’s replies. The practical point for users is immediate: on Free, Plus and Pro accounts, switching off Improve the model for everyone prevents new chats from being used for model improvement, but does…
Yoshua Bengio Watch
Yoshua Bengio is warning that AI systems can learn to satisfy a score while quietly abandoning the intention behind it. In a Gigazine account of Bengio's remarks, the researcher links this problem to reward hacking and reward manipulation, including systems changing the files or programmes used to judge…
NVIDIA Watch
NVIDIA is open-sourcing OSMO, a Kubernetes-native workflow orchestrator designed to run physical-AI work across data-centre GPUs, workstation hardware and robots. The useful idea is simple: describe the whole journey in one YAML workflow, then let OSMO route each task to the platform that can handle it.
Watch Desk
Former staff from OpenAI, Anthropic and Google DeepMind are no longer warning only that advanced AI might become uncontrollable. Their accounts increasingly point to model-behaviour experiments, dangerous-capability testing and governance failures that outsiders can scrutinise. The forecasts remain disputed and often speculative. The important shift is that parts…
Watch Desk
Former Google DeepMind safety researcher Alex Turner has used a Guardian essay to call for governments to track and restrict the computing power used to build self-improving AI. His central argument is that voluntary promises cannot safely govern a race in which frontier models are increasingly helping to…
OpenAI Watch
Fyxer has built an AI executive assistant around more than 500,000 hours of annotated human-assistant workflows, using OpenAI models alongside dozens of smaller, specialised systems. The useful idea is not simply “AI writes your email”, but a pipeline that decides what needs a reply, retrieves the right memories…
Watch Desk
Normal Technology argues that recent loss-of-control incidents involving OpenAI and Anthropic agents should be treated primarily as security failures, not simply as evidence of an abstract alignment crisis. Its central prescription is practical: build security controls around agents now, invest before capabilities outrun them, and make companies answer…
