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 V4.1 Flash is dramatically cheaper than Qwen3.8 Max on published token prices, and OrcaRouter’s telemetry puts it far ahead on time to first token. The catch is that one model is built for quick, interactive traffic while the other may suit slower, heavier jobs, so the cheapest…
Read the report & discussion ↗Watch Desk
The Netherlands’ new strategic-innovation agency NADI is launching a €40m challenge with Germany’s SPRIND to accelerate chip design using artificial intelligence. The important point is not another lab demo, but public money being aimed at shortening one of computing’s slowest and most expensive development cycles.
Watch Desk
OpenObserve has released version 1.0 for its cloud and self-hosted platforms, adding AI observability to the same system that handles logs, metrics, traces and real-user monitoring. The practical pitch is simple: teams can follow an AI agent from a user’s screen through model and tool calls to the…
Watch Desk
Okta has launched the Blueprint Alliance, a coalition of technology companies proposing a shared security architecture for autonomous AI agents. The important point is practical: the group wants agents treated less like clever software features and more like powerful identities that need discovery, permissions and ongoing supervision.
Zhipu AI / Z.ai GLM Watch
Zhipu AI says it will introduce a no-data-retention option for its model-as-a-service platform, meaning company and developer prompts, along with model outputs, would not be stored after processing. For businesses weighing whether to send contracts, customer information or source code to an AI system, that is a meaningful…
Watch Desk
Shopify says it has built a continual-learning system that turns production failures into daily model updates, using PyTorch and vLLM to improve performance while sharply reducing serving costs. The company’s case study claims the approach cut those costs by 96% and produced quality it says exceeded frontier models.…
Watch Desk
NOFire AI has open-sourced Brig, a microVM sandbox designed to run AI coding agents on macOS and Linux. The practical point is isolation: agents run inside a Linux kernel while users choose which projects and credentials they share, although internet access is enabled by default.
Watch Desk
The UK AI Security Institute is working with the EvalEval Coalition to publish AI evaluation results in a common, inspectable format. The important shift is practical: benchmark scores can now come with the setup details needed to reproduce and interpret them, rather than arriving as mysterious numbers in…
AWS AI Watch
AWS has launched CloudWatch Omni, a new observability service for applications and AI agents that brings tracing, evaluation and investigation into one shared workspace. The important bit is not another dashboard. Omni is designed to show developers and operators what an agent actually did, where it went wrong…
Watch Desk
Unreal Labs has launched Unreal Agent, a coding-agent harness that manages tool calls asynchronously. The practical promise is straightforward: developers can steer an agent while tools are still running, rather than watching the workflow pause after every command like a computer with stage fright.
Watch Desk
Skild AI says its S1 robot foundation model can learn new multistep tasks from one video demonstration, while a related Skild system is now being used on robots assembling Nvidia Blackwell GPU systems in Houston. The reports offer a rare glimpse of physical AI moving beyond a lab…
Fei-Fei Li Watch
Fei-Fei Li says World Labs’ Atlas is designed to understand and generate three-dimensional environments, pointing AI beyond language models and towards robotics, design and science. The practical significance is clear: machines that can reason about space could be more useful in the physical world, not merely more eloquent…
AWS AI Watch
AWS has made OpenAI’s GPT-6 Sol and GPT-6 Luna generally available through Amazon Bedrock, giving organisations cheaper model options for coding, document work and high-volume automation. The practical shift is not just another pair of model names: teams can now match heavier reasoning to difficult jobs and lighter,…
Watch Desk
Google has begun rolling out a Model Context Protocol integration that lets third-party AI agents work with Google Home devices and event history. The practical shift is bigger than another chatbot gaining a thermostat button: compatible agents can summarise camera activity, inspect household patterns, trigger automations and build…
AWS AI Watch
Trane Technologies says it has cut a building-maintenance diagnostic workflow from about 20 minutes to 20 seconds using an AI agent built on Amazon Bedrock AgentCore. The practical point is not another chatbot with a cheerful name. Technicians can ask questions in natural language and draw on live…
AWS AI Watch
AWS has published an architecture for helping Salesforce Agentforce search body-camera footage, surveillance video, photographs, audio and scanned documents through AI. The practical point is not another chatbot perched on top of a filing cabinet: the design turns messy evidence into structured results that an agent can retrieve…
