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
TypeSafe’s Jev is designed to return typed decisions, scores and probabilities instead of generated prose, and a new coding guide shows developers how to wire those results into real software. The important point is practical: Jev can inform a workflow, but the application still decides what happens next.
Read the report & discussion ↗Microsoft AI Watch
Microsoft is previewing an Integrated Security Operations Centre in Defender, a security platform designed to let people and AI agents detect, investigate and disrupt attacks from one shared foundation. The important change is architectural: Microsoft is joining security signals, context and protective controls so agents can act across…
Watch Desk
Xiaomi has released MiMo-V2.6, an open-weight AI family that combines frontier-level benchmark claims with unusually detailed training disclosures. The practical significance is not just another leaderboard result: developers can inspect the weights, training framework and more than 7,000 task environments under an MIT licence.
Meta AI Watch
Meta has announced VR glasses that weigh about 100 grams, use a pocket-sized compute puck and are due to launch in spring 2027 for $1,299.99. The important shift is not merely making virtual reality lighter: Meta is combining a high-resolution display, hand and eye controls, an AI assistant…
Microsoft AI Watch
Microsoft is urging teams to treat AI models, inference endpoints and retrieval pipelines as live dependencies in resilience planning, not invisible extras. Its new Azure guidance argues that a system is not resilient merely because the diagram shows three zones and a failover arrow. Someone has to prove…
Watch Desk
Andrew Dai, a former Google Brain and DeepMind researcher and co-founder of Elorian, argues that frontier AI remains unreliable at detailed visual tasks such as counting objects, tracking robot actions and interpreting complex designs. The practical point is less glamorous than another benchmark victory: systems that recognise a…
llama.cpp Watch
llama.cpp 0.5.0 is out with support for three new model families, a substantial ggml refresh and a long list of fixes across local inference, multimodal inputs and serving. The practical change is breadth: more models and messier real-world workloads now have a route through the same open-source stack,…
AMD Watch
AMD has filled in much of the picture around its Zen 6-based EPYC Venice server processors, including a 256-core flagship, faster memory and a direct performance challenge to NVIDIA’s Vera CPUs. The important point is not one spectacular benchmark, but a broader platform aimed at general-purpose computing, HPC…
NVIDIA Watch
NVIDIA’s Accelerated Quantum Research Center in Boston is due to receive IonQ’s Superion 256 in 2027, placing a 256-qubit quantum processor directly alongside an NVIDIA GB200 NVL72 system. The significance is practical rather than purely numerical: the companies are building a testbed for hybrid quantum-classical computing, where quantum…
AMD Watch
Viridien says it is using AMD CPUs and GPUs to extend its production-scale high-performance computing operation from energy-sector geoscience into materials discovery, molecular modelling and drug development. The important detail is practical rather than glossy: the company says its systems run more than one million HPC workloads a…
Watch Desk
WorkOS has launched Airlock, an access-control system that checks whether an AI agent’s proposed API call matches its declared task and the organisation’s security policies. The practical shift is important: agents can now be stopped, blocked or sent to a person for approval before an action reaches a…
Microsoft AI Watch
Microsoft researchers say robots can gain speed, accuracy and battery life by sending demanding AI inference to nearby edge hardware or the cloud instead of running everything on the robot. The catch is wonderfully physical: network delay and bandwidth now become part of the robot’s nervous system.
Google DeepMind Watch
Google is adding persistent, cross-device memory to its Private AI Compute platform, using encrypted cloud storage while keeping the keys on a person’s devices. The aim is to give AI assistants continuity without asking users to surrender the privacy advantages of local processing.
MIT CSAIL Watch
MIT has appointed computer scientist and Two Sigma co-founder David Siegel as its next Innovation Fellow, giving him a year to explore how artificial intelligence can accelerate scientific discovery. The appointment matters because it places a technologist with experience in research, finance and AI infrastructure inside the university…
MLCommons Watch
MLCommons has joined a €16.9 million European research project to develop generative AI that models disease mechanisms, with the organisation responsible for building the benchmarks that may decide whether the system deserves anyone’s trust. AIRIS brings together 21 partners from Europe, Canada and the US for a four-year…
