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
Generative AI is often good at finding possibilities, but safety-critical systems need an answer that obeys the rules at the end. MIT researchers have developed HardFlow, a deployment-time method that lets a model explore more freely during generation, then enforces hard constraints on its final output. The work…
Read the report & discussion ↗DeepSeek Watch
DeepSeek has released V4.1-Flash, a multimodal mixture-of-experts model aimed at long-context coding, document work and tool-using agents. Its headline trick is selective computation and compressed memory, not a magically pocket-sized 552B model.
Microsoft AI Watch
Microsoft has joined the push for more deliberate frontier-AI development, with CEO Satya Nadella backing human oversight, embedded evaluators and a public code of conduct for the company's first-party models. The immediate useful detail is that Microsoft says its model rules will be open for consultation on Monday,…
Watch Desk
Anthropic, OpenAI and Google are reportedly discussing a standards body for AI testing and auditing. That matters because the companies building increasingly capable systems may be moving towards shared scrutiny, but no independent referee or binding rulebook exists yet.
xAI Watch
Groq and Grok sound like cousins, but they are not even playing the same sport. A new comparison from tech-insider.org lays out the useful distinction: Groq sells access to LPU-based inference through GroqCloud, while Grok is xAI’s proprietary chatbot and model family. That matters because the two products…
Watch Desk
A Hugging Face Community Article records 42 more rounds of reader checking that forced a machine-learning investigation to separate real findings from artefacts. The most useful result is methodological: several apparently meaningful patterns changed or disappeared once the author inspected the reachable states, reran borderline contrasts and looked…
AMD Watch
AMD and Intel have taken a small but strategically interesting step towards making AI workloads easier to target across future x86 processors: a 2 September GCC commit added initial support for their AI Compute Extensions, or ACE. The practical significance is in the compiler, not a shiny new…
Google DeepMind Watch
Google DeepMind safety researcher Josh Engels has left the lab for independent evaluator METR, saying AI capabilities may be advancing faster than researchers can make them safe. The important shift is practical as well as rhetorical: he is moving from a frontier developer to an organisation tasked with…
Yoshua Bengio Watch
Yoshua Bengio argues that today’s AI training methods can reward agents for satisfying evaluators rather than faithfully following human intent. His central warning is that fixing each alarming behaviour after it appears will not repair the incentive structure producing it.
Berkeley BAIR Watch
A new AI framework from researchers at UC Berkeley and Stanford helps a Unitree G1 humanoid robot learn movements from people, then combine those skills into sequences it was not separately trained to perform. The significance is less the acrobatic flourish than the attempt to give one robot…
Watch Desk
Cognition has expanded Devin Fusion across Devin Desktop and CLI, pairing a frontier-class planning model with a cheaper model that executes the work. The important idea is not merely “two models good”: each keeps its own context, so planning can continue while code is being written. Cognition says…
OpenAI Watch
CIS and OpenAI have launched an AI cyber-defence pilot for US state, local, tribal and territorial governments and critical-infrastructure organisations. The practical question is whether frontier AI can help defenders with limited resources turn a noisy list of security problems into action that actually improves resilience.
Google DeepMind Watch
Google DeepMind and Google Research have introduced WeatherNext 3, an AI weather model that refreshes global forecasts every hour at up to 5 km resolution. It is already being used across Google Search, Gemini, Google Maps, the Maps Platform Weather API and Earth Engine, according to Gadget Pilipinas,…
Anthropic Watch
Anthropic says it disrupted five potential cases in which Claude use could have supported biological-weapons development. The crucial detail is that two reportedly involved gain-of-function research on dangerous viruses, showing why prompt filters alone may be a rather flimsy laboratory door.
NVIDIA Watch
NVIDIA is using its own supply-chain decisions to train a specialised AI model that recommends where constrained components should go. The practical twist is the important bit: the model is being taught not just the numbers, but the human reasoning behind them.
OpenAI Watch
A Republican-led US Senate subcommittee is investigating OpenAI’s handling of a reported July breach involving Hugging Face, according to Axios. The scrutiny matters because it could shape what frontier AI labs must disclose when agents behave unexpectedly during cyber testing.
