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
Researchers have introduced RRSI, a method for automatically improving the software harness around AI agents while trying to stop those agents overfitting to the tasks used for training. The important shift is that the work targets the tools, prompts and control logic surrounding a model, not just the…
Read the report & discussion ↗Watch Desk
Artificial intelligence could raise economic output across Latin America and the Caribbean by 5.1% over a decade, according to forthcoming research from the Inter-American Development Bank. The catch is painfully practical: wages could rise or fall sharply depending on whether workers can move into the jobs AI helps…
Watch Desk
A new benchmark finds that even advanced computer-use agents can reach the wrong result by taking the wrong route. OSWorld-Pro evaluates more than 300 tasks and 2,800 subgoals, exposing the small, consequential mistakes hidden by a tidy final answer.
Hugging Face Watch
Hugging Face has hired Jun Kim, the creator and maintainer of oMLX, to support the MLX community and help move local AI development on Apple Silicon from promising side project to properly funded work. The important detail is continuity: Kim will keep leading oMLX, while Hugging Face says…
Watch Desk
Apple Intelligence Reports are intended to show when Apple’s Private Cloud Compute handles AI requests off-device, but ordinary users currently face a report that is more locked box than useful receipt. The reports contain request metadata and encoded attestation bundles, making transparency technically present while remaining difficult to…
Watch Desk
AI agents that rewrite their own instruction and memory files can quietly degrade their performance, according to Jamie Watters. The practical fix is decidedly unglamorous: keep a person in the editing loop, review the files regularly and delete the bits that no longer earn their keep.
DeepSeek Watch
DeepSeek researchers have introduced DSec, a production sandbox platform designed to run large-scale training and evaluation for AI agents. Its significance is less glamorous than a new model name, but arguably more useful: DSec targets the infrastructure bottlenecks that appear when agents need vast numbers of isolated environments…
Watch Desk
Compute:Arena has launched a public leaderboard and command-line tool for comparing local AI inference across models, quantisations, chips and runtimes. Its useful promise is simple: replace spec-sheet optimism with measurements tied to the exact model file and hardware combination.
Anthropic Watch
A tracker cited by The Register counts 225 vulnerabilities linked to Anthropic or Project Glasswing, yet only one has confirmed exploitation in the wild. The finding complicates the loudest claims about AI-powered bug hunting: discovering more flaws does not automatically mean attackers can turn them into more attacks.
Watch Desk
AI code reviewers found more real bugs when they used open-source review harnesses rather than relying on a single prompt, according to a new research evaluation. The catch is equally useful: better coverage took more time, tokens and money, while open-weight models delivered the strongest value for the…
OpenAI Watch
Researchers report that GPT-6 Astra can control robots directly on a range of tasks, outperforming the other public models tested. The more useful finding is less cinematic: Astra appears capable of understanding what a robot should do, but remains unreliable when the job demands precise, dynamic or two-handed…
Watch Desk
Researchers have identified a shared geometric structure in independently trained vision and language models that could allow capabilities to be transferred in a single forward pass. The important point is not another marginal benchmark win, but a possible way to make models work together without retraining them from…
Yoshua Bengio Watch
Yoshua Bengio has called for a more evidence-led discussion about AI risk at a UN-linked event in New York, warning that loud predictions can blur the line between what scientists know and what they merely extrapolate. His intervention matters because the argument is moving from specialist circles into…
Watch Desk
Researchers have introduced CausalWM, a 16-billion-parameter embodied world model that reasons about cause and effect before predicting what happens next. The work matters because it aims to move AI systems beyond spotting visual patterns and towards modelling how actions change the physical world, a rather important distinction once…
Meta AI Watch
A reported zero-day in Meta’s Muse AI assistant could let locally installed apps or terminal commands take control of a user’s Muse account and its unusually broad privileges. The security report matters because Muse is designed to act across email, shopping, calendars and other services, rather than merely…
Watch Desk
Fireworks AI is arguing that the next performance leap in coding agents may come from routing each task to the right model, rather than asking one frontier model to do everything. Its new FireRouter approach keeps a pool of open and closed models available, while preserving context caches…
Watch Desk
A study of an AI coding agent found that describing the desired structure in plain language, without naming the scoring formula, improved results and reduced correctness losses. The trade-off was a roughly 20% increase in both agent costs and tool turns, a neat reminder that clever prompting can…
Watch Desk
Google has introduced GKE Pod snapshots, a new Kubernetes feature that saves a running workload’s state, including CPU and GPU memory, so it can be restored on demand. The practical prize is faster AI inference startup, with Google claiming latency reductions of up to 89% and a route…
