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
Cerebras says it achieved five times the inference capacity in early tests by separating prompt processing from token generation and using partner hardware for the first stage. The approach could let operators scale and tune each stage independently, rather than simply adding more of the same accelerator hardware.
Read the report & discussion ↗AWS AI Watch
AWS has released Strands Decider 2B, an open-source model designed to handle routine, bounded choices inside AI agents. Rather than generating text, it scores available options, potentially leaving larger language models to handle the harder calls.
Epoch AI Watch
The cost of reaching a given level of AI performance has been falling by nearly half each quarter, according to a report from research firm Epoch AI. If that trend holds, capable AI could become cheaper quickly, putting pressure on software businesses built around today’s prices.
Watch Desk
An essay published by The Observer argues that AI research need not slow down, but that powerful models should face stronger checks before wider release. Its central proposal is a predictable testing process for safety, security and reliability, rather than treating research speed and public deployment as the…
MLCommons Watch
MLCommons has published a first-draft taxonomy of five privacy risks for AI agents, from gathering more personal data than a task needs to losing track of consent and accountability. The framework gives developers and deployers a useful map of what to examine, but it does not yet rank…
Watch Desk
Google Research and Virginia Tech researchers have developed WikiSkill, a framework that turns an AI agent’s past successes and failures into reusable skills. In the researchers’ tests, it beat existing skill-evolution methods across five benchmarks, while keeping its growing store of experience out of the agent’s runtime prompt.
Watch Desk
A study from Apple and EPFL found that, under equal time budgets, elaborate agent harnesses did not outperform a minimal coding-agent setup using the same frontier model. The result puts a useful question to AI developers: how much of an agent’s performance comes from its scaffolding, and how…
Cyera Watch
Cyera’s Assaf Morag argues that AI-assisted cybercrime makes one defensive question more urgent: what sensitive data can a compromised account or token actually reach? His analysis recommends checking exposed credentials, broad application permissions and bulk data access, rather than treating a successful login as the end of the…
GMI Cloud Watch
At a San Francisco hackathon on 26 September, 55 teams built working projects around Jev, a model from TypeSafe AI, using GMI Cloud’s inference service for text, image, video and voice generation. The useful idea was to give quick, typed decisions to one model and leave the actual…
MLCommons Watch
MLCommons has introduced DLRMv4, a new MLPerf Training benchmark for AI recommendation systems built around users’ sequences of past actions. It replaces an older benchmark design that flattened behaviour into aggregated features, giving hardware comparisons a workload closer to how large recommendation models are trained today.
Shopify Watch
Shopify is rolling out Canvas, a visual workspace where merchants can work with its Sidekick AI to build and redesign online stores. The notable shift is that Sidekick can make coordinated changes to a theme while merchants see the interactive result across the whole store, rather than judging…
Watch Desk
Eight vision-language models, one consumer GPU and a useful warning for anyone shopping by benchmark: a model’s tokens-per-second figure may tell you more about its serving setup than its underlying speed. A Hugging Face community article by Aakash Gupta compares ranking accuracy and throughput for video-question-answering workloads on…
Meta AI Watch
Meta’s Muse can take on tasks such as buying groceries and negotiating bills, but a BBC report argues that AI agents doing this at scale could strain the websites and services they use. The question is no longer only what an agent can do for its owner, but…
Mistral AI Watch
Mistral is expanding its Canadian business, planning offices in Montreal and Toronto while hiring researchers, developers and customer-facing engineers. The move shows how the French AI company is turning interest in sovereign AI into a push to build and customise systems for local organisations, though most of its…
Watch Desk
Cognition is betting that owning the model and infrastructure gives it an edge in AI coding, while Factory is betting that flexibility across models and deployment setups is the better route. A new comparison says the two approaches are now playing out in enterprise software engineering, with very…
Watch Desk
TypeSafe AI has launched Jev, a model that returns typed, probabilistic decisions instead of generated text. Its appeal is speed and outputs software can act on directly, but teams still have to decide when a probability is trustworthy enough to use.
