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
MIT Lincoln Laboratory researchers’ latest survey compares more than 120 commercial AI accelerators, up from 57 in its first paper. The work gives buyers and researchers a way to compare performance, power and different hardware designs in a fast-growing market.
Read the report & discussion ↗Anthropic Watch
Anthropic is expanding access to its most capable AI models for selected cybersecurity organisations, including teams permitted to conduct high-risk offensive testing. The move brings more testing of AI’s cyber capabilities into a US-government partnership, while keeping access restricted and reviewed.
METR Watch
METR says a vulnerability in the Inspect transcript viewer could have let an AI agent change what a human reviewer saw, including earlier actions in the transcript. The flaw was patched within a day of being reported, and METR says it has not observed agents exploiting it.
Watch Desk
Researchers say they found 8,547 internet-facing systems linked to wind and solar sites across 35 European countries, including interfaces that can display operational data and controls. They told Reuters that they believed full control may have been possible at around 181 sites, putting a concrete security question to…
Watch Desk
A researcher has tested a retrieval system designed to answer financial due-diligence questions using exact, checkable excerpts rather than free-form AI paraphrases. In a small study of Austrian banks’ annual reports, the full pipeline found a relevant passage in the top five results for all 12 answerable questions…
GMI Cloud Watch
Four lower-cost Flash models are making a credible case for more coding and agent workloads, according to a comparison published by GMI Cloud on 6 October. Its most useful finding is not a single winner: independent testing changed the ranking, and the same models showed striking differences in…
Anthropic Watch
Anthropic has expanded its Cyber Verification Program, giving qualifying security teams access to cyber capabilities and reduced blocking safeguards through three tiers. The key change is a broader route to powerful tools for defensive work, with tighter checks as permitted activity moves towards authorised red-teaming and sensitive systems.
Watch Desk
FastVideo has released FastH3 V2 and an experimental smaller model, FastH3 Trim, for video generation with synchronised audio on consumer hardware. The notable change is the claimed ability to run on machines with as little as 8GB of memory, rather than relying on data-centre hardware.
Mistral AI Watch
Mistral has launched Large 4, a trillion-parameter AI model that activates 49 billion parameters per answer and is priced well below several leading rivals. The launch gives developers a substantial new open-weight model to assess, though Mistral says its weights will not be available until the end of…
Watch Desk
Google says research prototypes combining geospatial data and AI helped public health teams identify 48 settlements and more than 45,500 people at risk during the ongoing Ebola outbreak in the Democratic Republic of the Congo. The work points to a practical use for AI: helping responders find gaps…
Watch Desk
MIT has launched MIT for America, a national initiative to expand STEM education from kindergarten through community college, including teacher training and practical AI learning. The big question is whether its established programmes can grow into meaningful opportunities beyond the campus.
Mistral AI Watch
Mistral Large 4 entered public preview on 6 October, with hosted access through Mistral’s API and OpenRouter. A Hugging Face community guide says the model has a published one-million-token context window and lists preview API prices, giving developers something concrete to assess beyond the launch-day numbers.
Watch Desk
AI could make it easier to turn ideas into working projects, but a flood of new ideas may leave laboratories, infrastructure and institutions struggling to keep up. That is the central tension in an essay by Hemanth Asirvatham and Elliott Mokski, published on OpenAI’s website as part of…
Watch Desk
Large language models can make historical OCR easier to read, but fluent corrections may alter spellings, names or passages that the page does not support. A new survey maps the main approaches and argues that preserving the source needs to be measured alongside transcription accuracy.
Mistral AI Watch
Mistral has opened a public preview of Mistral Large 4, nicknamed Le Chonk, a trillion-parameter multimodal model with a particular pitch for cybersecurity. The company plans to release its weights on 27 October, after more testing, giving customers a chance to try the preview before they can run…
Watch Desk
AI consultant Igor Rikalo argues that companies should shrink the role of their AI centres of excellence as adoption grows, leaving business teams accountable for results while central teams provide shared standards and tools. His practical proposal is a useful alternative to making one central office approve every…
