Latest in AI
Every brief, newest first. Each one links to its primary source.
-
OpenAI sets out how it will disclose model misalignment, and publishes six cases
OpenAI has published a standing process for investigating and disclosing misaligned model behaviour, together with six cases observed over the past six months.
-
Perplexity brings its local-first Portable Computer to Windows
Perplexity’s on-device agent now runs on Windows, but it needs an RTX GPU with at least 24GB of, a floor well above a typical…
-
Manchester researchers forecast UK air pollution with Nvidia Earth-2 models
A University of Manchester team has produced UK-wide air quality forecasts at two-to-three-square-kilometre resolution, training the model in about two days on a single…
-
An AI agent scoring 77% on average solves only 53% of tasks every time
IBM Research puts a figure on agent flakiness: an agent scoring 77.4% averaged across five runs completed only 53.0% of tasks on all five,…
-
MLPerf Inference v6.1 adds RAG and agentic tests as Rubin arrives in preview
MLPerf’s latest round adds retrieval-augmented generation and edge agentic benchmarks and sets a participation record, but Nvidia’s Rubin silicon appears in the preview category…
-
A preprint claims 200,000-token context on a 24 GiB laptop
A single-author preprint reports a 27B model running at 212,992 tokens of context on a 24 GiB MacBook. Specific numbers, but, none of them…
-
Mistral’s models go into Firefox’s Smart Window
Mistral Small 4 lands in Firefox’s Smart Window beta, with the two companies’ own posts not quite matching on regions or on data retention.
-
OpenAI puts sponsored agents inside ChatGPT
Click an ad in ChatGPT and you can now talk to a business-sponsored agent, with HubSpot and Shopify distribution, and no stated answer yet…
-
OpenAI opens its managed agent harness in public beta
OpenAI’s managed agent harness reaches public beta with no platform fee, alongside GPT-Live-1 general availability at $0.05 a minute of voice session.
-
IBM and NASA release an open lunar foundation model
A model trained on nine instruments across four lunar missions, open on Hugging Face, with no parameter count or licence named in the announcement.