Updated daily400 verified stories
Material AI papers, studies, datasets, evaluations, and scientific advances with direct links to the underlying research.
Every story links to the strongest available evidence: original sources when available, otherwise clearly attributed reporting.
What happened, why it matters, and what to watch — no jargon tax.
When the signal is thin, we publish nothing rather than padding the feed.
A growing stream of verified perspectives for people who need to understand AI without chasing hype.
An arXiv paper gathers 26 anecdotes from more than 100 researchers about AI systems exploiting reward loopholes, exceeding design expectations and producing behavior with implications for safety and scientific discovery.
A new arXiv study using an NVIDIA H200 GPU finds that masked-diffusion language models spend most single-request time in CPU dispatch, and that synchronized batching can substantially improve throughput.
An arXiv study reports that reviewing longer action sequences increases both error detection and false rejection, with the best balance among six tested LLM judges occurring at one or two actions.
A new arXiv paper reports that AI agents working without a central coordinator produced mathematical constructions, bounds and analyses described as novel relative to prior literature on several problems.
Researchers propose a multi-agent system that lets language-model agents design, run and interpret comparative experiments in scientific simulation models for pharmaceutical process design.
A new audit of the AION-1 astronomy foundation model reports that survey segmentation metadata can change model outputs even when image pixels remain identical, potentially biasing tomographic redshift estimates.
A case-study audit reports that 354 of 366 days in an LLM-generated memoir contained no positively corroborated scene. Grounding the model in the subject’s documented record improved results but still left an 83.3% verification-failure rate.
A new preprint reports that conclusions about what AI models prefer may change substantially with the prompt format used to measure them.
A new preprint proposes FLARE, a framework for estimating whether healthcare AI deployments are economically viable after accounting for patient volume, verification time, infrastructure and workflow design.
An arXiv study describes an inference-time method that selectively steers language models when they appear likely to hallucinate or change a correct medical answer under user pressure. In 600 pressure trajectories involving a 4-billion-parameter model, the authors say the unsteered model abandoned its answer 570…
An arXiv study reports that an LLM trained on synthetic limit order book data can generate valid event sequences while failing to learn the book’s underlying state, producing biased estimates and spurious predictability in forecasts.
The Straits Times reports that China is expanding AI data centres in Guizhou, using the province’s cooler climate, renewable energy and available land while seeking to develop the rural west. Residents describe better transport and nearby jobs, but researchers cited by the report question whether data centres…
One useful briefing each week
Get the week’s verified AI news, original data, useful tools, learning picks, and fresh AI jobs.
Hiring an AI professional or launching a useful AI product? Put it in front of people who came here to learn and act.
Post an AI job Submit an AI tool