Updated daily177 verified stories
New frontier, open-weight, multimodal, reasoning, and specialist AI models—plus the evidence behind their capability claims.
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.
A multi-site prospective study reports that a large language model matched 87% of human-consensus answers when extracting information from unprocessed medical records, but accuracy fell to 62% on questions requiring event timing and greater clinical reasoning.
A paper accepted by ECCV 2026 proposes MV2GF, a multi-view pedestrian-detection system designed to generalize better when camera arrangements differ from those seen during training. Its authors use a visual geometric foundation model to estimate 3D structure and place image features in more appropriate world-space…
Companies are evaluating downloadable AI models as a hedge against rising usage costs, but hosting, governance and security concerns are slowing adoption.
VentureBeat reports that Anthropic updated Claude Tag so its Slack agent evaluates entire conversations, rather than individual messages, before choosing whether to respond, start work, route a task, or remain silent.
The Straits Times reports that Chinese state-affiliated hacking groups are using DeepSeek and other AI models to automate reconnaissance, exploit development and other parts of cyberattacks, citing Taiwanese researchers and cybersecurity firms.
A revised arXiv paper proposes a framework for using synthetic data in scientific research while quantifying when conclusions remain statistically valid, including for LLM-based evaluation.
An arXiv preprint reports that GPT-4o’s simulated persona patterns depend partly on how personality questions are ordered and aligned, suggesting that single aggregate scores may miss important behavior.
A new arXiv preprint argues that enterprise finance systems using large language models should be judged not only by answer accuracy but also by whether every claim can be traced to authoritative evidence. Its proposed framework improved citation traceability in a 145-question evaluation, while producing…
An arXiv preprint introduces VortexChat, a framework that uses a large language model to coordinate photonic-device generation, refinement and electromagnetic simulation. The authors report that it autonomously designed and helped fabricate a terahertz vortex-beam multiplexer whose measured performance matched…
A new preprint proposes a hierarchical memory system that decides which parts of an AI agent’s execution history should remain active. On GAIA-Text, the authors report higher accuracy, lower prompt-token use and less persistence of poisoned information than linear memory.
36Kr reports that DeepSeek released DeepSeek Harness, a developer-preview environment designed to let AI agents modify, evaluate and roll back software components. The report says the system uses a mechanism called Cordis to manage changing plugins and dependencies, but its release status, capabilities and broader…
NVIDIA says a preview feature in its Dynamo inference platform can restore an LLM worker in seconds by keeping a preinitialized standby engine on the same GPUs and sharing model weights in memory. In the company’s benchmark, recovery took 7.3 seconds instead of 283 seconds after a worker failure.
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