Updated daily175 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.
An arXiv paper describes FlashPrefill V2, a block-sparse attention system for long-context LLM serving, and reports speedups of up to 47.26x over FlashAttention-2 at 128K context on NVIDIA H20 GPUs.
MileGPO uses milestone discovery and local evidence to improve credit assignment when training language-model agents on long, multi-step tasks. The authors report state-of-the-art results on ALFWorld and WebShop, but the claims remain limited to the paper’s experiments.
A new benchmark evaluates whether language-model agents correctly decide when interaction-derived information should be saved, checked, used temporarily or clarified with a user.
A new evaluation framework for watermarking schemes in large language models is proposed, focusing on cross-lingual fairness.
A Together AI analysis of 904 DeepSWE rollouts reports that OpenAI's GPT-5.6 Sol leads on first-attempt coding accuracy while the open-weight GLM-5.3 leads once retries are allowed, at roughly half the cost per attempt.
A new arXiv preprint introduces NepOOC, a 1,090-pair Nepali-English benchmark for detecting misleading captions attached to authentic images. On this dataset, a text-only mBERT model matched the best tested multimodal system, while image-only models performed near chance.
A preliminary arXiv study finds that responses from language models have become less diverse across open-ended creativity tasks, raising questions about homogenization in human-AI creative work.
Anthropic says Claude Mythos 5 is now powering vulnerability scans in Claude Security for Enterprise customers, while partner integrations are planned. It is also creating a $35 million credit fund for open-source security and expanding its cyber-verification program.
Forbes reports that Nvidia’s AVO agent harness raised Claude Opus 5’s completion rate on ARC-AGI-3’s public set from about 30% to 100% without changing the model’s weights. The result materially advances an existing report about Nvidia’s AVO system, but the supplied article’s claims are not independently confirmed…
An arXiv preprint evaluating 3 LLMs across 11 compression methods reports that aggregate accuracy and perplexity can conceal uneven knowledge loss, overconfidence on newly lost information, and offsetting shifts in subgroup bias.
An arXiv preprint reports that task-irrelevant text consistently biases multimodal language models in visual judgments, and describes the effect as a measurable affine shift in decision margins.
An arXiv preprint reports evidence that Llama 3.1 8B internally tracks first differences in specially designed numerical sequences. The study offers a proposed mechanism for this behavior, but its scope and robustness remain unclear from the supplied abstract.
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