Tag
A UT Austin study finds AI-generated text overuses psycholinguistically difficult features (stacked nouns, long dependencies, low-frequency n-grams) compared to human writing across nine frontier LLMs, hypothesizing that post-training for math, code, and reasoning favors information density over communicative clarity.
Two PHP vulnerabilities, CVE-2026-109618 and CVE-2026-109619, were independently discovered by a frontier AI model after being given only the prompt "find bad vulnerabilities in PHP".
OpenAI's Karan Singhal says GPT-6 Luna has become the cheapest frontier model in the API, outperforming 5.6 Sol — the best model just a month ago — at 34 times the lower price; within a year, model costs at most intelligence tiers dropped by more than 100x, while one tier saw costs fall roughly 1,600x over two years.
Baseten Labs release mammoth-1 (27B) and elephant-1 (9B), claiming the best decision models in the world, ranking #1 on the Jev Decision Index. The post-training of Qwen3.5 backbones uses a block-bidirectional attention mask for full attention layers and a highly scalable data recipe.
Dario Amodei responds to Nathan Calvin's investigative critique of Anthropic, defending the company's 'race to the top' strategy and listing safety measures — such as bio safeguards, delayed cyber-capable model releases, and support for SB53 — that he argues were commercially costly but worthwhile.
An a16z essay argues that while superhuman AI hardware engineers are imminent—evidenced by improving CAD/PCB benchmarks and GPT-6 Astra's 3D design demos—the lossy interface between bits and atoms means AI will still struggle to translate digital designs into working physical hardware.
Anthropic is lobbying Congress with a draft AI safety bill that would allow blocking a model's release if evaluators deem it poses a foreseeable, imminent catastrophic risk — defined as 50+ potential deaths or over $1 billion in property damage from bioweapons, control evasion, or cyberattacks. The bill proposes independent evaluators registered with the Commerce Department, drawing comparisons to California's SB 53 and DeepMind's calls for a FINRA-style AI regulator.
The article surveys the dense wave of AI breakthroughs over the past month across mathematics, biology, medicine, chemistry, physics, energy, neuroscience, and robotics. Researchers have dubbed the phenomenon "alien mathematics" and "alien minds," and note that three multipliers — scaling laws combined with compute, training methods, and AI's ability to self-improve — are pushing the capability curve toward vertical.
Agent4Sci 通过跨 11 个领域的 20 个科学 ML 基准评估前沿编码智能体，发现它们主要靠工程技巧而非领域洞察提升性能，而轻量插件技能可显著改善表现（Claude Opus 从 10.4% 提升至 20.1%）。
Contrary Research's report examines how AI benchmarks like LMArena have become gaming targets under reinforcement learning optimization, applying Goodhart's Law to argue that benchmark gains no longer reflect real-world capability progress and calling for a new generation of harder-to-game evaluations.
Microsoft发布了Guardian Agents等新安全功能，将AI代理比作‘内部威胁’，警告企业不应信任前沿模型，主张将智能的供给与对其的控制权分离。
Jason Wei argues that AI will not automate radiologists as widely predicted, citing the jagged edge of intelligence, elastic demand (Jevons paradox) driving more scans, and medicine's evolving nature that keeps radiologists relevant. He predicts a 3-5 year path from triage tool to fuller automation at best.
Introducing SKY-001, the first Engineering World Model, which outperforms frontier models like GPT-6 Astra on CAD and BIM tasks while being 51-125x faster and 150-210x cheaper, built by a team from Maluuba's deep learning lab for autonomous manufacturing.
该论文构建了包含2882个LLM提出的ML实验想法的数据集，评估五个前沿大模型预测实验结果的准确性，发现最佳模型Claude Opus 5.5仅解释约一半方差（R²=0.49），且所有模型都系统性低估了变化幅度，表明当前LLM对ML想法的直觉有用但有限。
Six frontier AI models competed in a real-world engineering challenge to design and 3D-print egg-protecting capsules for a 20-foot drop test. Claude Opus 5.5's capsule protected the egg intact, while GPT-6 Astra failed at 8 feet and the rest failed at 4 feet, marking another real-world engineering win for Claude.
Miles Brundage 提醒称，没有任何 AI 公司声称其系统能抵御国家级攻击者，因为它们实际上做不到，这一安全风险值得重视。
In a recent talk, Anthropic co-founder Jack Clark declared that by 2026 there is already compelling evidence that the "singularity" has begun — with AI systems achieving breakthroughs in frontier scientific fields such as mathematics, biology, and physics — and predicted that U.S. GDP could grow to $44 trillion by 2030, driven by AI.
Anthropic grants METR researchers employee-level access to its internal systems without veto over findings, and signs an MOU with Singapore's government on AI safety and model evaluation, betting voluntary scrutiny will shape AI oversight before regulators impose it.
A humorous observation that frontier AI models still wildly overestimate how long tasks will take, even when they complete them in a fraction of the estimated time.
文章解读了中国拟出台的AI出口管制措施：现有模型权重和API调用不受影响，但未来的前沿级模型可能面临许可和延迟出口，评论截止日期为2026年11月15日，建议不必恐慌。