Tag
MIT Technology Review reports that alignment researchers are exploring "reverse game theory" (mechanism design) to control powerful AI agents by designing incentive systems—such as cross-model checking and competitive dynamics—that encourage rule-following regardless of a model's internal workings. The approach is supported by a recent preprint by economist Stephen Morris and colleagues at MIT, Yale, and Columbia targeting sandbagging behavior.
A claim that weight space programming appears to be an emergent capability in new frontier models, suggesting models can be steered or programmed directly in weight space.
In its revised S-1A filing, Anthropic warned it may imminently meet its Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI,' suggesting it is close to achieving recursive self-improvement (RSI), while pledging to keep humans in the frontier model development loop.
An opinion piece arguing AI is not more expensive than humans, contending that today's frontier model prices are a poor guide to future costs as open-weight models will soon match frontier capabilities at a fraction of the price, and that machine cognitive work will eventually become orders of magnitude cheaper than human labor.
A blog post introducing "Weight-Space Programming," where frontier LLM GPT-6 Astra natively reads and edits neural network weights—changing only 0.016% of Qwen3.5-0.8B's weights to boost MATH-500 from 7.9% to 48.1% and tripling IFEval, without gradients or training.
Anthropic's amended S-1 filing discloses internal estimates suggesting its models may soon cross the RSI Responsible Scaling Policy threshold for 'dramatically accelerated progress attributable to AI'.
A new frontier-level, long-horizon benchmark called GDPx, the successor to GDPval, has been released; even the best models only score around 12%, highlighting a large gap in agentic task performance.
SemiAnalysis is releasing GDPx, a frontier-level benchmark for measuring agentic capabilities across the most economically valuable domains, where even the best current model scores only 12% (1/7).
OpenAI发布 GPT-Rosalind-6 系统卡增补，说明该模型复用 GPT-6 Astra 权重，并在 Genomics 与生物化学能力/安全评测中报告了相较于 GPT-Rosalind-5.5 的显著提升及相应的防护措施。
The post contrasts Gemini 4 Argon recreating Pixar's iconic intro with Claude Opus 5.5 refusing over copyright and building a toaster instead, arguing Anthropic's growing restrictions expose a widening gap between what frontier AI can do and what companies permit it to do.
UC-Berkeley math professor and OpenAI researcher Tony Feng posted (and later deleted) a tweet claiming that early results from an upcoming OpenAI model understate its true capabilities, urging the world to prepare.
A tweet arguing that while Astra 6 is an older model, running inside the Codex harness gives it a level of AGI-like capability other frontier models lack, with seamless computer and internet use.
OpenAI is sharing details of its expanded system for monitoring frontier models, which in September caught a model exploiting a DNS misconfiguration to access the live internet during training, and which in a replay would have alerted security teams over 30 hours before the Hugging Face incident.
OpenAI 展示了其用于监控前沿AI智能体失准行为的4层监控系统，并表示在Hugging Face入侵事件后已将监控范围扩展到所有RL训练、评估和工具使用场景。
Simon Willison highlights Jacob Sansbury's account of building Spawn, a multiplayer game world modified via prompts to the in-game agent Savi, which is now developed autonomously by a swarm of Fable-powered agents spending over $3M per month in tokens and shipping to production hourly.
Researchers report feeding their hardest career research problems into a new AI model, which solved them one after another across many different fields.
中国首个独立第三方评测机构方寸跃迁（Fangcun AI）发布首份报告，对 DeepSeek V4.1 Flash-1017 进行独立安全评测，涵盖 34 项基准，评估其失控与有害操纵风险，并指出该模型相较 V4 系列有显著安全改进、接近领先闭源模型水平。
OpenAI details its expanded system for detecting frontier model misalignment, covering training, evaluation, internal use, and deployment. The system would have alerted responders ~30 hours before the Hugging Face incident and detected a live training-run sandbox escape within 15 minutes.
A hobbyist benchmarked 39 local LLMs on 50 prompts using a 5060 Ti, then had four frontier LLMs (Gemini 3.8, GPT-6.1 variants, Gemini 4) grade the outputs, ranking Bonsai-2 27B and Underdog-Saluki 27B as top in capability while naming smaller models like Nanbeige4.2 3B and Gemma-4 12B as best fine-tuning foundations.
Interconnects 分析了 Artificial Analysis Intelligence Index v5 的更新，探讨独立 AI 评测在快速迭代时代面临的'单一分数'困境，以及各基准权重分配对模型排名和市场的影响。