Tag
A Wired reporter deployed a swarm of AI agents to manage a fantasy Premier League squad, finding they still underperform humans, while new research shows frontier models like GPT-5 increasingly excel at real-world prediction tasks and forecasting benchmarks such as Prophet Arena.
作者讲述了自己构建基于Lean的自动定理证明系统的经历，并亲身体验了"苦涩的教训"——模型能力的提升（如GPT-5.6 Sol）远比大量工程优化更能推动进展。
Microsoft AI CEO Mustafa Suleyman claims its MAI-DxO multi-agent diagnostic system diagnosed complex NEJM cases correctly 83% of the time versus 36% for generalist physicians, per a study published in Science, marking a step toward 'medical superintelligence.' The study shows much of the gain comes from underlying models like GPT-5 and GPT-5.5, with the orchestration layer adding smaller boosts mainly on complex cases.
Two critical PHP vulnerabilities—discovered with the help of AI models like GPT-5.6 Sol—caused widespread site outages across the internet, affecting 4chan, Wikipedia, DeviantArt, government agencies, and Pokémon services.
OpenAI led a coalition to disclose critical PHP vulnerabilities first discovered by GPT-5.6 Sol, sharing a PHP version checker and a proposed global response framework while committing $1B in subsidized Daybreak access for frontline defenders.
一篇 every.to 的长文《How Codex Happened》，讲述 OpenAI 的编程代理 Codex 如何从 GPT-5 时期表现平平到逆袭成为公司核心产品的幕后故事，并对比了 Anthropic Claude Code 带来的全任务委派式工作流冲击。
Every 的长篇深度报道，讲述 OpenAI 的 Codex 如何从一个未找到产品市场契合的云端编码工具，转型为终端/编辑器内的 AI 软件工程师，并在 GPT-5 之后成为 OpenAI 复兴的核心产品。
XBOW 使用 GPT-5.6 Sol 耗时 6 天、约 4 万美元算力发现 Chrome 全链漏洞，获得 Google 25 万美元赏金，成为 2026 年 Google 首个全链奖励，影响约 10 亿用户。
Simon Willison 发布 ttok 1.0，将默认分词器从 GPT-4 切换为 GPT-5/GPT-6 系列，依据是实验确认 GPT-6 与 GPT-5 家族使用相同 tokenizer。
OpenAI case study on how Chatham Financial uses Codex and GPT-5.6 to power internal tools, its Chatham Onyx platform, and its Process Zero consulting service — cutting trade validation from 30 minutes to under 4 minutes. The work spans employee-built apps via Chatham Vibes and client-facing capital markets workflows.
Sam Altman notes that GPT 5.1 Sol was OpenAI's fastest-growing model ever and experienced some slowness under load, with performance improvements now rolled out.
论文提出 MetaPersona 框架与 MetaPersona-DB 数据集（11,000+ 实证社会科学研究），通过检索任务相关证据、构建文献衍生的 persona 依赖图并从经验先验采样合成人群，解决 LLM 社会模拟中的 persona 冷启动问题，在三个下游任务上验证其效果并提供 MetaPersona-Studio 交互原型。
The article reports initial results from OpenRSI's Marin-Scaling-Ladder experiment, where AI agents autonomously propose, implement, and evaluate new optimizers across model scales from 550M to 2.5B, discovering PSPR with Codex (GPT-5.6).
This study evaluates using GPT-5 to score teacher-child interactions in early childhood classrooms against human raters, finding partial alignment but limitations for full assessment.
At Dreamforce 2026, Sam Altman compared future AI models, stating GPT 5.5 matches an average math professor, GPT 5.6 is top percentile, and an internal model surpasses the best human mathematicians.
Benchmarked GPT-5.6 Luna vs GPT-6 Astra across 50 real PRs, showing Astra found 92 bugs vs Luna's 69, with Luna catching 75% of bugs at 3.6% of the cost. Includes detailed evaluation breakdown and upcoming comparison with Fable 5.1.
The article describes how MIT graduate student Beatriz Yankelevich used GPT-5.6 Sol with Codex to automate quantum computing experiments, saving time and enabling more focus on analysis and design.
Anthropic shared a prompt that significantly improved writing with Fable 5.1 and GPT-5.6 Sol, which the user now uses as a rule for AI editing and writing.
A benchmark called Simple Bench shows that the QWEN 3.8 27b model has common sense capabilities almost comparable to GPT 5.0 Pro.
A user reports that frontier AI models like Codex and GPT 5.6 Sol are often counterproductive in local AI setups, adding unnecessary restrictions and ignoring specifications, and seeks community experiences on the issue.