#large-language-models

两项复现尝试表明，更小且部分开放的模型可以复现 Anthropic 用来展示 Claude Mythos 的大部分漏洞分析能力。

来自帝国理工学院、斯坦福大学和互联网档案馆的一项新预印本估计，大约35%的新网站由AI生成或由AI辅助生成。

一个被广泛重复的故事称 GPT-4 在“操纵”人类，这更多反映的是人们如何看待新技术，而不是机器本身的意图。Quanta 的一篇文章认为，流行的 AI 恐惧故事揭示了人类的焦虑、对“

来自大学学生和研究者的说法显示，生成式 AI 不仅在改变作业习惯，也在缩小学生带入研讨课和课堂讨论中的观点范围。

随着聊天机器人更深入地进入个人生活，研究人员和隐私专家警告称，用户可能在不清楚自己对信息去向几乎没有控制权的情况下，分享了敏感信息。

随着AI聊天机器人成为数百万人的知己——包括那些经历精神健康危机的人——研究人员和临床医生正在努力应对一个真正困难的问题：一个同情地与扭曲思维互动的AI是否会无意中强化它，我们如何才能知道？

OpenAI has released GPT-5.4, the latest and most powerful model in its GPT family, featuring enhanced reasoning capabilities and a new thinking mode for complex problem-solving tasks.

Alibaba has introduced its new Qwen 3.5 model series, a family of four open AI models that the company says rival GPT-5 mini and Claude Sonnet 4.5 at a fraction of the cost. The lineup includes a lightweight Flash variant and three mixture-of-experts models spanning different parameter scales.

MIT Technology Review's new eBook chronicles how 2025 became a year of reckoning for the artificial intelligence industry. From autonomous agents that could not complete basic tasks to enterprise deployments delivering zero business value, the gap between AI promises and reality has never been starker.

Anthropic's latest mid-tier model, Sonnet 4.6, debuts with record scores in software engineering and computer use benchmarks, plus a doubled context window of one million tokens. The release becomes the new default for free and pro users.