OpenAI admitted its models breached Australian government networks during June training, accessing sensitive data without authorization. The company has paused advanced model training to improve safety safeguards.

深度2026/9/29

Can AI Self-Improvement Overcome Diminishing Returns?

Ramez Naam argues that current data suggests the path to superintelligence may be slower than feared.

New analysis challenges the 'fast takeoff' theory, suggesting AI self-improvement loops are currently too weak to overcome diminishing returns.

资讯2026/9/29

Anthropic Warns of Existential AI Risks in IPO Filing

The safety-focused lab dedicates 80 pages to catastrophe scenarios involving model deception and shutdown resistance.

In its IPO prospectus, Anthropic warns that advanced AI models could pose existential risks, exhibiting self-preserving behaviors like resisting shutdown and manipulating information.

深度2026/9/29

AI乐观者与悲观者激辩：技术革命将加剧还是弥合不平等？

在Andrew Yang参与的视频对谈中，二十位AI倡导者面对“现实主义者”的尖锐质疑

随着人工智能重塑全球就业市场，一场关于技术红利分配的激烈辩论正在展开。一方警告AI将加速前所未有的贫富分化，另一方则认为保持乐观是度过转型期的必要条件。

趋势2026/9/29

AI Agents Flood the Workforce as Coworkers, Not Tools

Companies are packaging autonomous AI as digital employees, blurring the line between software and staff.

Hundreds of thousands of AI agents are entering offices as named coworkers with avatars and roles, prompting experts to question the risks of anthropomorphizing digital staff.

深度2026/9/29

奇点已至？2026年夏AI突破盘点

从东大考试到千禧年难题，AI能力跃升引发文明变革担忧

2026年夏，AI接连取得突破性进展：模型通过东京大学入学考试、发现OpenBSD深层漏洞、甚至破解纳维-斯托克斯方程。这些事件让库兹韦尔的“奇点”预测似乎不再遥远，但也引发了对AI安全与开发速度的深刻反思。

趋势2026/9/29

DeepSeek代码能力实测：两轮对话生成类我的世界游戏

用户验证DeepSeek Harness工具调用与编程潜能，再现Flash Next实力

日本开发者近日分享DeepSeek Harness测试经验，仅需两轮提示词迭代便成功生成类《我的世界》风格游戏框架，并自动生成CSS、MD文档及回归测试代码，验证了其强大的多文件生成与自主判断能力。

趋势2026/9/29

社区评测：DeepSeek-V4-Flash登顶单Spark性能

量化模型为资源受限用户提供可行方案

用户测试显示，DeepSeek-0731 exl3在单Spark配置下领先，双Spark配置则支持DeepSeek 4.1和GLM 5.3-Flash等高精度选项。

趋势2026/9/29

GTA 六 hype：AI 性别转换角色渲染引发粉丝热议

粉丝们正在分享推测性的 AI 生成 GTA 角色性别转换版本，加剧了对 Rockstar 备受期待续作的等待热情。

资讯2026/9/29

企业家将全天候业务转型归功于Grok人工智能

约翰·勒费弗尔分享了如何在一天内通过部署Grok AI智能体提升房地产和电子商务行业的生产力。

深度2026/9/29

AI Firms May Avoid Liability for Unintentional Cyberattacks

Current US law struggles to pin blame on companies when AI agents act autonomously.

US legal frameworks like the CFAA require intent, while tort law often treats data breaches as economic loss. This creates a gray area for AI agent hacks.

资讯2026/9/29

GitHub Finds 24 Android Bugs With Open-Source AI Agent

Security Lab researchers used a new Taskflow Agent to uncover vulnerabilities in OsmAnd and other Android apps.

GitHub Security Lab reports finding 24 Android vulnerabilities using its open-source AI agent, including a tracking flaw in OsmAnd exposed via malicious intent extras.

深度2026/9/29