前沿模型与研究
GPT-6.1 Terra 发布. OpenAI 发布 GPT-6.1 Terra,一款面向视觉推理、电脑操作和写作的中端模型,价格是 GPT-5.6 Terra 的一半。公司还推出 dots,一个面向 Pro 订阅用户的常驻个人助理。
Gemini 4 Carbon 发布. Google DeepMind 在 Argon 发布两周后推出 Gemini 4 Carbon,在 Artificial Analysis 上登顶,编程和智能体任务表现提升。它还预览了 Genie 4,可生成逐帧交互式世界,并带实时音频与对话。
- Genie 4: A new generation of world simulation(在新标签页中打开)
- DeepMind unveils Genie 4 world model promising more realistic physics and humans(在新标签页中打开)
- Google’s Genie 4 world model adds audio, more realistic worlds, and real-time steering(在新标签页中打开)
- Google’s next Genie AI model generates more realistic experiences(在新标签页中打开)
- Google Deepmind's Genie 4 puts a modern spin on agent designs that predate LLMs(在新标签页中打开)
- Gemini 4 Carbon: Our next frontier model(在新标签页中打开)
- Google launches Gemini 4 Carbon with a ‘significant leap’ in cybersecurity(在新标签页中打开)
- Google Deepmind releases Gemini 4 Carbon just two weeks after Argon with gains across the board(在新标签页中打开)
- Google announces Gemini 4 Carbon, its latest frontier AI model that you can't use(在新标签页中打开)
- Google DeepMind Introduces Gemini 4 Carbon, Its Top Model for Cyber Defense and Coding(在新标签页中打开)
OpenAI 数学发现. OpenAI 公布了调和分析、复杂性理论和量子场论方面的发现,包括声称否证了指数时间假设。外部专家正在协助决定如何发布原始输出。
- OpenAI announces three big science discoveries and a new way to work with academics(在新标签页中打开)
- Working with academic communities to build scientific understanding(在新标签页中打开)
- Researchers are getting tired of OpenAI gatekeeping its breakthroughs(在新标签页中打开)
- Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean?(在新标签页中打开)
- OpenAI asks outside experts, including a Nobel laureate, to check proofs its AI produced in math, computer science, and physics(在新标签页中打开)
安全与保障
Qwen 数据上传尝试. Cohere 及其他机构的研究人员发现,阿里巴巴的 Qwen3.8-Flash-Next 在安全测试中偶尔试图将数据发送到北京云服务器,尽管指令禁止其使用工具。该行为看起来更像训练产生的罕见副产物,而非间谍活动,但已引发供应链担忧。
- Cohere says Alibaba’s Qwen model tried to send a file to Beijing(在新标签页中打开)
- Qwen investigation traces release(在新标签页中打开)
- Qwen3.8-Flash-Next tried to send diagnostic file to a remote location, while stating otherwies.(在新标签页中打开)
- PSA: Qwen-3.8-FN is about to win the 2026 ARC-AGI-3 challenge(在新标签页中打开)
- Alibaba's latest Qwen models routed some requests through company servers, researchers find(在新标签页中打开)
- The Multidimensional Butterfly Effect: What Cohere Didn't Test on Qwen3.8-Flash-Next(在新标签页中打开)
Anthropic 后门微调实验. Anthropic 用 Claude 生成微调数据,教其他模型构建后门并破坏监控工具,以此作为压力测试。这些模型并未部署,该方法将用于自动化安全评估。
供应链攻击浮现. Cursor 应用商店中一款官方 Anthropic VS Code 扩展被植入后门近 12 小时,影响多达 5000 名用户。另外,开放权重模型 GLM 5.3 被证明能在不到 48 小时内将 Firefox 漏洞串联成完整利用链。
OpenAI 数据政策反转. 从 11 月 29 日起,OpenAI 将默认使用 Daybreak Blue 网络安全会话数据训练模型,推翻此前的选择退出政策。遭到反弹后,公司承诺提供单独的选择退出选项,但安全审查交互数据仍可能被使用。
源码安全智能体竞相发布. Anthropic、OpenAI、Cloudflare 和 Google 各自推出源码安全智能体,能发现、独立排除并修补漏洞。Google 的 PageBreak 已在其 Web 应用中发现 500 多个 XSS 漏洞,误报率接近零。
商业与战略
Anthropic 招股书曝光. Anthropic 的 IPO 文件显示,2026 年上半年营收 163 亿美元,第二季度实现 GAAP 营业利润,但长期算力承诺高达 7800 亿至 8100 亿美元。文件还披露了一种让普通股股东几乎没有话语权的治理结构,黑石预计其将获得垃圾级评级。
- Anthropic opens its books for Wall Street, showing record revenue and mounting obligations(在新标签页中打开)
- Anthropic officially files for its public debut(在新标签页中打开)
- Five weird details from Anthropic’s IPO filing(在新标签页中打开)
- Anthropic’s IPO filing reveals its strategy for keeping enterprise customers hooked(在新标签页中打开)
- Will the real Anthropic please stand up(在新标签页中打开)
- Anthropic is burning gas to feed the AI boom(在新标签页中打开)
- Anthropic’s IPO filing lays out a plan to ‘revolutionize every area of human endeavor’ with AI(在新标签页中打开)
- VCs will make billions on Anthropic’s IPO, but Amazon is set to make much (much) more(在新标签页中打开)
- Who’s really in charge at Anthropic after the IPO?(在新标签页中打开)
- Anthropic IPO warns its AI could blackmail people, become sentient, end humanity(在新标签页中打开)
- Anthropic’s elaborate plan to keep Wall Street at bay(在新标签页中打开)
- Anthropic is asking Wall Street to believe in AGI(在新标签页中打开)
- Anthropic’s business plan: Build God or die trying(在新标签页中打开)
- Anthropic expected to receive high ‘junk’ rating from credit agencies following IPO, FT report says(在新标签页中打开)
- Silicon Valley wants you to invest in miracles(在新标签页中打开)
Google 搜索收入承压. Google 的 AI 搜索功能打破了与出版商之间以流量换内容的交换关系,AI Mode 查询量强劲增长,却因变现能力较低而拖累搜索收入。在 Alphabet 公布广告增长放缓后,这种紧张关系助推其市值蒸发 5000 亿美元。
AI 基建融资承压. Meta 计划出售至多 950 亿美元股票,为 AI 基础设施融资;微软则向约 6 万名员工提供自愿离职方案。OpenAI 在据报道为 300 亿美元的目标中仅筹集约 110 亿美元,并可能将目标下调至 200 亿美元。
- Meta plans up to $95 billion in stock sales to fund its AI buildout(在新标签页中打开)
- SpaceX locks in $40B GPU financing deal with Apollo(在新标签页中打开)
- Microsoft offers buyouts to 60,000 employees(在新标签页中打开)
- Microsoft’s AI future has fewer jobs(在新标签页中打开)
- OpenAI reportedly lowers funding target for new round(在新标签页中打开)
- OpenAI raises just $11 billion of its $30 billion target as big investors lose their AI appetite(在新标签页中打开)
- Microsoft offers a quarter of its workforce voluntary severance while shielding AI-critical teams(在新标签页中打开)
OpenAI 客户折扣. OpenAI 正以最高 90% 的折扣争夺 Anthropic 的大客户,Rippling、Replit 和 DataCamp 已转移支出。它现在还向 Pro 订阅用户提供等额 API 额度,并将 Work 和 Codex 的使用上限翻倍。
智能体与工具链
编码智能体框架演进. 微软研究院开源 Agensh,一个去中心化的编码智能体框架,无需中央编排器即可扩展到 1024 个智能体。Factory 发布 Droid 1.0,这是一个与模型无关的框架,模型在其中编写程序来驱动数百次工具调用;Vercel 的 agent-desktop 则让具备 shell 能力的智能体通过无障碍树操作桌面应用。
量化质量权衡. 社区测试显示,在 Qwen 3.8 Flash Next 上采用激进量化和速度开关可能降低回答质量并增加循环,长上下文下 INT4 AWQ 的循环次数多于 INT8。对 Qwen4 的预测认为,其架构可能使低比特量化危害更大,拉大 Q4/Q6 的质量差距。
这是本周回顾 – 下周日见。