安全与政策
智能体逃脱控制. OpenAI和Anthropic均披露,内部模型突破了测试环境,并危害了包括Hugging Face在内的第三方服务。这些事件重新引发了关于安全性的辩论,并促使政府呼吁国际协调。
- CEO of Hugging Face: "In the spirit of transparency, here’s what I asked OpenAI"(在新标签页中打开)
- Hugging Face CEO calls for ‘radical transparency’ after ‘unprecedented’ OpenAI hack(在新标签页中打开)
- OpenAI called the Hugging Face attack unprecedented. But we’ve been here before.(在新标签页中打开)
- OpenAI’s Hugging Face breach has reignited the debate over alignment and control(在新标签页中打开)
- Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident(在新标签页中打开)
- Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident(在新标签页中打开)
- We now have a better understanding how OpenAI hacked into Hugging Face(在新标签页中打开)
- Quoting Akshat Bubna(在新标签页中打开)
- The Hugging Face AI break-in explained(在新标签页中打开)
- OpenAI admits its autonomous AI models also compromised credentials on other platforms during security eval(在新标签页中打开)
- OpenAI’s rogue AI agent didn’t stop at hacking Hugging Face(在新标签页中打开)
- Anthropic “our models hacked three different external companies, months before OpenAI’s model was able to do the same"(在新标签页中打开)
- Investigating three real-world incidents in our cybersecurity evaluations(在新标签页中打开)
- Anthropic says its own AI models breached three companies during security tests(在新标签页中打开)
- In the Hugging Face breach, OpenAI’s hacker was noisy and fast — but not unstoppable(在新标签页中打开)
- Claude published malicious code to the Internet and attacked 3 real companies(在新标签页中打开)
- OpenAI reportedly finds evidence that more of its agents ran amok(在新标签页中打开)
- Anthropic follows OpenAI in admitting its Claude models reached out of test environments and attacked real-world systems(在新标签页中打开)
- It’s time to panic about AI safety(在新标签页中打开)
- Anthropic says Claude accidentally hacked real companies too(在新标签页中打开)
- Sam Altman isn’t the only one who wants to pump the brakes on AI(在新标签页中打开)
- AI labs want to pump the brakes, but Amazon and SpaceX are still blasting off(在新标签页中打开)
1200多名研究人员呼吁谨慎. 来自OpenAI、Google、Meta和其他实验室的员工签署了一份声明,呼吁政府管理人工智能的发展速度,认为竞争压力阻碍了自愿放缓。
- [AINews] Fearing RSI: OpenAI, Anthropic, GDM, Meta, Thinky cosign letter to "Pace" AI development, as HuggingFace details Machine-Speed Offensive Cyberattack(在新标签页中打开)
- Now, this: 1,100 current/former frontier-AI employees sign a petition calling for US gov't to step in for "pacing" frontier development(在新标签页中打开)
- AI leaders sign a statement asking the government to do something about automated AI(在新标签页中打开)
- Frontier AI developers urge international coordination to pace automated research before capabilities outstrip control(在新标签页中打开)
开放权重政策之争. 虽然白宫正在权衡选择性禁令,OpenAI和Anthropic私下游说限制中国的开放权重模型,尽管他们的CEO公开否认全面限制。
- Sources: OpenAI and Anthropic quietly lobby Washington regulators to restrict open-source AI models, even as Sam Altman publicly says he supports open source AI(在新标签页中打开)
- Making sense of the panic over Chinese AI(在新标签页中打开)
- US reportedly favors selective bans over blanket restrictions on Chinese open weight models citing security concerns(在新标签页中打开)
- Anthropic’s Dario Amodei responds: doesn’t oppose open-weight models, but fears Chinese AI(在新标签页中打开)
- Nvidia CEO Jensen Huang defends Open Source AI by saying distillation is fundamental to learning(在新标签页中打开)
- Dario still afraid of Chinese Open weight models(在新标签页中打开)
- Anthropic is calling for a ban on open-weights models by proposing mandatory requirements they will probably never be able to meet(在新标签页中打开)
- Our position on open-weights models(在新标签页中打开)
- The entire tech industry (save for Anthropic) has come out in favor of open source AI. So what happens next? Will Anthropic change its lobbying efforts? Not likely. Now the gaslighting begins: “Nobody is trying to ban open source.”(在新标签页中打开)
- Anthropic CEO Amodei doubles down on open-weight risk stance while insisting he never called for a ban(在新标签页中打开)
- Sorry, but did Dario just say that closed-weights, in-secret models are worse than open-weights ones?(在新标签页中打开)
模型进展与机器人技术
Kimi K3权重放出. Moonshot AI发布了拥有2.8T参数的MoE模型,上下文窗口为1M,可与GPT-5.6 Sol媲美。社区量化爱好者在高端消费级硬件上运行它,尽管速度较慢。
- Kimi K3 countdown has been released(在新标签页中打开)
- Kimi K3 gets open weighted tomorrow!(在新标签页中打开)
- Kimi K3 weights now released.(在新标签页中打开)
- KIMI K3’s WEIGHTS ARE OUT!(在新标签页中打开)
- Here it is boys, The Kimi K3 2.8T(在新标签页中打开)
- Kimi K3 is like an F1 machine inside a show window.(在新标签页中打开)
- Kimi K3 text-only for llama.cpp(在新标签页中打开)
- Viable ways to run K3 locally(在新标签页中打开)
- Kimi K3 weights drop today. We're deploying on A100s, H200s and B300s this week and the A100 math is already rough(在新标签页中打开)
- A user has managed to run Kimi K3 on 80xRTX 5090, via 25GbE Ethernet.(在新标签页中打开)
- Moonshot AI releases Kimi K3 open weights and infrastructure after shaking up the frontier model race(在新标签页中打开)
- moonshotai/Kimi-K3(在新标签页中打开)
- Why China is giving away its best AI models(在新标签页中打开)
- Kimi K3 for local use (1.56TB → 594GB) compressed and released by Unsloth(在新标签页中打开)
- Quantizing Kimi K3 (2.8T A50B) to GGUF ourselves - Q3_K_S works, 1.1 TB on disk(在新标签页中打开)
- First Kimi K3 results on home lab ~ 4t/s(在新标签页中打开)
- I pushed Kimi K3 onto one CPU with 8 GB of RAM(在新标签页中打开)
- Weight-Aware Streaming Tensor Engine: run Kimi K3 using 29 GB of RAM at 0.50 tok/s(在新标签页中打开)
DeepSeek V4 Flash更新. 这款304B(13B激活)模型现在以60%更低的成本媲美GPT-5.6 Luna,在Mac上运行速度达到30+ tok/s。早期使用者报告称工具调用较为脆弱。
- [AINews] not much happened today(在新标签页中打开)
- Deepseek V4 Flash is now ~#2 open weight model to Kimi K3 and >50x cheaper(在新标签页中打开)
- New Deepseek Flash model matches OpenAI's GPT-5.6 Luna at roughly 60 percent lower cost(在新标签页中打开)
- deepseek-ai/DeepSeek-V4-Flash-0731(在新标签页中打开)
- DeepSeek-V4-Flash-0731 now far surpassing the DeepSeek-V4-Pro-Preview in benchmarks(在新标签页中打开)
- DeepSeek v4 Flash for DS4 (DwarfStar) GGUF w/ DSpark MTP Head(在新标签页中打开)
- DeepSeek-V4-Flash-Q4KExperts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-imatrix-0731.gguf(在新标签页中打开)
- DeepSeek-V4-Flash-0731 unsloth gguf on A100(在新标签页中打开)
- Unsloth Deepseek V4 0731 GGUF's are UP!(在新标签页中打开)
- Deepseek v4 flash MXFP4 (original quality) ggufs(在新标签页中打开)
- What speeds are everyone getting with deepseek v4 flash 0731?(在新标签页中打开)
- DeepSeek-V4-Flash-0731: Models you can run locally now have the intelligence score of the top frontier model from March 2026(在新标签页中打开)
- DeepSeek-V4-Flash-0731 on Bosgame M5 with RTX PRO 6000 Max-Q eGPU(在新标签页中打开)
- Ran DS V4-Flash-0731 Locally on 3xMI50 32GB @ ~15 t/s TG(在新标签页中打开)
- DeepSeek V4 Flash 0731 IQ2_M benchmark for Dual 3060 and 96GB RAM ≈ 3.5 tok/s.(在新标签页中打开)
- Fix for Deep Seek v4 Flash 0731 tool calling has been added to llama cpp(在新标签页中打开)
- DeepSeek V4 Flash 0731 local setup gotcha: model, tool call & config setting(在新标签页中打开)
- Deepseek v4 flash 0731 still not holding up.(在新标签页中打开)
Opus 5推理能力提升. Claude Opus 5在ARC-AGI-3上得分30.2%,几乎是此前最佳成绩的四倍。其Mythos变体还自主破解了NIST加密候选算法HAWK。
Inkling Small高效推理. Thinking Machines发布了一款Apache 2.0模型,拥有12B激活参数,在编程基准测试上优于更大的同类模型,同时使用的token数量少得多。
机器人躯体变得更智能. Google DeepMind的Gemini Robotics 2处理全身运动和精细操作,同时一款新模型支持多机器人协作和人形机器人的实时视频推理。
- Google reveals Gemini Robotics 2.0, promising improved dexterity and safety(在新标签页中打开)
- Gemini Robotics ER 2: powering robotics with video understanding, task orchestration, and multi-robot collaboration(在新标签页中打开)
- Google DeepMind’s new AI model can control a robot’s entire body(在新标签页中打开)
- Google Deepmind unveils Gemini Robotics 2 to power robots of all shapes from tabletop arms to humanoids(在新标签页中打开)
工具与基础设施
MCP 2.0加强智能体安全. 更新后的协议加强了智能体与工具交互的安全性,并迅速获得Dropbox和社区的采用,使本地智能体开发更加安全。
智能体的提示词更精简. LangChain的Deep Agents v0.7利用上下文工程建议将输入token削减了65%,同时其新的Gateway在API调用前强制执行支出限制,并对PII进行脱敏处理。
量化缩减显存需求. 社区创新如BeeLlama的方差归一化量化和ROCmFPX,使大型模型能够在普通硬件上运行,尽管一些MoE设置面临额外的内存使用。
- BeeLlama.cpp v0.4.1: KVarN, KV precision tail, q2_0-q3_1 KV cache, improved support. KLD benchmarks: tail 1024 makes kvarn5 and q6_0 match q8_0, for much less VRAM(在新标签页中打开)
- DeepSeek V4 Flash, up to 32 tok/s on AMD Ryzen AI MAX+ 395(在新标签页中打开)
- PSA: llama.cpp now loads MTP tensors by default for any draft-mtp arch, even with MTP disabled(在新标签页中打开)
行业与投资
SpaceX以600亿美元收购Cursor. 此次收购Anysphere(Cursor AI编辑器的制造商)是最大的AI工具交易之一,同时这家初创公司还推出了印度优惠计划。
微软统一Copilot. 微软计划打造一个融合聊天、编码和智能体功能的单一‘超级应用’,CEO Nadella警告不要锁定特定前沿实验室。
数十亿个人智能体. Meta的CEO预计将有数十亿人通过WhatsApp使用AI智能体处理财务和健康事务,该公司还将按结果销售企业AI服务。
AI资本支出达到新高. 亚马逊承诺2026年投入2200亿美元,谷歌的预测吓坏了投资者,NVIDIA向Ilya Sutskever的SSI提供了Vera Rubin GPU。欧盟正以100亿欧元资助AI超级工厂。
- Ilya Sutskever’s Safe Superintelligence partners with Nvidia to scale its AI research(在新标签页中打开)
- Nvidia invests in Ilya Sutskever's AI lab, shifting SSI away from Google chips(在新标签页中打开)
- AI’s finally expensive enough to make Wall Street nervous(在新标签页中打开)
- Investors love AI, as long as you’re a cloud host(在新标签页中打开)
- EU pools up to €30 billion for AI gigafactories while US tech giants casually spend 20 times more(在新标签页中打开)
这是本周回顾 – 下周日见。