2026年9月16日
AI 消息
这周值得知道的 AI 消息。
Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking
Google DeepMind released Gemini 3.8 Live and 3.8 Live Extended Thinking, real-time speech models that detect and switch between 97 languages in the middle of a conversation. Extended Thinking takes the top spot on the Artificial Analysis speech-to-speech index at 82.6. Every clip they speak carries a SynthID watermark.
看原文 →Siri AI, a profoundly more capable and personal assistant, is here
Apple's rebuilt Siri is rolling out in beta, in English first, with French, Japanese, Korean, Portuguese and Spanish arriving next month. It runs partly on the device and partly on Apple's Private Cloud Compute servers, on models Apple developed with Google.
看原文 →Perplexity's local agent comes to Windows, powered by NVIDIA RTX
Perplexity's Portable Computer agent now runs on your own Windows PC instead of in the cloud, on RTX cards with 24GB of VRAM or more. It uses a local Qwen 3.8 27B model, keeps your files on the machine, and work done locally doesn't spend your cloud credits.
看原文 →Inside 'Project Lily': the humans reading your ChatGPT chats
404 Media obtained internal material showing OpenAI hired hundreds of contractors to read real ChatGPT conversations and score the answers on a 1-to-7 scale. Most of the service's 900 million-plus users have no idea a person may read their chat. Anthropic confirmed it also uses human review.
看原文 →Google opens Anthropic's Claude to all of its engineers
Google gave every engineer access to Claude Opus 5 through its internal Antigravity IDE, breaking with a rule that pushed staff toward Gemini. Google says Gemini is still its primary model for internal development and that outside models come with a per-user quota.
看原文 →OpenAI boss says world 'right to be afraid' but should trust AI firms
OpenAI's Sam Altman told the BBC the world is right to be afraid of AI, but argued people should trust companies like his to build it responsibly.
看原文 →We don’t need AI regulation — leave safety to us, Nvidia’s Jensen Huang says
Nvidia's Jensen Huang says AI doesn't need new rules: "Safety is an engineering problem, not a legal one. We're developing software after all." He argues companies already avoid shipping unsafe products to protect their own reputation, and that having to choose between speed and safety is a false choice.
看原文 →配音与开源
现在值得一看的开源语音、配音工具。
stronghamjji/PersoDub
一款桌面应用，用原说话人自己的声音把视频配成另一种语言。它不会为了对上时长硬把声音拉长或压短，而是让每句话都落在原本说话的时间段里。里面还带了一个配音助手：跟它说"把第4句重新翻译一下，让长度合适"，它就只重新翻译那一句，并重新生成语音。全部在你自己的电脑上运行。
看原文 →
Vincentwei1021/anything2explainer
A new GitHub tool called "anything2explainer" generates narrated explainer videos from any topic using Claude Code and Codex, drawing each frame with code.
看原文 →
kruzovic7/ai-data-extractor
A free, open-source tool that pulls your chat history out of AI coding assistants, including Claude Code, Cursor, Windsurf, Aider and Cline/Roo Code.
看原文 →
本周 X 上的讨论
这周 AI 圈在 X 上聊得最多的。
Google DeepMind (@GoogleDeepMind)
We're launching AlphaGenome Atlas: an AI-powered searchable database mapping the predicted impact of all 9 billion possible single-letter DNA changes. Here's how it could help researchers better understand our biology
看原文 →OpenDesign (@OpenDesignHQ)
We benchmarked DeepSeek V4.1 Flash by @deepseek_ai. It reached 98% of GPT-6 Astra's score at 1.4% of the cost on everyday design tasks based on user requests. Every model except Astra scored lower AND cost more. Are open models overtaking closed ones?
看原文 →Akshay (@akshay_pachaar)
Brilliant paper by NVIDIA. they found a way to make KV cache transferable between models. the target model skips prefill entirely, and the conversion runs 2.7 to 25x faster than processing the context again.
看原文 →el.cine (@EHuanglu)
never thought AI could do video editing. now i think my 14 yrs editing experience is done. GPT-6 Astra watched 84 clips, edit, trim, find music in my library and even generated subtitles on davinci resolve all by itself.. this is so scary
看原文 →Cognition (@cognition)
Introducing Devin Voice. Your favorite AI software engineer just got a landline. You say it, Devin ships it. Powered by GPT-Live and our new SWE-2 model.
看原文 →用邮件接收下一期
8 种语言任选，随时可以退订。