2026年9月23日
AI 消息
这周值得知道的 AI 消息。
Meta's AI agent, Muse, is chasing ChatGPT's App Store rise — and hit No. 1 with fewer downloads
Meta's Muse, an AI agent that sends emails, books travel and makes purchases for you, hit No. 1 on the US App Store ten days after launch, even with fewer daily downloads than ChatGPT had at its debut. But in one survey, only 8% said they'd trust Meta with their passwords.
看原文 →Introducing Claude Opus 5.5
Claude Opus 5.5 performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. Anthropic also says it writes more naturally and gets to the point faster.
看原文 →Introducing GPT-6 Sol and Luna
OpenAI released GPT-6 Sol and Luna, cheaper models trained the same way as GPT-6 Astra. API prices are cut in half, and OpenAI says Sol beats Claude Opus 5 on a business workflow test at 9% of the cost per task.
看原文 →Better prompt caching for GPT-6
GPT-6 now reuses more of what you've already sent, so replies come back faster and cost less. OpenAI also added a dashboard and a tool that tells you why something didn't get cached.
看原文 →New Anthropic, OpenAI models make same promise: A little more for a lot less money
Anthropic and OpenAI shipped new models on the same day with the same pitch: a little better than before, for a lot less money.
看原文 →A new skyline of data centres is rising from the rolling plains of rural China
In Ulanqab, Inner Mongolia, once known as China's potato capital, data centres for AI are going up next to the potato fields. China wants AI in 90% of its industries by 2030, and these buildings supply the computing power.
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
A desktop app that dubs your video into another language in the speaker's own voice. The best part is the built-in Dub Agent: tell it "make line 9 fit" and it redoes just that line, both the translation and the voice. Everything runs on your own computer.
看原文 →
FireRedTeam/FireRedTTS3
An open-source voice cloning model that speaks 24 languages and 21 Chinese dialects. It can also design a brand-new voice from a plain description, like age, tone or accent, and edit parts of speech it already made.
看原文 →
0xNatoshi/jev-codex-router
A router for Codex that asks Jev, before every turn, which model to use, how hard to think and how fast to go. The goal is to be just smart enough for each step without burning extra quota.
看原文 →本周 X 上的讨论
这周 AI 圈在 X 上聊得最多的。
Elon Musk (@elonmusk)
"Cool" — quoting @aug5thmusic: "Grok 4.7 scored 100% on my music error detection test, the same as GPT-6 Astra."
看原文 →Jared Palmer (@jaredpalmer)
Kev-0.6B, 4B, and 8B are now available. Kev is a family of small open source Jev-like decision models you can train and run yourself. Out of domain: Kev-8B 79.6%, Jev 85.7%. Kev-4B serves on a 32 GB Mac. Apache 2.0.
看原文 →Mirage (@trymirage)
We rebuilt After Effects and Premiere for GPT-6 Astra and Sol. Introducing Tesseract. The video creative suite for your AI agent. Give your agent direct access to our creative engine. No more clicking through software built for people. → Edit footage, motion, and sound → Made with AI
看原文 →用邮件接收下一期
8 种语言任选，随时可以退订。