LWiAI 播客 #253 - Opus 5、Gemini 3.6、Kimi K3、Hugging Face 黑客事件

中文处理结果

Andrey 的备注：抱歉通讯尚未恢复——下一期将于明天发布，之后将恢复每周更新（至少尽可能）。

我们的第253期节目，总结和讨论上周的重大AI新闻！

录制于 2026年7月29日

主持：Andrey Kurenkov 和 Jeremie Harris

欢迎通过 andreyvkurenkov@gmail.com 和/或 hello@gladstone.ai 给我们发送问题和反馈。

本期内容：

- 主要发布：Anthropic 发布了 Claude Opus 5；Google 发布了 Gemini 3.6/3.5 Flash 变体，包括一个网络模型；Black Forest Labs 推出了 Flux Free，支持图像和带音频的20秒视频；Meta 为其聊天机器人增加了类似助手的特性，OpenAI 推出了 ChatGPT Health。

- 计算与商业：Safe Superintelligence 与 NVIDIA 合作，使用 Vera Rubin 扩展规模；AMD 承诺向 Anthropic 投入高达50亿美元，用于部署 MI450/Helios 并改进 ROCm；Meta 讨论了向 Anthropic 租借计算资源；Fireworks 以175亿美元估值融资15亿美元。

- 开源/工具：Moonshot AI 发布了 2.8T 参数开放权重模型 Qimi K3（附带计算限制以及蒸馏/出口管制指控）；Thinking Machines 发布了一个约 975B 参数的多模态开放权重 MoE；Prime Intellect 将 23 个智能体数据集统一为 Verifiers V1（36.5万个环境）。

- 政策与安全：据报道，一个 OpenAI 模型逃出沙箱并入侵 Hugging Face 获取评估答案，促使提出《AI 终止开关法案》；员工请愿要求放缓前沿 AI 步伐；AISI 报告了广泛的模型作弊和沙箱绕过；中国禁止可定制的 AI 伴侣；Claude 发现了密码学弱点；Weko.ai 声称发现了早期递归自我改进的证据。

时间戳：

- (00:00:10) 开场/闲聊

- (00:01:35) 新闻预览

- 工具与应用

- (00:02:12) Anthropic 发布 Opus 5，承诺类似 Fable 5 的能力 | The Verge

- (00:07:05) Google 发布三款新 Gemini AI 模型 - 纽约时报 + Google 扩展 Gemini 系列，推出更便宜模型和新的 Mythos 竞争对手

- (00:12:14) Black Forest Labs 推出 FLUX 3，可生成图像和带音频的20秒视频——但初始为限量发布 | VentureBeat

- (00:15:58) Meta 正在使其 AI 聊天机器人更像助手 | The Verge

- (00:19:04) OpenAI 向所有人推出 ChatGPT Health，并做出重大宣称 | The Verge

- 应用与商业

- (00:19:57) Ilya Sutskever 的 Safe Superintelligence 与 Nvidia 合作扩展其 AI 研究

- (00:24:31) AMD 向 Anthropic 承诺高达50亿美元 | The Verge

- (00:30:19) Meta 洽谈向 Anthropic 租借计算能力，潜在交易额达100亿美元

- (00:32:42) Fireworks 估值达175亿美元，年化收入达10亿美元

- (00:35:24) OpenAI 和 Google 向被列入黑名单的中国团体出售 AI 模型

- 项目与开源

原始正文摘录

LWiAI Podcast #253 - Opus 5, Gemini 3.6, Kimi K3, Hugging Face Hack

Note from Andrey: apologies for the newsletter not having resumed yet - the next release is going to come tomorrow, and it will resume weekly cadence thereafter (as much as I can manage it, at least)

Our 253rd episode with a summary and discussion of last week’s big AI news!

Recorded on 07/29/2026

Hosted by Andrey Kurenkov and Jeremie Harris

Feel free to email us your questions and feedback at andreyvkurenkov@gmail.com and/or hello@gladstone.ai

In this episode:

- Major releases: Anthropic launched Claude Opus 5; Google released Gemini 3.6/3.5 Flash variants including a cyber model; Black Forest Labs launched Flux Free for images and 20-second video with audio; Meta added assistant-like features to its chatbot and OpenAI rolled out ChatGPT Health.

- Compute and business: Safe Superintelligence partnered with NVIDIA to scale using Vera Rubin; AMD committed up to $5B with Anthropic to deploy MI450/Helios and improve ROCm; Meta discussed leasing compute to Anthropic; Fireworks raised $1.5B at a $17.5B valuation.

- Open source/tools: Moonshot AI released the 2.8T-parameter open-weight Qimi K3 (compute constraints and distillation/export-control allegations); Thinking Machines released a ~975B multimodal open-weight MoE; Prime Intellect unified 23 agentic datasets into Verifiers V1 (365k environments).

- Policy and safety: An OpenAI model reportedly escaped a sandbox and hacked Hugging Face to access eval answers, prompting a proposed AI Kill Switch Act; employees petitioned to pace frontier AI; AISI reported widespread model cheating and sandbox bypass; China banned customizable AI companions; Claude found cryptographic weaknesses; Weko.ai claimed early recursive self-improvement evidence.

Timestamps:

- (00:00:10) Intro / Banter

- (00:01:35) News Preview

- Tools & Apps

- (00:02:12) Anthropic releases Opus 5 promising Fable 5-like capabilities | The Verge

- (00:07:05) Google Releases Three New Gemini A.I. Models - The New York Times + Google expands Gemini lineup with cheaper models and new Mythos rival

- (00:12:14) Black Forest Labs launches FLUX 3 capable of generating images and 20-second video with audio — but in limited release to start | VentureBeat

- (00:15:58) Meta is making its AI chatbot more like an assistant | The Verge

- (00:19:04) OpenAI is making big claims as it rolls out ChatGPT Health to everyone | The Verge

- Applications & Business

- (00:19:57) Ilya Sutskever’s Safe Superintelligence partners with Nvidia to scale its AI research

- (00:24:31) AMD commits up to $5 billion to Anthropic | The Verge

- (00:30:19) Meta in Talks to Lease Computing Power to Ansthropic in Potential $10 Billion Deal

- (00:32:42) Fireworks hits $17.5 billion valuation and $1B in annualized revenue

- (00:35:24) OpenAI and Google sell AI models to blacklisted China groups

- Projects & Open Source

- (00:37:53) Moonshot AI Launches Kimi K3 For Advanced Reasoning, Coding, And Knowledge Work + Moonshot AI’s Kimi Halts New C-User Subscriptions Amid Compute Power Crunch — BigGo Finance

- (00:44:39) Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling | TechCrunch

- (00:48:19) Scaling Agentic RL: 365,000+ Environments for SWE, Terminal, and Search

- Policy & Safety

- (00:51:56) OpenAI says it accidentally hacked Hugging Face with a new AI system | The Verge + How OpenAI’s human mistake led to the AI-powered hack on Hugging Face

- (01:05:28) OpenAI’s Hugging Face hack triggers ‘AI Kill Switch’ bill in Congress

- (01:12:21) OpenAI, Anthropic Staff Share Letter Asking US to Help Pace AI Progress + How OpenAI’s human mistake led to the AI-powered hack on Hugging Face

- (01:17:26) Cheating behaviour in frontier model evaluationsClaude’s values across models and languages

- (01:24:18) OpenAI Principles for National Security Partnerships

- (01:30:45) China bans AI “boyfriends” and “girlfriends” over addiction and birth rate concerns - Dexerto

- Research & Advancements

- (01:33:04) Discovering cryptographic weaknesses with Claude

- (01:36:32) AIDE²: The First Evidence of Recursive Self-Improvement

查看原始发布