LWiAI 播客 #252：GPT 5.6、Grok 4.5、Nemotron-Labs-Diffusion 与 AI 2040

中文处理结果

来自 Last Week in AI（Andrey）的说明：我回来了！很抱歉让 Substack 暂停更新，工作有点太繁重，所以落后了。我会尽力恢复正常的发布，从补上这里缺失的播客节目开始（这不是上周的……但我想还是应该发布，抱歉刷屏！）

我也暂停了付费订阅，直到我能恢复持续更新。对于不稳定表示歉意，感谢你的订阅！

我们的第252期节目，总结并讨论了上周的重大AI新闻！

录制于2026年7月11日

主持人：Andrey Kurenkov 和 Jeremie Harris

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

本期内容：

- OpenAI 公开发布了 GPT-5.6（包括 Sol 和 Luna），并将其桌面智能体编码产品更名为 ChatGPT Work；与此同时，关于美国政府是否实际放行并推迟了发布存在争议，人们也担心前沿模型监管零散且随意，以及模型可被越狱的问题。

- 新模型发布加剧了定价和能力竞争：SpaceX AI 的 Grok 4.5 以极低价格、Opus 级编码模型的身份推出，安全文档极少；Meta 则发布了 Muse Spark 1.1，定价激进，在编码/网络基准上大幅提升，并附有冗长的安全评估。

- Meta 还预览了 Muse Video，并推出了 Muse Image，但在因可轻松生成公共 Instagram 账户图像而遭到强烈反对后迅速撤回；另外，随着成本压力上升，中国开源模型占 OpenRouter 每周 token 的比例超过 30%，同时讨论了个别内部人员威胁等风险。

- 基础设施、政策和安全方面的发展包括：Meta 探索将 AI 算力作为云业务出售；美国能源监管机构敦促电网运营商处理大负荷数据中心接入；Anthropic 发表了一种‘全局工作空间’可解释性方法，用于可口头化的内部表征；有报道称中国可能限制海外访问顶级模型；AI 2040 提议美中协调放缓进展，直到对齐得到改善。

时间戳（注意：这些时间戳未考虑动态插入的广告，因此可能偏移几分钟）：

- (00:00:10) 开场 / 闲聊

- (00:01:33) 新闻预览

- 工具与应用

- (00:02:03) OpenAI 在政府放行后推出 GPT-5.6，并宣布‘ChatGPT Work’ | The Verge + 新的 ChatGPT 超级应用瞄准 Claude Desktop + OpenAI 正在关闭其 Atlas 网络浏览器 + 英国机构称，OpenAI 最新 AI 模型可能具有与导致美国对 Anthropic 的 Fable 实施出口管制类似的网络漏洞

- (00:15:41) SpaceXAI 与 Cursor 为金融、法律应用推出 Grok 4.5 AI 模型 - Bloomberg + SpaceXAI 的 Grok 4.5 在编码智能体定价上低于 Anthropic 和 OpenAI

原始正文摘录

LWiAI Podcast #252 - GPT 5.6, Grok 4.5, Nemotron-Labs-Diffusion, AI 2040

Note from Last Week in AI (Andrey): I’m back! And i’m sorry for putting the substack on a silent pause, work got a bit too overwhelming so I fell behind on this. I’ll do my best to resume normal posting, starting with catching up on podcast episodes missing from here (which are not last week… but I guess I should post them still, sorry for the spam!)

I’ve also paused paid subscriptions until I can get back to consistent posting. Apologies for the flakiness, and thanks for being a subscriber!

Our 252th episode with a summary and discussion of last week’s big AI news!

Recorded on 07/11/2026

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:

- OpenAI publicly rolled out GPT-5.6 (including Sol and Luna) and rebranded its desktop agentic coding product as ChatGPT Work, amid disputed claims about whether the US government effectively green-lit and delayed the release and concerns about inconsistent, ad hoc frontier-model oversight and jailbreakability.

- New model releases intensified pricing and capability competition: SpaceX AI’s Grok 4.5 launched as a very low-cost, Opus-class coding model with minimal safety documentation, while Meta released Muse Spark 1.1 with aggressive pricing, large coding/cyber benchmark gains, and a lengthy safety evaluation.

- Meta also previewed Muse Video and rolled out Muse Image before quickly backtracking after backlash over easy generation of images of public Instagram accounts; separately, Chinese open-source models grew to over 30% of weekly OpenRouter tokens as cost pressure increased, alongside discussion of risks like potential insider threats.

- Infrastructure, policy, and safety developments included Meta exploring selling AI compute as a cloud business, US energy regulators pressing grid operators on large-load data-center connections, Anthropic publishing a “global workspace” interpretability method for verbalizable internal representations, reports that China may restrict overseas access to top models, and AI 2040 proposing US–China coordination to slow progress until alignment improves.

Timestamps (note - these don’t take into account dynamically inserted ads and therefore may be off by a couple of minutes):

- (00:00:10) Intro / Banter

- (00:01:33) News Preview

- Tools & Apps

- (00:02:03) OpenAI rolls out GPT-5.6 after government greenlight — and announces ‘ChatGPT Work’ | The Verge + The new ChatGPT superapp takes aim at Claude Desktop + OpenAI is shutting down its Atlas web browser + OpenAI’s latest AI model likely has similar cyber vulnerabilities to one that led to U.S. export controls on Anthropic’s Fable, British agency says

- (00:15:41) SpaceXAI, Cursor Launch Grok 4.5 AI Model for Finance, Legal Applications - Bloomberg + SpaceXAI’s Grok 4.5 Undercuts Anthropic and OpenAI on Coding Agent Pricing

- (00:20:29) Meta says its new AI model is ready to compete on coding | The Verge

- (00:27:21) Introducing Muse Image and Muse Video + https://www.nytimes.com/2026/07/10/technology/meta-muse-images-instagram-removal.html

- (00:29:01) Chinese AI models gain ground with U.S. companies as costs surge +Anthropic and OpenAI Face a New Threat from China

- Applications & Business

- (00:35:21) Meta Is Planning a Cloud Business to Sell AI Computing Power - Bloomberg

- (00:46:32) US energy regulator sets ultimatum for data centres + Grid operator PJM orders emergency steps to avoid large-scale US power outages

- Projects & Open Source

- (00:51:40) Nemotron-Labs-Diffusion: A Tri-Mode Language Model Unifying Autoregressive, Diffusion, and Self-Speculation Decoding

- (00:57:44) Tencent Releases Hy3: An Open 295B Mixture-of-Experts (MoE) Model with 21B Active Parameters and 256K Context - MarkTechPost

- Policy & Safety

- (00:58:30) Verbalizable Representations Form a Global Workspace in Language Models

- (01:09:29) Beijing is looking at curbing overseas access to China’s top AI models, sources say

- (01:12:53) The ex-OpenAI employee behind ‘AI 2027’ recommends a rosier path - The Washington Post + AI 2040: Plan A

查看原始发布