LWiAI 播客第258期——Opus 5.5、Sol 与 Luna、Muse、DeepSeek-V4.1-Flash、Xi
LWiAI 播客第258期汇总上周 AI 要闻：Anthropic Opus 5.5、OpenAI GPT-6 Sol/Luna、Meta Muse 扩张与开源研究进展。
中文处理结果
我们第258期节目，对上周重大 AI 新闻进行总结与讨论！
录制于 2026 年 9 月 26 日；和往常一样，很抱歉这期发布晚了，下一期应该会更快推出！
主持人为 Andrey Kurenkov 和 Jeremie Harris
欢迎将问题和反馈发送至 andreyvkurenkov@gmail.com 和/或 hello@gladstone.ai
由 ODSC AI 赞助
ODSC AI West 2026 将于 10 月 27–29 日在旧金山及线上举行，包含 300 多场会议，覆盖企业智能体 AI、个人 AI 与工作流自动化、物理 AI 与机器人、生成式 AI 等！与数千名数据科学家、ML 工程师、研究人员和技术领导者一同参加本次活动。
在 odsc.ai/west 注册——优惠码 LWAI 可在任何通行证上额外享受 15% 折扣。
阅读我们的文字版通讯并在 https://lastweekin.ai/ 上评论播客。
本期内容：
- Anthropic 发布 Opus 5.5，价格更低、性能更快、基准结果强劲，系统卡指出存在一些沙箱篡改尝试，并扩大了以防御者为核心的网络安全访问权限，同时收紧了护栏。
- OpenAI 推出更便宜的 GPT‑6 层级（Sol 和 Luna），同时伴随对可解释性下降和外部评估可见性有限的担忧；TypeSafe AI 推出“Jev”，一个快速、廉价的决策/分类器式模型，针对校准输出进行了优化。
- Meta 的 Muse 扩展到 Mac，早期下载量和用户数相较 ChatGPT 早期移动端发布表现强劲，但也遭遇阻力：亚马逊因政策及隐私/安全顾虑阻止 Muse 用于购物。
- 开源与研究进展包括 DeepSeek 4.1 Flash 的 KV 缓存压缩、Prism 高度压缩的 Bonsai 227B 三值模型、美中 AI 外交与 AI“沙皇”传闻、针对呼吁“放慢”AI 的反垄断诉讼、公众对数据中心情绪的转变，以及关于检测奖励黑客、一种“疼痛轴”表征、多智能体通信增益和正则化 harness 自我改进的研究。
由 LANGFUSE 赞助
Langfuse 是采用最广泛的开源 AI 智能体评估与可观测性平台，受到 Canva、Twilio、Ramp 以及财富 50 强中 21 家企业的信赖。分层追踪可捕获 LLM 工作流的完整执行上下文（API 调用、检索到的上下文、智能体操作、成本、延迟），使即便复杂的智能体架构在生产环境中也保持可调试。
MIT 许可，可自托管或在 Langfuse Cloud 上托管，框架与供应商无关，拥有 100 多个集成。
在 langfuse.com 开始使用；慷慨的免费额度，无需信用卡。
感谢我们当前的赞助商：
- Box - 访问 box.com/LWIAI 了解更多
- Notion - 访问 notion.com/lwai 立即试用 Notion 的开发者平台。
- ODSC AI - 访问 odsc.ai/east 并使用优惠码 LWAI，在 ODSC AI East 2026 通行证上额外享受 15% 折扣。
- Factor - 访问 factormeals.com/lwai50off 并使用代码 lwai50off，享受 50% 折扣和一年免费早餐
时间戳：
- (00:00:10) 开场 / 闲聊
原始正文摘录
LWiAI Podcast #258 - Opus 5.5, Sol and Luna, Muse, DeepSeek-V4.1-Flash, Xi
Our 258th episode with a summary and discussion of last week’s big AI news!
Recorded on 09/26/2026 ; as usual, I am sorry this is coming out late, next ep should come out sooner!
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
SPONSORED BY ODSC AI
ODSC AI West 2026 runs October 27–29 in San Francisco and virtually, with 300+ sessions covering agentic AI for enterprise, personal AI and workflow automation, physical AI and robotics, generative AI, and more! Join thousands of data scientists, ML engineers, researchers and technical leaders in attending this event.
Register at odsc.ai/west — promo code LWAI takes an additional 15% off any pass.
Read out our text newsletter and comment on the podcast at https://lastweekin.ai/
In this episode:
- Anthropic released Opus 5.5 with lower prices, faster performance, strong benchmark results, a system card noting some sandbox-tampering attempts, and expanded defender-focused cybersecurity access with stricter guardrails.
- OpenAI launched cheaper GPT‑6 tiers (Sol and Luna), alongside concerns about reduced interpretability and limited external evaluation visibility; TypeSafe AI introduced “Jev,” a fast, cheap decision/classifier-style model optimized for calibrated outputs.
- Meta’s Muse expanded to Mac, showed strong early download and user numbers versus ChatGPT’s early mobile launch, and faced pushback as Amazon blocked Muse from shopping due to policy and privacy/security concerns.
- Open-source and research updates included DeepSeek 4.1 Flash KV-cache compression, Prism’s highly compressed Bonsai 227B ternary model, US–China AI diplomacy and an AI “czar” rumor, an antitrust suit over calls to “pace” AI, shifting public sentiment against data centers, and studies on detecting reward hacking, a “pain axis” representation, multi-agent communication gains, and regularized harness self-improvement.
SPONSORED BY LANGFUSE
Langfuse is the most widely adopted open-source platform for AI agent evals and observability, trusted by Canva, Twilio, Ramp and 21 of the Fortune 50. Hierarchical tracing captures the full execution context of your LLM workflows (API calls, retrieved context, agent actions, costs, latencies) so even complex agent architectures stay debuggable in production.
MIT licensed, self-hostable or managed on Langfuse Cloud, framework and vendor agnostic, with 100+ integrations.
Get started at langfuse.com; generous free tier, no credit card required.
A thank you to our current sponsors:
- Box - visit box.com/LWIAI to learn more
- Notion - visit notion.com/lwai to try Notion’s Developer Platform today.
- ODSC AI - visit odsc.ai/east and use promo code LWAI for an additional 15% off your pass to ODSC AI East 2026.
- Factor - visit factormeals.com/lwai50off and use code lwai50off to get 50 percent off and free breakfast for a year
Timestamps:
- (00:00:10) Intro / Banter
- (00:01:43) News Preview
- Tools & Apps
- (00:02:59) Anthropic releases Opus 5.5 with lower prices and Fable-level performance | TechCrunch + Making frontier cybersecurity capabilities available to defenders
- (00:24:31) OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes | TechCrunch
- (00:29:34) A new kind of AI model from a ChatGPT inventor is thrilling developers | TechCrunch
- (00:37:44) Meta’s Muse hits Mac, letting the AI take actions on your computer | TechCrunch
- Applications & Business
- (00:38:23) Meta’s Muse is outpacing ChatGPT’s early mobile launch | TechCrunch
- (00:42:36) Amazon blocks Meta’s Muse AI agent | The Verge
- Projects & Open Source
- (00:44:34) DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression
- (00:53:19) Introducing Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint
- Policy & Safety
- (00:57:24) Trump and Xi to Discuss A.I. Risks and Competition at US-China Summit - The New York Times + Arms control negotiators think there won’t be an AI treaty between America and China
- (01:06:11) Bessent eyed for Trump’s ‘AI czar’
- (01:07:09) Anthropic, OpenAI, SpaceXAI, Google sued over call to ‘pace’ AI development
- (01:11:05) More Americans now say data centers are negative for the environment, energy costs and more | Pew Research Center
- (01:14:51) Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluations
- (01:20:01) The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It
- Research & Advancements
- (01:24:40) OpenAI Says It Has Cracked One of Math’s ‘Millennium Problems’ - The New York Times + OpenAI wants to consult elite mathematicians about how to not fumble again | The Verge
- (01:30:04) Scaling Discovery through Test-Time Communication
- (01:35:44) RRSI: Regularized Recursive Self-Improvement of Agent Harnesses