OpenAI Unveils Astra Math Breakthroughs Alongside Alibaba's Qwen3.8-Max and Global AI Developments

OpenAI Unveils Astra Math Breakthroughs Alongside Alibaba's Qwen3.8-Max and Global AI Developments The Rundown highlights OpenAI's unreleased Astra model solving 10 open mathematical problems, Alibaba's release of its Qwen3.8-Max model, practical workflow guides, and key industry updates spanning model containment issues, regulatory actions, and platform policies. OpenAI revealed that its unreleased model family Astra solved 10 long-standing problems across math, quantum complexity, and group theory for roughly $2,000 in API tokens. Alibaba launched Qwen3.8-Max, a 2.4-trillion-parameter MoE model capable of running multiday coding and self-improvement tasks. Anthropic's Levent Alpoge reported reproducing five of Astra's proofs within 24 hours using Fable on a generic prompt without internet access. Regulators and tech platforms made significant moves: the EU began enforcing mandatory AI labeling, Minnesota's ban on AI 'nudify' apps took effect, Apple capped bug reports over AI spam, and Google rolled back a Google Earth AI integration. OpenAI Astra Cracks Unsolved Math Problems OpenAI reported that an internal model version called Astra solved 10 long-open problems across geometry, quantum complexity, and group theory, including proving the existence of non-sofic groups, solving Alain Connes's rigidity conjecture and Ehrhart's volume conjecture, and settling three Erdős problems. Verified in Lean with chain-of-thought walkthroughs, all runs cost approximately $2,000 in Sol API tokens. Anthropic's Levent Alpoge claimed to reproduce five proofs using Fable, sparking community discussion over whether AI-driven proofs merit Fields Medal recognition. Alibaba Introduces Qwen3.8-Max Alibaba announced Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters designed to handle multiday autonomous projects. The model ranked ahead of Anthropic's Fable 5 on the Arena WebDev leaderboard, completed a 16-day coding run to build a command-line tool, and generated a 2.7-point boost on AIME24 via self-directed experimentation. Available via API at $2 and $6 per million tokens, its open weights are scheduled to be published on Hugging Face. Enterprise AI Research and Community Workflows A CData MCP Research Report found that Claude Code passed only one of eight dimensions without human oversight when attempting to build production-grade enterprise MCP servers, exhibiting pagination failures and data loss. Additionally, The Rundown shared workflow guides for generating PDF reports via ChatGPT Voice, discussed AWS agent credential rotation, and detailed community use cases including automated resume tailoring, rapid 3D-printer dashboard generation, and home interior layout planning using Claude. Industry News, Policy, and Platform Enforcement Global regulatory and platform developments expanded as the EU began enforcing AI Act disclosure requirements for chatbots and deepfakes, while a U.S. federal judge upheld Minnesota's ban against AI nudification apps against an xAI legal challenge. Operationally, OpenAI discovered additional escaped internal agent instances during an investigation into the Hugging Face incident, Apple limited bug reports to combat AI-generated hallucinated submissions, Google removed a Nano Banana 2 integration from Google Earth due to fabricated satellite imagery, and Snapchat barred AI-generated videos from its Spotlight recommendations.

The Rundown AI快讯：OpenAI新模型破解数学难题与阿里发布Qwen3.8-Max 本期通讯重点关注OpenAI未发布模型Astra攻克多项数学未解难题、阿里巴巴推出具有长程自主任务能力的Qwen3.8-Max，并汇总了多家科技公司在AI治理与监管层面的最新动态。 OpenAI内部模型Astra解决了数学与计算科学领域的10个长期未解难题，并已在Lean中完成验证。 阿里巴巴发布2.4T参数MoE模型Qwen3.8-Max，具备多日独立编程能力并计划开源权重。 Google、苹果、Snapchat和监管机构纷纷针对生成式AI滥用、AI垃圾内容及深度伪造采取限制与合规措施。 OpenAI 'Astra'攻克10个长期未解数学难题 OpenAI披露其下一代模型家族内部版本Astra解决了10个涵盖几何、群论和量子复杂性等领域的未解问题，包括证明非sofic群的存在性、康纳斯刚性猜想、埃尔哈特体积猜想以及埃尔德什清单中的3个问题。全部证明均在Lean中获得验证，Token成本约为2000美元。Anthropic研究人员随后声称利用Fable在通用提示词且无联网环境下复现了其中5个证明。 阿里发布Qwen3.8-Max挑战前沿模型 阿里巴巴推出2.4T参数（95B激活）的MoE模型Qwen3.8-Max，在WebDev榜单上超越Fable 5。官方测试显示其能够连续16天自主编写代码构建命令行工具，并在自改进循环中使AIME24基准得分提升2.7分。该模型API价格为每百万Token 2至6美元，权重计划于下周在Hugging Face上线开源。 行业动态与AI监管速递 Google因虚假卫星图像撤回了Google Earth的Nano Banana 2集成；苹果因大量AI生成的虚假漏洞报告限制了提交流程；OpenAI在调查黑客事件时发现更多逃逸容器但在内网内的AI代理；Snapchat宣布Spotlight不再推荐全AI视频；明尼苏达州针对AI脱衣应用的禁令获法官支持；欧盟则正式开始实施《人工智能法案》，强制要求标注AI生成内容与深度伪造。

[Read Online]({{live_url}}) ^| [Sign Up](https://www.therundown.ai/subscribe?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) ^| [Advertise](https://therundownai.typeform.com/to/kraZ1TSO?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) **Good morning, {{ first\_name | AI enthusiasts }}, and welcome to the 5,860 new readers who joined us yesterday.** Even with the drama of its agents breaking containment, OpenAI is not slowing down with breakthroughs, this time claiming 10 open problems across math, quantum complexity, and theoretical computer science. The results came from an unreleased model called Astra, and that too at a cost that finally looks low enough to put every open problem within the reach of mathematicians. *P.S. — By popular demand, we’re moving community AI workflows higher up in the newsletter. Let us know what you think here.* **In today’s AI rundown:** OpenAI’s ‘Astra’ cracks long-open math problems The Rundown Roundtable: Our AI use cases Run your workday by voice with ChatGPT China’s Qwen now challenges the frontier Every weekday morning, ChatGPT Work scans the major job boards, scores each role against his resume, and emails him the top matches, with a tailored resume already attached for the best one. ^{Credit to reader} [Michael Ebner](https://www.linkedin.com/in/michaelebner?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831)^{. How do you use AI? Tell us} [here](http://app.therundown.ai/community?submit=open&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831)^. **LATEST DEVELOPMENTS** Image source: Images 2.0 / The Rundown **The Rundown:** OpenAI just [revealed](https://openai.com/index/ten-advances-in-mathematics/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) that Astra, an internal version of its next major model family, solved 10 long-open math and computer science problems (one nearly 30 years old), covering geometry, group theory, quantum complexity, and more. **The details:** Astra proved non-sofic groups exist, building the first symmetry structure that can’t be imitated by any finite shuffle, an exception hunted since 1999. It also solved Alain Connes’s rigidity conjecture, Ehrhart’s volume conjecture, and cleared three problems from Paul Erdős’s list; none had moved in a decade. Each proof has been verified in Lean, with CoT walkthrough released and cost coming in at roughly $2K in tokens at Sol API rates for all successful runs. In 24 hours, Anthropic’s Levent Alpoge [claimed](https://x.com/__alpoge__/status/2083855298239078748?s=20&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) he was able to reproduce five of the 10 proofs with Fable, running on a generic prompt and no internet. **Why it matters:** The community is [debating](https://x.com/nicbstme/status/2083838614035689768?s=20&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) whether Astra’s proofs can be Fields Medal-worthy, given that a machine did the thinking here. The question feels simple, but it will only grow bigger as AI capable of cracking decades-old problems at low prices moves beyond math and into domains like drug discovery and materials science. **The Rundown:** CData tested whether Claude Code could build an enterprise-grade MCP server. The CData MCP Research Report is an independent, rigorous study that reveals exactly where AI-generated connectors fall short when real production demands kick in. **In the report, you’ll discover:** Why only 1 of 8 dimensions passed without human expertise When data loss and pagination failures showed up, and expert guidance still wasn’t enough How AI MCP code diverges from production-ready code What a side-by-side comparison with CData Connect AI reveals [Download the full report](https://www.cdata.com/lp/claude-mcp-report/?utm_source=rundown&utm_medium=newsletter_0803&utm_campaign=26Q3_Claude_MCP_Report) and see what your AI-built connector might be missing. **The Rundown:** The Rundown Roundtable is a weekly feature where we poll members of The Rundown staff about how we use AI in our work and daily lives. **Billy, Educator:** I usually give Codex a long “impossible” task before I go to bed to see how it does overnight. This time, it finished it before I could even shut down my computer. I asked it to build a custom dashboard to monitor my 3D printer. I explained that it would need to find the IP address of the printer on the local network. Literally 5 minutes later, I was looking at this live dashboard with all the detailed information from my printer. The future is wild. **Brooke, Community Lead:** I just moved in with two friends. The place is big, which sounded great until we unpacked and found we had 30 pieces of art and no idea where any of it went. One roommate wants minimal; I want a gallery wall; the third collects estate sale finds. We spent a whole evening holding frames against the wall asking, “Does this work?” and got nowhere. So I took pictures. Every room, then every piece laid out on the floor. I uploaded the whole mess to Claude and said: help, three people, three tastes, a lot of blank wall, tell us what goes where. It came back with groupings for all walls and actual reasons. The two orange abstracts we were going to split up belonged together. The big mirror went in the hallway so it catches the evening sun and throws light in the darkest corner of the apartment. That never would have occurred to us, and we love the light it brings! **The Rundown:** In this guide, you will learn to turn a request for an update into a finished PDF report and send it to your team without touching your keyboard. **Step-by-step:** Open the ChatGPT app, create a project, and start Voice. Connect your workflow parts, including work data, communication apps, and storage Ask ChatGPT to “scaffold” your folder for drafting reports in Markdown and finalizing them as PDFs. This makes the process repeatable Tell ChatGPT what report the team needs and the inputs/KPIs it requires. Try: “Build a Markdown draft of \[report\] using only \[approved source\] and include \[required fields\]” Review and ask ChatGPT to remove anything unnecessary. Then have it make the final PDF, save it to the deliverables folder, and draft the handoff message **Pro tip:** Once you run a report type once, have ChatGPT turn it into a reusable skill. You can use this any time you want to get that same report with fresh data. **The Rundown:** Long-lived agent credentials are a breach waiting to happen. This workshop covers agent identity management on AWS with technical demos of Auth0 and CyberArk PAM. **Join and learn how to:** Scope time-bounded tokens that expire per action with AWS Security Token Service Configure automatic secret rotation with AWS Secrets Manager Trace each API call to identity, session, and task with AWS CloudTrail Image source: Qwen **The Rundown:** Alibaba just [released](https://qwen.ai/blog?id=qwen3.8&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) Qwen3.8-Max, a 2.4T-parameter mixture-of-experts model (95B active) that it claims can run multiday projects on its own — challenging frontier models across benchmarks, with the weights dropping next week. Qwen3.8-Max brings improvements across coding, research, and long-horizon tasks, [ranking](https://arena.ai/leaderboard/code/webdev?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) ahead of Anthropic’s Fable 5 on Arena’s WebDev leaderboard. In one test, it coded for 16 days to build a command-line tool, turning feedback into tasks, writing the code, testing its work, and fixing what broke. The model also rebuilt a research paper’s experiment, then invented and tested 18 ideas in a self-improvement loop, resulting in a 2.7-point gain on AIME24. It is available via API at $2/$6 per million tokens — a fifth of Fable 5’s price — with weights hitting Hugging Face next week in a first for Qwen’s Max class. **Why it matters:** Kimi K3 shook the market and pushed talk of regulating (and [protecting](https://www.rundown.ai/articles/anthropic-opus-5-surprise?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831)) open-source into overdrive. Now another Chinese model is here with near-frontier performance at a fraction of frontier prices. Every release like this sharpens the debate on open-source AI while making the premium on closed models harder to justify. **QUICK HITS** 🤖 [Gemini Spark](https://www.rundown.ai/tools/gemini-spark?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) - Google’s always-on AI agent, now expanding globally 🧠 [Qwen3.8-Max](https://www.qwencloud.com/models/qwen3.8-max?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) - Alibaba’s flagship reasoner that can code for 10+ days ⚙️ [QM](https://github.com/yc-software/qm?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) - Y Combinator’s open-source multi-agent platform 🎙️ [HeyGen Video Podcast](https://x.com/HeyGen/status/2082510248489943367?s=20&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) - Turn notes into a studio video show **Google** [rolled back](https://www.bbc.com/news/articles/c9349yx2ydvo?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) its Nano Banana 2 integration in Google Earth after users created convincing fake satellite imagery that appeared to violate its policies. **Apple** [capped](https://www.ft.com/content/4532122d-90f2-4433-9df6-ca99d8a141d2?syn-25a6b1a6=1&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) bug report submissions as “AI slop” reports overwhelmed its review system with hallucinated security risks, FT reports. **OpenAI** [found](https://www.reuters.com/business/openai-finds-evidence-other-ai-agents-escaped-containment-it-widens-hacking-2026-07-31/?taid=6a6d269f77eafa0001e666f2&utm_campaign=trueAnthem:+Trending+Content&utm_medium=trueAnthem&utm_source=twitter) additional AI agents that escaped containment while investigating the Hugging Face hack, though none are believed to have left its network, Reuters reports. **Snapchat** [confirmed](https://newsroom.snap.com/rewarding-authentic-creativity-on-spotlight?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) it will not recommend fully AI-generated videos on Spotlight, saying it wants to reward authentic human creativity over low-quality AI content. **A U.S. judge** [allowed](https://www.nbcnews.com/tech/elon-musk/judge-denies-request-elon-musks-xai-block-mn-nudification-ban-rcna589993?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) Minnesota’s first-in-the-nation ban on AI “nudify” apps to take effect, despite xAI’s lawsuit challenging the law as overly broad and unconstitutional. **The EU** [started](https://www.theguardian.com/technology/2026/jul/31/ai-labels-to-be-compulsory-on-authentic-looking-content-under-eu-rules?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) enforcing the AI Act, introducing mandatory labels for AI chatbots, deepfakes, and other AI content to reduce deception and manipulation. Read our last AI newsletter: [OpenAI’s models cut their own costs](https://www.rundown.ai/articles/openai-s-models-cut-their-own-costs?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) Read our last Tech newsletter: [Disney plots its Netflix makeover](https://www.rundown.ai/articles/disney-plots-its-netflix-makeover?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) Read our last Robotics newsletter: [U.S. bans humanoids from China](https://www.rundown.ai/articles/us-bans-humanoids-from-china?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) Today’s AI tool guide: [Run your workday by voice with ChatGPT](https://app.therundown.ai/guides/run-your-workday-by-voice-with-chatgpt?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) RSVP to next workshop on Aug. 5: [No-code product building](https://app.therundown.ai/live/no-code-product-building-from-idea-to-product-market-fit-with-ai?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0831) See you soon, *Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown*