据我所知，目前没有任何聊天机器人或 API 能让你访问 IMO 2025 金牌得主级别的模型。但今天，这一现状将被打破！你现在可以在 @huggingface 上通过 Apache 2.0 许可证下载 @deepseek_ai Math-V2 的开源版本，获取模型权重！
想象一下，你可以免费拥有世界上顶尖数学家的大脑，并用它来：
- 探索研究
- 进行微调
- 优化性能
- 在你自己的硬件上运行
没有限制，没有削弱，没有任何公司或政府可以将其收回。这才是 AI 和知识民主化的最佳体现，简直太棒了！🤯🤯🤯
你可以在这里下载权重：huggingface.co/deepseek-ai/De… 。AI 的前沿阵地就在开源社区！
想象一下，一个软件团队在做一个大项目，但有个奇怪的规定：每个工程师只能工作几十分钟，最多几小时，干完就要换一个新的工程师。所以让这个团队完成简单项目任务还行，复杂一点需要长时间运行的项目，比如你让它克隆一个 claude .ai，它就做不到。
这其实就是 Coding Agent 的现状：没有记忆，上下文窗口长度有限。所以要它执行长时间任务，它还做不好。
Anthropic 的这篇博客：《Effective harnesses for long-running agents》，专门讨论了如何让 Agent 在跨越多个上下文窗口时依然能持续推进任务。
先看 Agent 在长任务中遇到的主要问题是什么？
主要三种：
第一种叫一口气干太多。比如你让 Agent 克隆一个 claude .ai 这样的网站，它会试图一次性搞定整个应用。结果上下文还没用完，功能写了一半，代码乱成一锅粥。下一个会话进来，面对半成品只能干瞪眼，花很多时间猜测前面到底做了什么。
第二种叫过早宣布胜利。项目做了一部分，后来的 Agent 看看环境，觉得好像差不多了，就直接收工。功能缺一大堆也不管。
第三种叫测试敷衍。Agent 改完代码，跑几个单元测试或者 curl 一下接口就觉得万事大吉，根本没有像真实用户那样端到端走一遍流程。
这三种失败模式的共同点是 Agent 不知道全局目标，也不知道该在哪里停下来、该留下什么给下一位。
那么 Anthropic 的解决方案是什么呢？
其实就是软件工程的一些现成的解决方案：引入类似人类团队的分工协作机制，将复杂任务拆解成小的可跟踪验证的任务，清晰的交接机制，并严格验证任务结果
一个初始化 Agent，它只在项目启动时出场一次，任务是搭好项目运行环境：有点像架构师的角色，写一个 init .sh 脚本方便后续启动开发服务器，建一个 claude-progress.txt 记录进度，做第一次 git 提交，最关键的是生成一份功能清单。
这份功能清单有多细？在克隆 claude .ai 的案例中，列了超过 200 条具体功能，比如用户能打开新对话、输入问题、按回车、看到 AI 回复。每一条初始状态都标记为失败，后续 Agent 必须逐条验证通过才能改成成功。
而且这里有个细节，这个清单不是用 Markdown 来写的，是一个 JSON 数组，因为 Anthropic 实验发现，相比 Markdown，模型在处理 JSON 时更不容易随意篡改或覆盖文件。
另一个是编码 Agent。在初始化项目后，后续就是它干活了，核心行为准则只有两条：一次只做一个功能，做完要留下干净的环境。
什么叫干净的环境？想象你往主分支提交代码的标准：没有严重 bug，代码整齐有文档，下一个人接手能直接开始新功能，不用先替你收拾烂摊子。
每次开工前，它先做几件事：
– 运行 pwd 看看自己在哪个目录
– 读 Git 日志和进度文件，搞清楚上一轮干了啥
– 看功能清单，挑一个最高优先级的未完成功能
– 跑一遍基础测试，确保 App 还能用
然后专心做一个功能，做完后：
– 写清楚的 Git commit message
– 更新 claude-progress.txt
– 只改功能清单里的状态字段，绝不删改需求本身
这个设计的巧妙之处在于，它把“记忆”外化成了文件和 Git 历史。每一轮的 Agent 不需要依赖上下文窗口里的碎片信息，而是模仿靠谱的人类工程师每天上班会做的事。先同步进度，确认环境正常，再动手干活。
测试环节的改进值得单独说。
原来 Agent 只会用代码层面的方式验证，比如跑单元测试或者调接口。问题是很多 bug 只有用户真正操作页面时才会暴露。
解决方案是给 Agent 配上浏览器自动化工具，比如 Puppeteer MCP。Agent 现在能像真人一样打开浏览器、点按钮、填表单、看页面渲染结果。Anthropic 放了一张动图，展示 Agent 测试克隆版 claude .ai 时自己截的图，确实是在像用户那样操作。
这招大幅提升了功能验证的准确率。当然也有边界，比如浏览器原生的 alert 弹窗，Puppeteer 捕捉不到，依赖弹窗的功能就容易出 bug。
这套方案还留了一些开放问题。
比如，到底是一个通用 Agent 全包好，还是搞专业分工？让测试 Agent 专门测，代码清理 Agent 专门收拾，也许效果更好。
再比如，这套经验是针对全栈 Web 开发优化的，能不能迁移到科研或金融建模这类长周期任务？应该可以，但需要实验验证。
响马 @xicilion 说：
> ai 的尽头依旧是软件工程。
AI Agent 也不是魔法，它一样需要从人类软件工程中汲取经验，它也需要将复杂的任务进行分解成简单的任务，要有一个结构化的工作环境和清晰的交接机制。
人类工程师为什么能跨团队、跨时区协作？因为有 Git、有文档、有 Code Review、有测试。AI Agent 要想长时间自主工作，也得把这些东西搬过来。
Anthropic 的方案，不过是把软件工程的最佳实践变成了 Agent 能理解的提示词和工具链。不是让模型变得更聪明，而是给它提供更好的脚手架。
Anthropic 的思路值得借鉴。无论你用的是 Claude、GPT 还是别的模型，在设计多轮长任务时，都要想清楚，怎么让下一轮的 Agent 快速进入状态，怎么避免它重复造轮子或者把代码搞成一团乱麻。即使是单轮任务，也要清楚它是没有记忆的，你需要通过外部文件来帮助它“想起来”之前做过的事。
以现在模型的能力，Coding Agent 已经能做很多事情了，核心还是在于你是不是能像软件工程中那样，去分解好任务，设计好工作的流程。
原文：Effective harnesses for long-running agents anthropic.com/engineering/ef…
翻译：baoyu.io/translations/e…
Anthropic 发布新博文，探讨长时间运行 AI 代理在多上下文窗口下的挑战，并提出受人类工程师启发的解决方案。 anthropic.com/engineering/ef…
How OpenAI Builds for 800 Million Weekly Active Users: Model Specialization and Fine-Tuning
We sat down with Sherwin Wu, Head of Engineering at OpenAI Platform, to discuss OpenAI’s developer strategy, how to manage top ML teams, why they decided to start releasing open-weight models again, how prompt engineering has evolved over time, and how developer tools will shift to accommodate better agents moving forward.
00:00 Introduction
8:36 Horizontal vs vertical OpenAI
12:18 Why you can’t “disintermediate” the model
15:11 People build relationships with models
17:30 Not one AGI model but many
20:10 Fine-tuning, RFT, and customer data choices
24:44 Prompt engineering isn’t the point anymore
28:06 What an “agent” really is
31:55 How OpenAI thinks about pricing
36:46 Why open-weights don’t kill the API
42:57 Different stacks for text, images, video
45:47 How the agent builder actually works
@sherwinwu @martin_casado
We wrote a tutorial on extracting massive structured tables from documents 📃
Using naive LLM structured output for document extraction fails if the number of output tokens is large - the LLM will end up dropping or hallucinating results.
A lot of documents are basically rendering massive tables in PDF form, like this Blue Shield document showing network coverage across 380+ CA hospitals.
We created a new mode in LlamaExtract that lets you extract every single row from this document with 100% accuracy. This lets you:
* ETL it into a structured database
* Do structured queries over it.
* And more!
Check out our blog post:llamaindex.ai/blog/extractin…e
Stop losing 80% of your data when extracting from long documents with repeating entities like catalogs, tables, and lists.
Our new Table Row extraction target in LlamaExtract solves the core problem: instead of trying to extract everything at once (where LLMs get overwhelmed), we intelligently segment documents and extract entity by entity.
🎯 Long multi-page insurance directory? Extract all 380 hospitals (vs. only 40 with document-level extraction)
📋 Handle both formal tables and semi-structured content like product catalogs automatically
🔍 Define your schema for a single entity - we return the complete list with exhaustive coverage
⚡ Leverage LLM flexibility while achieving template-based reliability through smart segmentation
This approach identifies repeating patterns, segments documents at natural boundaries, then applies your schema to focused chunks. Works with tables, lists, catalogs, or any document with distinguishable repeating entities.
Read the full technical breakdown with code examples herllamaindex.ai/blog/extractin…Z7N
We've been close partners for months and we can’t wait to build with @jarredsumner and the Bun team!
Read more here: anthropic.com/news/anthropic…
𝐑𝐞𝐝𝐝𝐢𝐭 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 just shared a deep dive on their Vector Database selection journey.
Thrilled and honored to see 𝐌𝐢𝐥𝐯𝐮𝐬 stand out for reliability and scale. 🚀
They narrowed the field from many options (Vertex AI, Elasticsearch, etc.) to a final showdown: Qdrant vs. Milvus, based on strict criteria: Open Source, Proven 1B+ Scale, and High Performance.
🧪Reddit stressed-tested both systems, focusing on two critical operational areas:
1. 𝐒𝐜𝐚𝐥𝐢𝐧𝐠: While Qdrant required manual sharding for replication changes, Milvus excelled with 𝐚𝐮𝐭𝐨𝐦𝐚𝐭𝐢𝐜 𝐝𝐚𝐭𝐚 𝐫𝐞𝐛𝐚𝐥𝐚𝐧𝐜𝐢𝐧𝐠, significantly reducing operational toil.
2. 𝐃𝐞𝐛𝐮𝐠𝐠𝐚𝐛𝐢𝐥𝐢𝐭𝐲: Reddit found that Milvus's distributed architecture provided superior observability and control, making it easier to isolate and fix issues under stress.
This is precisely why Reddit chose 𝐌𝐢𝐥𝐯𝐮𝐬🏆.
💡There is no "standard answer" for selection, but here are some lessons learned:
𝐃𝐢𝐠 𝐃𝐞𝐞𝐩𝐞𝐫: Challenge requirements and avoid bias—don't get trapped by existing solutions.
𝐒𝐜𝐨𝐫𝐞𝐬 ≠ 𝐃𝐞𝐜𝐢𝐬𝐢𝐨𝐧𝐬: Use matrices to clarify needs, but don't get fooled by flashy documentation.
𝐁𝐚𝐥𝐚𝐧𝐜𝐢𝐧𝐠 𝐀𝐫𝐭: Don't obsess ovemilvus.io/blog/choosing-… maintenance, debugging, and usability over niche performance metrics.
Know the details: https://t.co/qaffu0rG2H
———
👉 Follow @Milvus, for everything related to unstructured data!
I talked to a startup today that had been forced to take so much from investors and whose expenses were so low that they could put the money in bonds and fund their operations off the interest. I told them they should.
Now that you’ve had the chance to get to know GPT-5.1, we pull back the curtain on how training took shape.
On this episode of the OpenAI Podcast, @christinahkim and @Laurentia___ join @andrewmayne to talk about reasoning in GPT-5.1 Instant, personality controls, and how they refine model behavior at scale.
What world-class GTM looks like in 2026 with Jeanne DeWitt Grosser (@jdewitt29)
Jeanne built @stripe's legendary sales team from the ground up, and is currently COO at @Vercel, where she oversees marketing, sales, customer Success, and rev ops.
In our conversation, we discuss:
🔸 Why GTM is becoming a huge differentiator in the AI era
🔸 How AI is changing sales
🔸 The rise of the "GTM engineer" role
🔸 How Vercel made their SDRs 10x more efficient
🔸 A primer on segmentation
🔸 How to build a sales org that product teams love
🔸 Much more
Listen now 👇
• Yyoutu.be/RmnWHz8HD74yB79mREN
• Sopen.spotify.com/episode/33oMOu…LRrS4IDU
•podcasts.apple.com/us/podcast/the…4bgvKLPh
Thank you to our wonderful sponsors for supporting the podcast:
🏆 @datadoghq — Now home to Eppo, the leading experimentation and feature flagging pdatadoghq.com/lennysR16CMQo1
🏆 @Lovable — Build apps by simply chattinglovable.devBR0CyiVN59
🏆 @stripe — Helping companies of all sizes grostripe.com/zOscBW98BX
Great article from @jeremyakahn in Fortune on the huge impact of AlphaFold on biological and biomedical research.
Five years on, Google DeepMind’s AlphaFold shows why science may be AI’s killer app fortune.com/2025/11/28/goo…
