春秋带剑 | 公众号 chunqiudaijian
1. 技术思考分享
- 开源AI视频复现工具：Hypit 官方使用指南
2. 新闻
1. 科大讯飞发布讯飞星火X2.6大模型：全国产算力训练，端云一体架构
“只有真正地在国产算力上训练出自己的模型，才是真的自主可控。”10月22日，在2026科大讯飞全球1024开发者节主论坛上，科大讯飞董事长刘庆峰强调，全栈自主可控不仅要能够在国产算力上进行推理，更要能够在国产平台上完成训练。
梁文峰：
参考资料：科大讯飞发布新一代国产大模型，刘庆峰最新发声|中国基金报|讯飞星火|自主可控|推理|AI_新浪新闻
2. 微软发布多智能体harness Agensh
去掉了现在流行架构中负责统筹、规划安排任务的主智能体（orchestartor），让AI去中心化、自组织的协作。
论文：Agensh: Scaling Organizational Intelligence to 1,024 Agents
3. 有个X英文博主今天一口气狂发4条帖子爆Kimi AI虚假宣传
博主直接把PR发的宣传稿po出来了：
博主拒绝参与宣传涉及中美矛盾的内容：
博主拒绝宣传并且在别的宣传帖底下吐槽，被要求删评：
宣传策略是贬低Claude：
吃瓜！
4. 英伟达推出AI工厂运维平台Minerva
参考资料：英伟达推出AI工厂运维平台Minerva：每小时分析数十亿个数据点_新浪科技_新浪网
5. 国内首部拿到院线公映许可的人工智能（AI）电影《三星堆：未来往事》于23日登陆全国影院
参考资料：中国首部AI院线电影主创接受《环球时报》专访：AI电影不是速成片，观众期待一个“值得看的故事”_新浪财经_新浪网
6. OpenAI CEO奥尔特曼对数学界的演讲
OpenAI CEO奥尔特曼近日在斯坦福大学面向担忧职业前景的学生演讲。他描述数学家面对AI进军数学研究时的心理变化：第一天"我们完蛋了"，第三天"OpenAI太过分了"，第五天开始接受AI成为数学研究不可或缺的工具。
AI给出证明大量且迅速，数学界人工消化能力跟不上；数学学术领域对学者的评判标准将会发生巨变，甚至直接失去可靠的发展路径；AI巨头目前使用的AI模型仍然私有，数学家而且可能侵占了现有学者的知识产权，这是一场不公平的竞争。
AI数学发生了很多事，还将继续发生很多事。人们从不同的视角来看这一切。
AI数学领域对路人来说还是偏神仙打架了一点，对于我个人来说，数学，学点数学挺好的，锻炼脑子，减少得阿尔兹海默症的概率
参考资料：OpenAI 奥尔特曼：面对 AI，数学家们起初会觉得“我们完蛋了”，最终却还是接受现实_新浪科技_新浪网
7. OpenAI的GPT-Rosalind仅用30分钟分析数百万个小鼠脑细胞的已有实验数据，为罕见神经退行性疾病佩里综合征找到潜在治疗线索。人类完成同类数据分析与文献审查估计需要数月甚至一年
继AI数学之后，AI生信领域也传来噩耗！
看了一下，好像AI主要是做的文献综述工作，整理可能的研究方向，还没做实验。但是AI做这么大量的文献整理确实也已经超越了人类的能力，还是很有发展潜力的。不过生信是实验科学，学者们的主要工作——发现有价值的研究方向、做实验、确认应该还没有被颠覆（吧）
参考资料：生信人瘫坐🤯AI仅32分钟分析百万小鼠脑细胞 - 小红书
8. 字节跳动大模型投毒事件，受害者也出了调查报告
第2期的时候我写过田柯宇本人做客elsewhere播客访谈，讲述自己是年少轻狂、见义勇为过激行事。今天我才发现原来他的同事也出了长文，来反驳他：2024 年模型训练攻击事件：受害者陈述与证据整理
包括两点：第一是持续在破坏，而且他自己就在调查会里，还根据调查会的进展修改破坏方案。一开始被攻击者用的算力很少，后来被破坏的Goku项目才有占用GPU，这是个多人协作项目。所以他声称自己是因为看别人的项目占了别人的GPU才见义勇为，大概率是巧言令色。
第二是他部署用于内部测试的自己的项目VAR text-to-image实际上就是开源模型PixArt，自己的模型其实并没有达到声称效果。
珍贵的双视角罗生门。最后还留下了受害者们的实名联系方式。感觉这篇长文的作者在尽可能表示克制了，但是对田柯宇导致的莫名其妙的bug仍然透露出了无法遏制的恨意。
当然这也是程序员之常情！
看了一下GitHub的issue区，意思是之所以现在还是两头打嘴仗，而没有直接提起刑事诉讼，是因为是TikTok，字节跳动不敢打官司……
不过无论如何，投毒事件是真的，反正如果我是投资者的话，是不敢投田柯宇的创业项目的。年少轻狂做错了事的聪明人，如今真的长大了吗？不过也许投资者有他们自己的道理。毕竟创业投资这行并不太讲究道德叙事。至于田柯宇的创业项目的产品，我不确定，看看他们能做出什么吧。得验证一下真的是他们自己做的产品才行。
田柯宇在播客中的发言和受害人的报告更详细的矛盾对比可以参考：深度丨双面田柯宇：一次无法自圆其说的「行侠仗义」
还有一篇莽AI的非常乐子的锐评：基座模型田柯宇从不缺卡
最后我们留一个思考题，田柯宇到底能不能算字节系创业者？
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
WeChat Official Account cunqiudaijian
1. Technology thinking sharing
- Open-source AI video reproduction tool: Hypit official usage guide
Public information
1. Koda Z66ZpublishZ66Z Starfire X2.6LLM: National computing exercise, cloud-like architecture
“It is only by truly training its own model in national production techniques that it is truly autonomous.” On 22 October, at the 2026 Grand iFlytek Global Developer's Forum, President Liu Qingfeng of Big Z66Z emphasized that full-fledged autonomy could be controlled not only by reason of national productivity but also by training on national production platforms.
Liang Wen Feng:
References: iFlytekZ94Z, the new generation of national production LLM, Liu Qingfung's latest voice, the China Foundation, iFlytek, autonomous and controllable reasoning, Ai New Wave News
2. Microsoft publish Multi AI agentharness Agency
AI agent (orchestartor), the main body in the current popular architecture responsible for integration and planning, has been eliminated, leaving AI decentralised and self-organised.
Thesis: Agesh: Scaling Organic Inteligence to 1,024 Actes
3. An X-English blogger who's been blowing up Kimi AI's false propaganda with four posts today.
The blogger has published a press release from PR:
The blogger refused to engage in propaganda about the Central American conflict:
The blogger, who refused to promote it and gave it to others, was asked to delete it:
The propaganda strategy is to devalue Claude:
Eat melons!
4. Minerva, the launch of the AI Plant Transport Platform
References: Minerva: analysis of billions of data points per hour — New Wave — New Wave Network
5. The country ' s first film, " The Three Stars: The Future ", which received a public screening license, landed at the National Cinema on 23rd.
Reference: China's first AE-line filmmaker received an interview with The Globe Times: the AI film is not an instant film, and the audience looks forward to a “story worth watching” New Wave New Wave New Wave
6. Open AI CEO Ortman ' s speech to the mathematics community
Open AI CEO Ortmann recently addressed students at Stanford University who were worried about career prospects. He described the psychological changes that mathematicians had in the face of AI: first day, "We're done," third day, "Open AI's gone too far," and the fifth day began to accept AI as an indispensable tool for mathematical research.
AI has proven to be large and fast, and the mathematics community is unable to adapt to artificial digestion; the mathematical academic world will experience a major change in the standards of appreciation of scholars, or even a direct loss of a reliable path of development; the AI model currently used by AI giants remains private, mathematicians may have encroached on the intellectual property rights of existing scholars, which is an unfair competition.
There's a lot going on in Ai maths, and there's still a lot going on. People look at it from different perspectives.
It's a little bit of an A.M.T. fight for the passers-by.
References: Open AI Altmann: In the face of AI, mathematicians would initially think, "We're finished," but eventually accept reality.
7. The GPT-Rosalind of Open AI uses only 30 minutes to analyse available experimental data on the brain cells of millions of mice and to find potential therapeutic clues for the Peri syndrome of rare neurodepressive diseases. Humans are expected to complete the same data analysis and literature review for months or even a year.
After alms, there's bad news coming from the AI field!
It's like AI's mainly doing a literature review, sorting out possible research directions, and not experimenting. But it is true that AI has gone beyond human capabilities and has great potential for development. But biopsy is experimental science, and the major work of scholars -- finding valuable research directions, experimenting, confirming that it hasn't been subverted.
Reference: The biopsy is only 32 minutes to analyze millions of rat brain cells - REDnote
8. byte beat LLM poisoning incident and victim investigation report
On the second issue, I wrote about Tane's own interview as a podcaster for elsewhere, describing myself as a young man who behaved too harshly. I found out today that his former colleague had written a long article to counter him: Model training attack of 2024: victim statement and evidence.
There are two points: first, the destruction continues, and he himself is at the investigative conference, and the destruction programme is revised in the light of the progress of the investigation. It was only after the destruction of the Goku project that the GPU was occupied, a multi-person collaborative project, that the attackers initially used little credit. That is why he claimed that he was acting in good faith because he saw other people's GPUs taking over.
The second is his own project, VAR text-to-image, deployed for internal testing, which is actually an open source model, PixArt, whose own model does not actually produce the alleged results.
Precious double perspective. The victims ' real contact details were left. It was felt that the author of the long article was showing the greatest possible restraint, but the inexplicable bugs caused by Tan Koyu still revealed an unstoppable hatred.
Of course it's the programmer's routine!
Look at GitHub's assue district, which means that now there's a fight between two sides and no criminal proceedings are brought directly because TikTok, byte, is afraid to take a case...
In any case, however, the incident of poisoning is true, and if I were an investor, I would not be able to go to Ko Woo's business project. Are you really grown up now? But perhaps investors have their own reasons. After all, the business of investing in entrepreneurship is not very ethical. I'm not sure what they're doing with Tan Ko-woo's business project. We have to check that they actually made it.
A more detailed comparison of the statements made by Tan Ko-woo among the podcasts and the victim's report can be found: Ko-woo, in the depths of the two-faced field: an inability to justify what he said about "soldierness."
There's also a very happy comment from the rash AI: Ko-woo has never missed a card.
And finally, we'll leave a question of whether Tane Ko Woo is an entrepreneur.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.