Yinglai Yang

Posting about AI, Coding and Cars.

[Yinglai Yang @yangminded]

@mattparlmer Coal plant make bill go down. Datacenter make bill go up. Simple as that.

[Yinglai Yang @yangminded]

@PKUCXK 眼来了！这会有paper么？

[Yinglai Yang @yangminded]

@htihle @teortaxesTex Interesting. That would fit to the recent discussion of scaling by @jietang here.

[Yinglai Yang @yangminded]

@tolstoybb Chinese would literally riot if there was such a high error rate.

[Yinglai Yang @yangminded]

@MaxForAI 抗议之前也不想一想，把rogue agent表演的成萌宝宝。

[Yinglai Yang @yangminded]

@zephyr_z9 Does not bet his whole life savings in leveraged memory ETFs.

Cannot eat Sichuanese food from Hongkong.

Jukan might be a European fellow 😄

[Yinglai Yang @yangminded]

@niccruzpatane Amazing, hardcore, gung-ho! Shows that at least the Chinese automakers still have balls.

[Yinglai Yang @yangminded]

Time for anyone who works with Anthropic models to self-train a small OCR model with completely private data. Only way to “air-gap”.

[terminally onλine εngineer @tekbog]

soon you will have anthropic strike your content with a copyright for wrongful thinking and misuse of their models

not to add to all the business documents that will now have an AI watermark anyone can check, even in codebases - does this change ownership?

[Yinglai Yang @yangminded]

@DeepDishEnjoyer I - education, Cape Cod, Maine, Vermont

[Yinglai Yang @yangminded]

@paularambles Bugged me throughout the whole novel including other inconsequential names. Maybe Butler was slightly dyslexic.

[Yinglai Yang @yangminded]

@MaxForAI 现在跟不上这些年轻人了

[Yinglai Yang @yangminded]

@teortaxesTex AGI!

[Yinglai Yang @yangminded]

@teortaxesTex Yeah, that’s a case of “It’s just not for you.”

People need to take Liang Wenfeng verbatim. DeepSeek wants to push technological frontiers and experiment. Consumer UX is just not in their interest at all (very notable).

If you want a great daily driver, just use Kimi.

[Max For AI @MaxForAI]

推荐大家读一下DeepSeek Harness开源同时发布的这个他们北大合作的Agent新论文🔥

这次他们研究的不是模型本身，而是一个可能直接关系到下一代 Agent 的问题：

当 Agent 开始持续修改自己的 Harness，怎么保证它不会把自己改死？

论文名叫《A Programming Paradigm for Spatiotemporal Composability》，作者来自北京大学和 @deepseek_ai ，并给出了一套围绕 Cordis 构建的动态组件理论。

论文提出两个核心概念。

第一是 Temporal Composability。

一个组件加载后产生的资源注册、状态修改和其他副作用，都需要带有对应的 inverse。组件被移除时，Runtime 可以自动追踪并撤销这些影响。

目标不是简单执行一个开发者手写的 cleanup，而是让组件的副作用从机制上具备可恢复性。

第二是 Spatial Composability。

组件可以声明自己依赖哪些其他组件。当 Provider 出现、消失或被替换时，系统自动重新计算依赖关系，只重新激活真正受影响的组件。

论文进一步给出了完整的 component lifecycle、dependency ordering、failure recovery 和 transactional HMR，并对这套动态组合机制进行了形式化证明。

其中一个尤其值得注意的结论是 Confluence。

在论文规定的独立性、无环依赖等条件成立时，一个系统即使经历过多次组件加载、卸载和替换，最终达到稳定状态后，可以与「直接从最终配置重新组装一次」得到的状态等价。

换句话说，中间发生过的动态演化，在最终系统状态中可以不留下痕迹。

这对于未来 Self-Evolving Agent 很关键。

论文专门把 Self-Evolving Agent Harnesses 列为核心应用场景：未来 Agent 可能会在持续处理任务的同时，生成、部署和替换自己的 Tool、Memory、Sandbox、Orchestration 等 Harness 组件。

如果一次错误修改必须重启整个进程，自我进化很难真正持续进行；更严重的是，一次错误修改甚至可能破坏负责恢复系统本身的进程。

Cordis 想解决的正是这个基础设施问题：

让 Agent 可以修改自己，同时让修改本身可撤销、依赖可重连、故障可以局部隔离。

论文还用拥有 4000+ 社区插件的 Koishi 作为实际案例，展示了组件动态卸载、依赖重解析和 HMR 等机制。不过作者也明确说明，目前 Koishi 使用的是 Cordis v3，而论文描述的是进一步重新设计后的 Cordis v4。

所以现在还不能直接得出「DeepSeek Harness 已经基于 Cordis」这样的结论。

但这篇论文透露出的研究方向已经足够明确：

当 Agent 开始拥有长期运行、自我修改和自我演化能力之后，竞争可能不再只是模型能不能写出新的代码，而是谁能先解决一个更基础的问题：让 AI 安全地修改自己。

祝贺 @tianyi

[Tianyi Cui @tianyi]

DeepSeek Harness was just released with MIT license. The current 0.1.0 version is a developer preview, and may still have many rough edges. Feedback is welcome!

DeepSeek Harness 已经以MIT协议开源发布。现在的0.1.0版本是一个面向Harness开发者的预览版，还很不完善。恳请大家多提提宝贵意见。

https://t.co/aBToa3b3L9 https://t.co/VMZC5i2sG0

[Yinglai Yang @yangminded]

Nice. The Cordis framework looks especially interesting.

The main idea is how an agent can modify its own harness multiple times without damage. This also enables very long-running agents.

[DeepSeek @deepseek_ai]

🧩 DeepSeek Harness v0.1 is now available in Developer Preview!

🔹 We’re opening it up to developers building agent harnesses worldwide and open-sourcing the codebase in MIT license.
🔹 Powered by the Cordis meta-framework, DeepSeek Harness is an agent harness built around one core idea: Everything is a plugin. Models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, and UI are ALL implemented as plugins, and can be mixed, matched, replaced, and extended.

Try it now!
https://t.co/2YWSvJHhKA

[Yinglai Yang @yangminded]

What’s the point of traveling if not to mold your mind with new impressions from what you see and experience on the ground?

Content for social media? Guess what - it’s 2026 and you can easily generate short videos of you rising horses in Mongolia, ice-skating in Harbin. Heck, even parachuting down 东方明珠 in Shanghai as if you’re James Bond.

Or eating local food? Just take your phone and open 美团/淘宝闪购. Even Americans will have great options to travel culinarily with Doordash.

More people than ever have access to travel as leisure, and more than ever do it. But more than ever we can ask ourselves: What for?

[Yinglai Yang @yangminded]

My god, people in the comment section would do well to do the same as Joe and travel China without those stereotypes in mind.

Only then will you find interesting impulses from different places and people.

Otherwise you will just reinforce your stereotypes. Then you could have just staid at home and be no wiser.

[Robert @Baiguan @robert_baiguan]

They call Deng Xiaoping the great reform and opening-up of the Chinese economy's "chief architect", but Zhu is the actual engineer who made it work

[Yinglai Yang @yangminded]

Spotify is easily 10 times as technologically advanced and forward looking as any other company based in continental Europe.

[Spotify Engineering @SpotifyEng]

We just launched Xirp, a vendor-neutral agentic development environment. One place to manage agent sessions across @ClaudeDevs, @GeminiApp CLI, and @OpenAI Codex. 1,300+ @Spotify engineers already use it. Now it's available for you to try. Learn more at https://t.co/Hwo8Qqx4OI.