Yesterday, open A, I took a page out of the anthropic c playbook and announced, every journalist on earth is that the next AI model is super deeper, scary and dangerous. 昨天，OpenAI 我借鉴了 Anthropic 的剧本，宣布——地球上每个记者都在说——下一个 AI 模型超级深、可怕又危险。
And because Sam open cares about your safety more than anything, he decided to hit the pause button on frontier reinforcement learning for the next two weeks. 而且因为 Sam Altman 比什么都更关心你的安全，他决定在未来两周按下前沿强化学习的暂停键。
And the controversial, yet so brave. 这一举动颇具争议，但又如此勇敢。
And the official reason is that open a eyes next model code named Astra, may have cross the critical cybercafes threshold, and maybe that not too far fetch, considering one of model recently escaped in evaluation sandbox because because it could have hugging face Eesti to cheat on the benchmark. 官方理由是 OpenAI 的下一代模型，代号 Astra，可能已经跨越了关键的网络能力阈值，而这也许并不算太牵强，考虑到其中一个模型最近在评估沙箱中逃逸了，因为它可能利用 Hugging Face 数据集在基准测试中作弊。
But not everybody is by the story and because nobody would have a brain actually believes they’re stop in the largest plan training run in history for a quote, safety. 但并不是所有人都买账，因为没有一个有脑子的人会真的相信他们会为了所谓的“安全”而停止历史上最大规模的训练运行。
There’s an an imaginable amount of power and wealth stake in the AI race right now. 当前 AI 竞赛中涉及的权力和财富难以想象。
And some armchair experts say this is proof that AI is Platov, while others say it’s all about regulatory capture. 一些纸上谈兵的专家说这证明 AI 已经达到瓶颈，而另一些人则说这完全是监管俘获。
But perhaps the most likely theory is what happened in 6000 miles away in China, because coincidently, Deepa is returned and release the fastest start GitHub bo in history. 但也许最有可能的理论与 6000 英里外的中国发生的事情有关，因为巧合的是，DeepSeek 回归并发布了史上最快获得 GitHub star 的项目。
And today’s video will take a look at deep sea hardness and find out if it lives up to the hype or if it’s just a cheap team on knock off of cloud code. 今天的视频将看看 DeepSeek harness，看看它是否名副其实，或者只是 Claude Code 的廉价山寨版。
It is August 28, 2026. And and or watching the code report, if you know your AI history, you might remember that the entire type scrip code base of cloud code and lick to the net t anthropy c. 今天是 2026 年 8 月 28 日，你正在收看 The Code Report。如果你了解 AI 历史，你可能还记得 Claude Code 的整个 TypeScript 代码库被 Anthropic 泄露到了网上。
Accidentally shift AA, A eab source map file to P, P, M. 他们不小心把一个 source map 文件发布到了 npm。
And Thropp is famously smug about the anti open source and anti open wait. 而 Anthropic 以对反开源和反开放权重态度自鸣得意而闻名。
So so this leak was like the invisible hand of God finally clicking the popes. 所以这次泄露就像上帝的无形之手终于按下了按钮。
And we analyze cloud code source code on this channel. It is pretty mod. 我们在这个频道分析过 Claude Code 的源代码，它相当一般。
It is surprising to see Chinese knock off harnesses pop up and wait a minute, you be wondering what it tia horse, well, to get AI to write your code, you first need a model to spit out and predict kens that the brain of the operation, and to do everything else you need to strap on a horse. 看到中国山寨 harness 冒出来很令人惊讶，等等，你可能想知道 harness 是什么。好吧，要让 AI 写代码，你首先需要一个模型来输出和预测 token，这是操作的大脑；而要做其他所有事情，你需要套上一个 harness。
And that allows the the in in to the tools plugin, the file system, manage context and so on. 它允许接入工具插件、文件系统、管理上下文等等。
And the hornets runs on this things in a loop that decides when to keep going and when to stop. 而 harness 在一个循环中运行这些东西，决定何时继续、何时停止。
Is some popular harnesses or opening ec codex, cloud code, open code and many others. 一些流行的 harness 有 OpenAI Codex、Claude Code、OpenCode 等等。
But the deep sea corners takes a much different architectural approach that can summarized in the words, that is, a pluggin. 但 DeepSeek harness 采用了一种截然不同的架构方法，可以用一句话概括：一切都是插件。
The model adapter is a plug in the tools or plugins and the sandbox in the U. Y. And the wild. 模型适配器是插件，工具是插件，沙箱在 UI 和外部环境中也是插件。
Look at the center of the code and in self pluggin or in other words, they’re just constant pages and the spot on with a mind. 看看代码的核心，它本身就是一个插件，换句话说，它们只是常量页面，并且恰到好处。
Hmm. And and really cool of your developer because it means you have a ton of control to customized your plugin and that are not possible on other platforms. 嗯，这对开发者来说真的很酷，因为这意味着你有大量控制权来定制你的插件，这在其他平台上是不可能的。
Like instead of using the sandbox and Thropp told you a safe, you have the freedom to use custom sandbox from some random half and and maintain geography. 比如，与其使用 Anthropic 告诉你的安全沙箱，你还可以自由地使用某个随机黑客的自定义沙箱，并自行维护安全性。
The whole thing feels like Linux. But for AI agents at a deeper level, though, the tools built TA, A app that was released by deep sea about spatial temporal composition. 整个东西感觉就像 Linux，但针对 AI 代理。在更深层次上，DeepSeek 发布的应用中内置了关于时空组合的工具。
But the general idea is the Deonna should be hot, swapped the, the, the, the, the baby in the honest and over time. 但总体思路是模型应该可以热插拔，就像 harness 中的婴儿一样，随着时间推移可以更换。
And they actually built a small framework, Curtis, to facilitate this plugin architecture. 他们实际上构建了一个小框架 Curtis 来促进这种插件架构。
Now, I ve used many different terrible plugin systems in my days, like a word, Press and website. 我这些年用过很多糟糕的插件系统，比如 WordPress 和网站。
But this paper from deep sea is easily the most Laboratoire and have ever seen for a deluging model. 但 DeepSeek 的这篇论文无疑是我见过的语言模型中最精巧的。
But the big question is, can the deep sea hardness right better code than codex or code code? 但最大的问题是，DeepSeek harness 能比 Codex 或 Claude Code 写出更好的代码吗？
But at the same time of the release of the desk model, they also released version for for pro of the flagship model, along with a massive price increase in the API. 但在发布 DeepSeek 模型的同时，他们还发布了旗舰模型的 v4 Pro 版本，同时 API 价格大幅上涨。
I have been to have a demo city on my deep sec account, so let’s find out if you can build a production version of horse ender, but this one shot prompt and deep sea model with the mac setons. 我碰巧在我的 DeepSeek 账户上有一些演示额度，所以让我们看看能否用这个一次性提示词和 DeepSeek 模型以最大设置构建一个生产版本的 horse render。
But one thing to keep in mind is that you can point any different model to this harness. It’s not just limited to deep six model. 但要记住的一点是，你可以将任何不同的模型指向这个 harness，它不仅仅限于 DeepSeek 模型。
The first, the resting thing you’ll notice is that there is different modes for running this pro. 你会注意到的第一件有趣的事情是，运行这个程序有不同的模式。
I’m just going to go a standard mode, but there is also a minimal mode that will speed things up and a creator mode if you actually want to dive into the plugins and create your own mode. 我打算使用标准模式，但还有一种可以加速的最小模式，以及一种创造者模式，如果你真的想深入研究插件并创建自己的模式。
After submitting the prompt and immediate a model, as is doing its work, we can go to the project c pannel. 提交提示词后，模型立即开始工作，我们可以进入项目面板。
You can see all the reasoning tool calls and results in a way that feels like a stack trace for your AI models thinking process. 你可以看到所有推理工具调用和结果，感觉就像 AI 模型思考过程的堆栈跟踪。
But pretty cool. And after staring at the, after about 29 minutes and 58s, I finally had a working application. 相当酷。在盯着屏幕大约 29 分 58 秒后，我终于得到了一个可运行的应用程序。
But it AA, A, A million output token. S it. I used a grand total of $0.30. 但它消耗了一百万个输出 token。是的，我总共花了 0.30 美元。
It built the entire thing with no js and react. 它用 Node.js 和 React 构建了整个应用。
And at first, I was a little bit disappointed with AU. I. You’ll definitely get more spectacular results with favor or codex. 起初，我对 UI 有点失望。用 Vercel 或 Codex 你肯定会得到更惊艳的结果。
But this deep sic application is still pretty solid, but the wipe animation was implemented well. 但这个 DeepSeek 应用仍然相当扎实，擦除动画实现得很好。
It has a nice chat feature and got the job done well on a bunch of other small details. 它有一个不错的聊天功能，并且在许多其他小细节上也完成得很好。
So Yeah, if I were Sammer Dara right now, I would be totally terrified and just give up on the whole AI d ts. 所以，是的，如果我现在是 Sam Altman，我会非常害怕，干脆放弃整个 AI 事业。
But I want to actually understand AI issues at a deep level, but need to check out blue diet impact, the sponsor of today’s video, that there a non profit whose mission is to get more people involved in making AI go better for humanity. 但如果你想真正深入理解 AI 问题，你需要看看今天的视频赞助商 BlueDot Impact，他们是一家非营利组织，其使命是让更多人参与让 AI 更好地造福人类。
The main way they do that is by offering free on my courses, like the future of AI course, but which provides an unbiased introduction to where AI is today and things could be heading over the next few years. 他们做到这一点的主要方式是提供免费在线课程，比如“AI 的未来”课程，该课程公正地介绍 AI 的现状以及未来几年可能的发展方向。
And they also have more technical courses on things like AI governance and biosecurity belong with personalize career support for people interested in working on AI safety. 他们还有关于 AI 治理和生物安全等更技术性的课程，以及为有兴趣从事 AI 安全工作的人提供个性化职业支持。
All of blue dots resources are free to use because they’re funded by Phillip project donations, and they’ve helped over two people get jobs. BlueDot 的所有资源都免费使用，因为它们由慈善捐款资助，并且已经帮助超过两千人找到工作。
And organizations like DeepMind, the Stanford HA and and Apollo research try out the future of AI course right now at the link below. 像 DeepMind、斯坦福 HAI 和 Apollo Research 等组织的人员都可以通过下方链接立即试用“AI 的未来”课程。
And this been the code report. Thanks for watching, and I will see you in the next one. 以上就是本期 The Code Report。感谢观看，我们下期再见。