DeepSeek Harness Review — Full Breakdown & Transcript

[00:02] anthropic playbook and announced to every journalist on Earth that their next AI model is super duper scary and dangerous. And because Sam Alman cares decided to hit the pause button on Frontier Reinforcement Learning for the

[00:15] next 2 weeks. That's so controversial yet so brave. And the official reason is that OpenAI's next model, Cenamed Astra, may have crossed the critical cyber capability threshold. And maybe that's not too far-fetched considering one of

[00:28] their models recently escaped an evaluation sandbox so it could hack cheat on a benchmark. But not everybody is buying this story because nobody with half a brain actually believes they're stopping the largest planned training

[00:41] run in history for quote safety. That there's an unimaginable amount of power and wealth at stake in the AI race right now. And some armchair experts say this others say it's all about regulatory capture. But perhaps the most likely

[00:54] theory is what's happening 6,000 mi away in China. Because coincidentally, Deepseek just returned and released the fastest starred GitHub repo in history. In today's video, we'll take a look at Deep Seek Harness and find out if it

[01:07] lives up to the hype or if it's just a cheap T-Mu knockoff of Claude Code. It watching the code report. If you know your AI history, you might remember that the entire TypeScript codebase of Claude Code was leaked to the internet after

[01:20] Anthropic accidentally shipped a 57 megabyte source map file to npm. Anthropic is famously smug about being anti-open source and anti-open weight. So, this leak was like the invisible hand of God, finally clicking public

[01:33] source code on this channel and it was pretty mid. So, it's not surprising to see Chinese knockoff harnesses popping up now. But wait a minute, you might be wondering what even is a harness. Well, to get AI to write your code, you first

[01:46] tokens. That's the brain of the you need to strap on a harness that allows the brain to use tools, plugins, the file system, manage context, and so on. And the harness runs all these

[02:00] keep going and when to stop. Some popular harnesses are OpenAI codeex, Claude Code, Open Code, and many others. But the Deep Sea Carnis takes a much different architectural approach that can be summarized in three words.

[02:13] Everything is a plugin. The model adapter is a plugin. The tools are plugins. Even the sandbox and the UI and the while loop at the center of the other words, they're just ordinary packages that you can swap out with one

[02:26] line of YAML. And that's really cool if you're a developer because it means you harness in ways that are not possible on other platforms. Like instead of using the sandbox Anthropic told you was safe, you have the freedom to use a custom

[02:39] sandbox from some random half-assed unmaintained GitHub repo. The whole thing feels like Linux but for AI agents. At a deeper level though, this released by Deep Seek about spatiotemporal composability. The

[02:52] general idea is that components should be hot swappable both as dependencies in the harness and over time. And they actually built a small framework called architecture. Now, I've used many different terrible plug-in systems in my

[03:05] days, like in WordPress and Webpack, but this paper from DeepSeek is easily the most elaborate justification I've ever seen for a plug-in system. But the big question is, can the Deep Seek Harness write better code than Codeex or Claude

[03:17] release of the Deep Seek Harness, they also released version 4 Pro of their flagship model along with a massive price increase in the API. I happen to have $20 sitting on my DeepSeek account. So, let's find out if it can build a

[03:30] production version of Horse Tinder. For this OneShot prompt, I'm using V4 Pro with the max settings. But one thing to keep in mind is that you can point any different model to this harness. It's not just limited to DeepSeeks models.

[03:42] notice is that there's different modes for running this prompt. I'm just going also a minimal mode that will speed things up and a creator mode if you and create your own stuff. After submitting the prompt, it immediately

[03:55] gets to work. And as it's doing its work, we can go to this trajectory panel where you can see all the reasoning, tool calls, and results in a way that model's thinking process. But pretty cool. And after staring at that for

[04:08] about 29 minutes and 58 seconds, I finally had a working application built with 2.6 million output tokens, which cost a grand total of 30. It built the at first, I was a little bit disappointed with the UI. You'll

[04:21] definitely get more spectacular results with Fable or Codeex, but this Deepseek application is still pretty solid. The swipe animation was implemented well. It done well on a bunch of other small details. So yeah, if I were Sam or

[04:34] Daario right now, I would be totally terrified and just give up on this whole understand AI issues at a deep level, sponsor of today's video. They're a

[04:46] people involved in making AI go better for humanity. The main way they do that is by offering free online courses like their future of AI course which provides an unbiased introduction to where AI is today and where things could be heading

[05:00] over the next few years. They also have more technical courses on things like AI governance and biocurity along with personalized career support for people interested in working on AI safety. All of Blue Dot's resources are free to use

[05:13] because they're funded by philanthropic donations. And they've helped over 8,000 people get jobs at organizations like DeepMind, the Stanford, HAI, and Apollo Research. Try out their future of AI course right now at the link below. This

[05:26] watching, and I will see you in the next one.