The Chinese LLMs are good!
Source rabbitholes.gardenI’m addicted to vibe-coding, and the reason is simple: thanks to coding tools like Codex, Claude Code, and others, I sincerely think that I can build the next Amazon, Google, or Notion from my bedroom—and also sometimes from when I’m sitting on the potty, because I can still send remote commands from my phone.
Coding tools have truly democratized the ability for any random idiot to create a unicorn. In that sense, aspiring to be a billionaire used to feel nice, but now the thrill is gone. If I can easily vibe-code unicorns at whim, what’s the point?
But anyways, I still continue to vibe-code because I have a long list of things that I always wanted to do but couldn’t. Now, thanks to these tools, I can—as the motto of the moment puts it, “just do things.” What these tools have also unlocked is a golden age of fucking around and finding out, and a golden age of people doing random side projects.
A lot of people pejoratively dunk on all the vibe-coded apps and websites and call it slop. A lot of it is well deserved, but not all slop is slop. A lot of these things are small things that people always wanted to do for their own fun and enjoyment. A lot of them are very narrow and have a lot of utility—be it somebody trying to digitize a rare archive, somebody building a tool to explore local city or municipal finances, somebody building a tool to help people understand how the Indian government gets money and where it goes, or people building tools to visualize historical maps, and so on.
Calling all of these things slop feels utterly stupid. All the snobby pricks who just throw this term with reckless abandon deserve to be kicked in the nuts by a horse.
Anyways, my vibe-coding adventures continue. I was working on a way of using old historical maps as a gateway for people to explore history, and I wanted to experiment with using all these Chinese models because there’s so much hype on them. I’ve been using Kimi K3, GLM 5.2, DeepSeek, and Qwen.
Qwen, for some reason, wasn’t… I didn’t have an aha moment with it. But for the rest, I’m happy to report that the hype is justified.
All of this I say strictly as a vibe-coder. I think somebody who actually knows coding will probably be able to get 100x the productivity of these tools, but I speak strictly as a vibe-coder.
Ironically, somebody at work mentioned that using these large language models is like using a slot machine, and I think that’s a very apt description. While I was initially awed at the capabilities—or rather the quality—of Claude Code and Codex, in the last few months their quality has steadily degraded.
At the same time, using Kimi K3, DeepSeek, and GLM 5.2 in limited use cases was as good as using the earliest versions of Claude Code and Codex. They’re just really good. The back and forth required to get whatever you want is very little. They pretty much start delivering the best version of whatever you ask for, and they also thoughtfully push back on some of the dumber things that I wanted to do.
So much so that I’m thinking of canceling my Claude Code and Codex subscriptions and just switching to Kimi Code and the GLM—along with checking out DeepSeek. They’re similarly priced, and the quality is, in my context, better than Claude Code and Codex, so I have no reason to stick around with them.
By the way, if you’re looking for some reasonably priced alternatives to Claude Code and Codex, I highly recommend checking them out.
Nearby leaves, connected by subject and form.