AI YouTube channel
Discussion in Non-Morons
https://www.youtube.com/@matthew_berman
And the contrary view:
https://www.youtube.com/@kiraa_ai
I found [this video](https://youtu.be/dwD3Gd0Cq1w) against AI harnesses interesting.
Interesting. I’ll watch it when I have time!
mmHg
edited
> [fschmidt](/thread.html?thread=1376&post=1431#p1431)
> I found [this video](https://youtu.be/dwD3Gd0Cq1w) against AI harnesses interesting.
I don't know how many tech entrepreneurs would say that "the value of AI comes from the harness, not the LLM or from prompt engineering," but probably close to none of them. Whoever thinks those things are mutually exclusive is a moron. The first guy in the video seems to think they're mutually exclusive, so he seems like a moron to me. Anyway, I would think that tech entrepreneurs generally recognize LLMs, prompt engineering skills, and AI harnesses as complementary benefits that go hand in hand (and I would agree with them).
For AI-driven software development, one needs an LLM, and not all LLMs are created equal. Prompt engineering skills are also beneficial regardless of which LLM or harness one uses. Are AI harnesses beneficial? I think so. But not all AI harnesses are created equal, either.
With a ChatGPT Plus subscription, you can use "Work Mode" from your browser or from your desktop via Codex, which is itself an AI harness. (For what it's worth, Claude Code is also an AI harness, so an argument against AI harnesses should, by default, apply to Codex, Claude Code, etc.) Codex is just far more capable than ChatGPT's browser interface for software development. It can use your computer for you by taking screenshots and then writing Python scripts that point and click things on your computer based on those screenshots.
For example, I use Linux Mint, but I run MultiSim on a Windows 10 virtual machine. Codex was literally able to construct a complete and fully functional ALU circuit for me in MultiSim through my virtual machine. The fact that Codex can use your computer and run commands also means that it can work directly on your local files without requiring you to download ZIP files, merge or replace old files with new ones, and so on.
Anyway, I mostly use Pi and supply it with AI models via Codex. If I worked with a small team of software developers who had their own AI subscriptions, I could connect their subscriptions to the same Pi harness and coordinate multiple models working simultaneously, and so multiply our work efficacy. Pi is just simple, flexible, and fully functional software. [And its creator seemingly has a reactionary programming ethos](https://mariozechner.at/posts/2026-03-25-thoughts-on-slowing-the-fuck-down/). The Pi harness offers many extensions made by the community and allows you to create your own extensions. I use my own Luan extension, as well as an extension for a fork of CodeGraph that reads Luan. This helps save a lot of credits that would otherwise be burned unnecessarily.
Another advantage of Pi is that it's the type of harness that would be perfect for running AI agents locally on your own machine instead of through a provider. If AI models become too modern and incestuous in the future, then we may need to run relatively old models locally or something like that. Pi would provide the perfect environment for making this possible.
I have never tried omp, Prime Agent, or jcode, but I am not really interested in trying them. Yes, they appear to burn far more tokens than usual and produce messy code. Those are examples of generic harnesses that conform to modernity, which Pi does not. Pi is in a category of its own among AI harnesses. I would anticipate myself arguing against omp, Prime Agent, jcode, etc., whilst arguing for Pi and maintaining a somewhat neutral attitude toward Codex and Claude Code.
The first guy failed to offer any convincing response to these benefits of AI harnesses, so I am disappointed by the video overall.
Okay, I can't comment because I haven't used a harness. But so far I can do everything I want through Claude's web interface.
Claude-web can use a headless browser. It can use Python for whatever it wants. I can spawn agents. I can't think of any files on my computer that I would want to give it. To share with a small team, I would use a shared login. Then I can read their chats. I don't know anything about CodeGraph but I assume that Claude-web could use such extensions.
I will look at harnesses eventually, but I focus on what I need, and right now I don't feel any need for a harness.
Log in to reply.