Comment by spapas82 | Hacker News Reader

Comment by spapas82 | Hacker News Reader

I'm 100% on this since Electron is probably the worst thing that happened to the desktop development.

Hiwever, I'd like to hijack this front page article to for a related concern/question I've got.

It seems that all (popular) opensource LLM harnesses are either in JS or python (pi, hermes, opencode, paseo). The same seems to be true for closed source harnesses that I have used (deepseek, chatgpt).

I'd really like to ask why? I mean why somebody would use technologies like JS or python to implement a harness application instead of golang or rust or even C? I understand that before 1 year JS or python would be tempting but right now? Especially the LLm related companies have unlimited tokens and re-writing the harness software to golang or rust should be trivial.

This would benefit so many people (especiall considering the prices and scarcity of RAM these days).

So, why nobody does it? This is an honest question, I don't want to offend interpreted languages but it should be a universal truth that a compiled language is faster and consumes less ram.

One good reason: extensibility. Pi is extremely malleable in ways that would probably be harder to achieve if it were in, say, Rust.

The resource usage of my harness is nothing compared to an electron app, and being able to easily customize it is more important to me than saving a few extra MB of ram.

I guess it depends on the harness, for example: chatgpt.exe right now is 1 GB on my windows machine. peseo.exe is 500 mb. and opencode.exe + bun.exe is 300 mb.

Chatgpt isn't a harness. OpenAI's harness is codex.

Because most of the recent improvements on correctness and 'memory' are from harnesses, my guess is that you want to iterate and test quickly on the software that seems to improve the most and differentiate your brand best. Once a harness is 'complete' (which might happen for specific harnesses, but probably never for code), the team might start to optimise it with compiled languages.

The problem is that they never reach this point where they put it into native code. So you always have a plethora of JavaScript and Python files littering your system. It's baffling.

Because it used to cost extraordinarily more money to build for a performance increases almost no one cares about.

chatgpt.exe right now is 1 GB on my windows machine. I have 16 GB and can spare none. So I don't think that nobody cares about such performance increases.

Most don't care or ever notice these things. Hacker News crowd does and always has, but the a non-technical user of a run-of-the-mill desktop app would rather have new features, and maybe some bug fixes, over performance.

Myself? I care. I regularly run a project that needs 6GB for the language server to provide type hints.

> but the a non-technical user of a run-of-the-mill desktop app would rather have new features, and maybe some bug fixes, over performance.

Ask anyone with a 5 year or older laptop that wasn't top of the line when it came out and can't afford a new one, and I think you'll find they're of the exact opposite opinion. And even if they aren't explicitly, they might have abandoned their PC and instead switched to their phone for their personal computing, just because that experience turned out more coherently for them.

Heck, even the normal user with a decent laptop probably would take better performance if given the option, but when have they ever been given that option? The current state just isn't unperformant enough that it crosses the threshold to the point where it's at the front of mind for people. If anything, that threshold is probably beyond the threshold of 'eh, I'll just use my phone for this' for any task that doesn't need a keyboard.

> chatgpt.exe right now is 1 GB on my windows machine. I have 16 GB and can spare none.

These are not the same units, because all modern OSes have compressed swap.

But you’re running it, aren’t you? So from the business perspective, you don’t care enough for it to matter.

One possible reasons is that all of those are compiled languages vs interpreted. It doesn’t sound like much but compiling your code for three plus platforms and distributing binaries adds just enough complexity and friction to make it feel a lot more difficult.

I agree with that yes but consider how many people would benefit. And how easy it should be for them to do the change (and compile). It's just tokens.

But "friction" is the job, and "complexity" is the job. That's why we're developers.

Because the (wall clock) execution speed / performance oft the harness is tiny compared to waiting for the LLM response.

Reply on news.ycombinator.com