Posted by thunderbong 10 hours ago
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.
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.
These are not the same units, because all modern OSes have compressed swap.
Myself? I care. I regularly run a project that needs 6GB for the language server to provide type hints.
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.
Raycast, originally a native macOS launcher, recently released v2 using a hybrid architecture. It’s not Electron, but it follows a similar approach to Spotify’s desktop app, using web technologies inside a native shell.
The memory usage is still pretty significant. It’s using around 350–450 MB at idle and can sometimes reach 1–2 GB depending on the workload. The upside is that they can now share most of their code across macOS and Windows and evolve beyond being just an app launcher.
It’s always been about trade-offs, and LLMs don’t fundamentally change that.
GPUI is one example: some of the downsides of electron (UI is not native), but it solves memory and performance side, and shared codebase.
Just writing native apps because LLMs makes it cheaper is the other option. Shopify did this after years of react native.
The opposite is also true. My argument is that you can also ship Electron apps with more features more quickly with LLMs.
You still have a finite token budget. With native apps, you have to split those tokens across multiple platforms, whereas with Electron, you can reuse most of the code.
Big companies have been going back and forth between native and cross-platform approaches even before LLMs. They generally have the resources (humans, tokens, compute) that small and medium-sized companies don’t have. So I’d take their rationale with a grain of salt.
Also the GPUI approach is still finite tokens, just better stack.
Also, pro tip, you can still just use it in browser. I hope that doesn't go away.
Discord is a server-side protocol (the client isn't doing much besides displaying the server-prescribed state), there's very little logic/computation going on on there that should explain a world of difference.
Nowadays, AI (LLMs, genAI, call it what you will) has become so cheap to handle more than one codebase, they no longer feel the need for web only approaches.
I'm well aware that Electron has a large RAM footprint and installation size, but doing a polyglot architecture instead puts more cognitive load on you. Projects that mix multiple languages are a pain just to manage. I'd rather hope Electron gets better optimized.
Discord is ditching Electron and its new app uses less RAM