Comment by delta_p_delta_x | Hacker News Reader

Comment by delta_p_delta_x | Hacker News Reader

Parent Full thread

delta_p_delta_x

Unironically, with LLMs, everyone should be looking to cut their system resource consumption (memory, disk, network, GPU, I/O traffic in general) and rewrite everything natively. Have a platform-independent business logic layer, and a platform-dependent UI/UX layer. Done. It's not even that hard, and has been done several times over; we quite literally have MVC/MVVM/etc that were invented two decades ago.

Operating systems developers should be looking to profile everything. Benchmarking and profiling used to be horribly time-consuming; nowadays, with agents, this is almost trivial. Then valuable engineer time can be spent looking at actually solving and optimising the workloads, from the kernel, to drivers, to user-mode libraries, to vendored applications.

About a decade ago programming was expensive, and we decided writing everything in JavaScript was OK, and wasted everyone's memory instead of using programmer time to write fast and efficient native applications. This is no longer the case.

Yeah I have to agree here - the days of being scared of complete rewrites are pretty much over.

> Unironically, with LLMs, everyone should be looking to cut their system resource consumption (memory, disk, network, GPU, I/O traffic in general) and rewrite everything natively.

Depends.

There is a lot of value in man-made software where the man knows exactly what happened where when why.

If you just revibe in rust, you get something that works at this point in time, but this deep architectural knowledge is completely lost and you're at the mercy of Dario not locking down your account for wrongthink.

That can still be fine for smaller tools where the cost of recreating them from scratch is alright, but revibing your core business thingy is likely a horrible idea, which I encourage people to do, as I'd love to have the business opportunity to take over their frustrated customerbase ca 5 years from now.

>and you're at the mercy of Dario not locking down your account for wrongthink.

Inference cannot be controlled. The models are already out there. It's too late.

Good luck doing the full-hands-off one-shot revibe stuff this all is referring to with GLM-5.3.

That needs claude-like capabilities.

It also should not be done at all. But that approach needs it.

Using open weights models like GLM-5.3 can also speed up rewriting, but then you're rewriting and not revibing, which is a lot slower.

I don't think we should just 're-vibe' it as you put it. Naturally the rewrite should be progressive and thoughtful, backed by test-driven development at all levels, and regression tests every step of the way. To reiterate: writing code used to be a lot more expensive, which justified foisting React, Electron, JS on everyone. No longer the case.

I've mentioned in another comment, but my biggest pet peeve is that AI agent applications (Claude, GPT, etc) are themselves just Electron wrappers over their web pages. What a waste.

Reply on news.ycombinator.com