AI writes code like that, but people get pretty upset hearing it.

The program that I use to manage my small business, basically a GUI wrapper on an sql DB is a total delight to use, because everything happens instantly. Even opening it is instantaneous.

Modern software is such a bloated mess that you kind of get used to a 20 second start-up and a .25-1 second delay on every action. Then you use something that isn't bloated junk and it feels like actual magic.

>> AI writes code like that, but people get pretty upset hearing it.

To this day I never saw an AI write anything less than a 3Mb binary. Do you have specific examples in mind, and if so, can you share them?

Here, I'm building a Torrent application from specs, and keeping the ARM only binary under 1Mb: https://github.com/josusanmartin/rustorrent

It's fully vibecoded, and quite functional, although I'm still working out some bugs.

AI currently writes median-good code which is by definition better than the worst developers' code.

In order to push beyond that (e.g. architecture, size, performance), the human has to bring the constraints. There's only so much AI can/should default assume from "write code for me" (which I expect a lot of bad / non-HN developers are doing).

It is exciting that there's an opportunity for global improvement though! After the general model improvement pace slows, expect there will be a lot of room for code gen AI optimization, and if "write code for me" can be made to generate more efficient code... suddenly all AI-generated code (read: most code) will improve.

Shit code is out in the wild because the company decided to optimize for programmer labor cost instead of code efficiency. If instead apples:apples because AI code gen is being used both ways, everyone wins.

Feels like what I imagine were the early days of optimizing compilers, when I'm sure there were a large number of really bad assembly programmers out there.

AI doesn't get tired. I can tell it to profile the application and find quick wins and it usually does a decent enough job.

I am sure that's true if AI writes it, but what if YOU use AI to type it?

What happens if you keep asking AI to rewrite it to use less and less resources?

It’s actually pretty good at this if you have some very specific constraints and objective.

I find I get great results from explicitly asking an LLM to write tests which profile memory and cpu usage. Then ask it to optimise to improve those metrics.

And a way for it to get actual measurable results / feedback loop, which is apparently quite tricky even for humans to get.

just be aware of the exact prompt:

"memory usage is important": creates a 1TB binary

"take multiple steps": uses so much power, extreme global warming, it will just print "hot"

"think outside of the box": it will enslave you so you will not care about weather

This: https://news.ycombinator.com/item?id=49226923

Fans went crazy when loading the browser demo

It does a pretty good job being put into a loop of "try to optimize for X" then "test and audit your code" then repeat.

tbh, it probably works. I'm vibe-coding a thing and just told it to figure out why it runs so slow: put in some optimizations. It ran some tests and stopped loading an entire-something at opening because there's no value to doing that.

> AI writes code like that, but people get pretty upset hearing it.

Yeah, people get upset when you spew out total bullshit.

I work as a performance specialist and "AI writes code like that" is utter nonsense. AI writes the same terrible performing code that was the mediocre developer standard on Stackexchange and other codebases it ate.

You still need a lot of vigilance and time to make well performing software and most software engineers just don't give a damn if we don't outright block their PRs/feature launches before passing performance tests.

LLM coding assistants have a bias towards writing standalone implementations instead of pulling in libraries to do a job, which does tend to sidestep probably the most common cause of bloated software: pulling in too many layers of abstraction. Most software can be pretty damn fast (at least as far as users are going to notice) without needing too much of a focus on optimization, but just by not accidentally screwing it up.

performance specialist, really?

Yes.

AI does this when it's not warranted, and even happily writes two functions with the same purpose.

Forty functions with the same purpose are still better than importing react.

What do you mean my hook to dynamically build <span contenteditable="true" style="border: 1px dashed #ccc; padding: 4px 8px; border-radius: 4px;">input your username here</span> with inline CSS is worse than using <input>? All I need to do next is to listen to form submit to build a new formdata object to POST to the server. Easy!

Is your UI kit open source? I do think calling into the native OS toolkits is the way to go, when they exist.

I'm curious if you use winit for windowing, or something lower level that's just mac specific.

I ported the leptos rust web framework to run on native platforms. Currently it works on ios, macos and linux (GTK). Eventually want to also add support for windows and TUI as well. It uses taffy for layout - which gives you the same flexbox and grid layout engines that web browsers use. Components are all fully native for the platform, via their corresponding binding crates (objc, gtk4, etc).

UI library is here: https://github.com/josephg/leptos-native

And here's the schedule viewer app I made with it. It runs silky smooth on device: https://github.com/josephg/dweb-sched

The whole project is currently experimental & vibe coded. It's missing a lot of components, and it has bugs. I'm currently rewriting it by hand to do it properly.

SwiftUI’s diffing is such a mess that I wouldn’t be surprised if fine-grained reactivity of a toolkit binding UIKit like this worked better. Do you have decent support for long lists via UITableView yet? Edit: it seems so!

Yeah! I added it for my ios scheduling app.

I tried to learn SwiftUI a few years ago. XCode was horribly slow, and it would hang and crash all the time, even when working on small example projects. I'm excited by the thought that I can use any rust IDE to write UI code. Rust's compiler, toolchain and IDEs seem way more reliable.

The view! builder macro at the moment is designed to look like JSX, since it was ported directly from leptos. But I'm considering dropping the jsx and using a fluent API. Something like this:

    let count = RwSignal::new(0u32);

    let view = vstack().padding(16.0).gap(12.0)
        .child(label().text(move || format!("Count: {}", count.get())))
        .child(
            hstack().gap(8.0)
                .child(button()
                    .on_click(move |_| count.update(|c| *c = 0))
                    .text("Clear"))
                .child(button()
                    .on_click(move |_| count.update(|c| *c -= 1))
                    .text("-1"))
                .child(button()
                    .on_click(move |_| count.update(|c| *c += 1))
                    .text("+1"))
        );

Code is a bit less readable like this, but autocomplete and cmd+click all works out of the box in your IDE.

Yeah, hence why for one kind of like the way RAM prices are going, maybe it forces the industry to re-learn best practices and "bio-programming" instead of "fast-food programming".