Hacker News Reader

Comment by jb_briant | Hacker News Reader

I have a decade of experience writing web code professionally, in my experience, LLM is a true waste of time regarding web.

On the other side, I'm switching to game dev and it became a very useful companion, outputing well known algorithms. It's more like an universal API rather than a junior assistant.

Instead of me taking time to understand the algo in the details then implementing, I use GPT4o to expand the Unreal API with missing parts. It truly expands the scope I'm able to handle and it feels good to save hours that compounds in days and weeks of work.

Eg. 1. OOB and SAT https://stackoverflow.com/questions/47866571/simple-oriented...

2. Making a grid system using lat/long coordinates for a voxel planet.

View on HN

> I have a decade of experience writing web code professionally, in my experience, LLM is a true waste of time regarding web.

As someone who knows web front end development only to the extent I need it for internal tools, it’s been turning day-long fights into single hour dare I say it pleasurable experiences. I tell it to make a row of widgets and it outputs all the div+css soup (or e.g. material components) that only needs some tuning instead of having to google everything.

It still takes experience to know when it doesn’t use a component when it should etc., but it’s a force multiplier, not a replacement. For now.

I can totally understand it! I would loved to have access to a LLM back in time. Especially since I started to learn just after the <table> era but we didn't have "flex" yet. It was true garbage...

Now I mostly JSX / tailwind, which is way faster than prompting, but surely because I'm fluent in that thing.

> Now I mostly JSX / tailwind, which is way faster than prompting,

Not saying that is not true, but did you measure that actually or is it a feeling or you didn't spend very much time on getting a prompt you can plop your requests in? jsx and tailwind are very verbose ; frontend is mostly very verbose, and, unless you are building LoB apps, you will have to try things a few times. Cerebras and groq with a predefined copy paste prompt will generate all that miserable useless slob (yes, I vehemently hate frontend work; I do hope it will be replaced very soon by llms completely; it's close but not quite there yet) in milliseconds so you can just tweak it a bit. I am fluent at building web frontends since the mid 90s; before I did DOS, windows and motif ones (which was vastly nicer; there you could actually write terse frontend code); I see many companies inside and I have not seen anyone faster at frontend than current llms, so I would like to see a demo. In logic I see many people faster as the llm can often simply not even figure it out even remotely.

I prompt a lot everyday for various reasons, including, but not exclusively coding. Because jsx/tw is indeed very verbose, it requires a lot of accuracy because there is sooo many places you have to be just perfect, this is something LLM are inherently incapable of.

Don't get me wrong, it will output the code faster than me. But overall, i will spend more time prompting + correcting, especially when I want to achieve special designs which are not looking like basic "bootstrap". It's also much less enjoyable to tweak a prompt than just spitting jsx/tw which doesn't require lot of effort for me.

I don't have a demo to justify my claim and I'm totally fine if you dismiss my message because of that.

I recon that I don't like front ending with LLM yet, maybe one day it will.be much better.

My website where I tried some LLM stuff and been disapointed https://ardaria.com

That's also my experience, the AI is very helpful for doing common things with languages you're not great at. Makes it easier to pick up a few extra tools to use now and then.

This I agree with. It is awesome when you don't know something. I have being doing some TypeScript recently and it's not something I have used professionally in anger before.

But, as I learn it, GPT 4o becomes less and less useful for anything but short questions to fill in gaps. I am already way better than it at anything more than a few pages long. Getting it to do something substantial is pretty much an exercise in frustration.

> it outputs all the div+css soup

Is there a difference between soup and slop? Do you clean up the "soup" it produces or leave it as is?

It’s soup in the sense you need it for the layout to work in the browser. I clean up when I know there’s a component for what I need instead of a div+span+input+button or whatever use case is being solved. I mostly tweak font sizes, paddings and margins, sometimes rearrange stuff if the layout turns out broken (usually stuff just doesn’t fit the container neatly, don’t remember an instance of broken everything).

> I tell it to make a row of widgets and it outputs all the div+css soup (or e.g. material components) that only needs some tuning instead of having to google everything.

As someone with a fairly neutral/dismissive/negative opinion on AI tools, you just pointed out a sweet silver lining. That would be a great use case for when you want to use clean html/css/js and not litter your codebase with a gazillion libraries that may or may not exist next year!

Given how bad most DOM trees look out there (div soup, thousands of not needed elements, very deep nesting, direct styling, HTML element hacks, not using semantically appropriate elements, wrong nesting of elements, and probably more), I would be surprised, if LLMs give non-div-soup and responsive HTML and CSS, the way I would write it myself. Some day I should test one such LLM as to whether it is able to output a good HTML document, with guidance, or the sheer amount of bad websites out there has forever poisoned the learned statistics.

Same here. Really liking it asking it Unreal C++ trivia.

Reply on news.ycombinator.com