[josephg]

I’m curious if - and when - LLMs change this. They’re very good at web apps. And they’re great at rewriting existing stuff. Just give them a well scoped /goal and go get coffee.

Theres lots of open questions about the future of our profession in the age of AI. But, playing with opus and fable, I think the future will be bright for our users. There is no reason any more for teams to put out junk that’s worse than what an LLM can do.

[xp84]

Unfortunately the LLMs are trained on what we've made, and there's going to be a ton more React garbage[1] in the training set than there are carefully-crafted websites like the article describes, so I don't expect a decrease in overengineered, bloated junk. If anything, I predict that the fact that you can shit one out in less time than before will have a different effect: A modest increase in bloat since an LLM won't mind adding a half dozen redundant and competing ways to do the same things in a large codebase, combined with a shorter mean-time-between-full-rewrites.

I think most of us have seen incredibly creaky codebases that are too buggy to be maintained any longer, where we make the hard choice to wipe the slate clean and build a new one.

We might find those rewrites happening every 12-24 months instead of after a decade.

[1] Frontend people, I mean no disrespect -- just that React & friends are (ab)used for nearly every website now, even those which map perfectly onto the "Simple document viewing with occasional submission of incredibly simple form data" model that plain HTML has always been perfect for.

[ChrisMarshallNY]

> Unfortunately the LLMs are trained on what we've made

This.

As I’ve pointed out in other posts, I work in two languages: Swift (my main language), for native app clients, and PHP, for backend work.

I have a lot of experience, with each language (12 years for Swift, and 25, for PHP).

My experience with an LLM, is that I get very good PHP, and fairly mediocre Swift. The Swift often looks like fancy tricks, that newer enthusiasts like to demonstrate. Lots of unnecessary threading, and complex, dogmatic, overengineered approaches to simple problems.

I suspect that this has a great deal to do with the availability of high-Quality public repositories. PHP is more than twice as old as Swift, and, by nature, is a much more open language.

[VincePlatt]

It doesn't take that much effort to put guardrails around your prompt to solve problems in a certain way and with certain frameworks and excluding certain others.

[xp84]

Who will be doing that? Only a small minority of developers pre-ai cared to attempt using HTML, so I don’t see them urging Claude to create efficient and lean websites in the future either.

[josephg]

Claude is remarkably good at performance engineering and ports. It only takes one person on your team to ask claude to do a round of performance profiling and tuning. Or ask claude to take your react site and set up a server to render parts of your site as static, cache friendly HTML.

You barely need domain expertise any more. Just ask claude to make it go faster and it will.

[N_Lens]

You also need to ask Claude not to make any mistakes /s

[tobyhinloopen]

You can order LLMs to use other patterns. I had an LLM recreate an app in a different stack with reasonable success

[XajniN]

You can, if you know what you want.

[lotsoweiners]

Lol

[josephg]

Better than a lot of web dev teams at least.