> Can't help to think of a recent HN post about most AI-generated projects being abandoned within months. Why?

Because we have a near infinite number of artificial monkeys figuratively typing at a computer keyboard and, once in a while, they spout out something that looks like an acceptable program.

But it's not acceptable: it's sloppy-pasta.

> Because value of a project is not in the code produced. It's in the amount of battle-testing that code has seen.

Yeah. I don't know if we live in the same world as some other HNers: I pay not one but three AI subscriptions: Google (through Google Workspace), Open AI and Anthropic. I'm switching from Claude Code to pi.dev atm.

I'm using LLMs daily. The person next to me too.

LLMs produce shit code. Pure unadulterated shit.

They're extremely useful for many things, including finding bugs. But, to me, they simply suck at fixing them. They write horrible, verbose, code full of bogus assertions and cases that are not handled or not handled correctly.

I do verify the output they produce: the number of time it "works" but I look at the code and see sheer horrors and say stuff like this: "wait, if you replace A by B like you suggest, shouldn't thing X that A makes be done too?". (it's a rethorical question: I know I'm right).

"good catch".

No. It's not a good catch. It's a totally obvious catch that any programmer who's been in this since more a few years (a few decades for me) would catch instantly.

Now if there's one area were they're semi okay'ish is for porting code from one language to another. At least you get a base to start from: if you're lucky. And I suspect it's only looks good if you don't know much about the target language.

Also don't be fooled: we're talking about a company with tens of billions investments that fully knows AI is not enough. You can be sure there are countless programmers fixing, behind the scene, the sheer mess that their AI (with unlimited budget btw) is creating. And then they pretend it's a 100% automated AI rewrite.

AI is not only definitely not enough: it's useful for finding bugs, it's a good rubber duck, but the more and more time passes, the more I'm appalled by the shit code it produces and by the errors they make all the time.

If I wanted to be facetious I'd say that at least we're already --how things moves fast-- long past the point were the kool-aid drinkers explain to us that these thing are intelligent.

It's summer 2026 and, as I type this, the SOTA models all suck at writing code.

Am I still deep into it? Sure am. For basically everything but coding (finding bugs, documentation, translation, generating assets, automating dumb stuff, ...).

Oh well, going back to pi.dev facepalming myself and rolling eyes. I know it's going to be "good".