>What I am saying is that once the high quality training data runs out, it will drop in its capabilities pretty fast.

That's more a misunderstood study that over time turned into a confidently stated fact. Yes, the model collapses if you loop the output to the input. But no, that's not how it's done.

The reality is that all the labs are already using synthetic training data, and have been for at least a year now. It basically turned out to be a non-issue if you have robust monitoring and curation in place for the generated data.

Why would the capabilities drop instead of stagnate?

>Is the world any better for them existing? The decline of coding and sw engineering skills in humans from outsourcing the practice of it to AI is it worth it and sustainable long term?

The world is going to be no worse than it was when humans transitioned from writing assembly to writing compilers for high level languages. Assembly is still necessary, but not that often. In the same way writing code is going to become less necessary as tools are going to be written in higher level language in standards and requirements documents instead of code most of the time, with more specific exact coding only occasionally.

Programmers were mostly solving the same plumbing problems over and over in secret because of "proprietary" needs to hide your code, but one million separate integrations of your billing backend with Stripe didn't really add to humanity. We're cutting out the boring middle drudgery and human effort is going to be freed up to work on the edges of human knowledge instead of tromping around in the middle.

You're fooling yourself. It's very easy to get demonstrably working results in an afternoon that would take weeks at least without coding agents. Demonstrably working, as in you can prove the code actually works by then putting it to use. I had a coding agent write an entire declarative GUI library for mpv userscripts, rendering all widgets with ASS subtitles, then proceeded to prove to my satisfaction that it does in fact work by using it to make a node editor for constructing ffmpeg filter graphs and an in-mpv nonlinear video editor. All of this is stuff I already knew how to do in practice, had intended to do one day for years now, but never bit the bullet because I knew it would turn into weeks of me pouring over auto-generated ASS doing things it was never intended to do to figure out why something is rendering subtly wrong. Fairly straightforward but a ton of bitch work. The LLM blasted through it like it was nothing. Fooling myself? The code works, I'm using it, you're fooling yourself.

I also did a native implementation of git so I could use an S3 compatible data store, your rails guru can't do that.

Objectively, my GitHub clone is still shitty, BUT it got several ways github is shitty out of my way and allowed me to add several features I wanted, no small one of which was GitHub not owning my data.

I don't know the shit out of Rails and I don't want to, I know the shit out of other things and I want the tools I'm using to be better and Claude is making that happen.

It's a little odd the skepticism to the level that people keep telling me I'm delusional for being satisfied that I've created something useful for myself. The opposition to AI/LLMs seems to be growing into a weird morality cult trying to convince everybody else that they're leading unhappy immoral lives. I'm exaggerating but it's looking like things are going in that direction... and in my house, so to speak, here on HN there are factions. Like programming language zealots but worse.