My question is how far does it go - are the gains going to peter out, or does it keep going or even accelerate? Seems like one of the latter two thus far.

I feel like this comes about because it's the optimal strategy for doing robust one-shot "point fixes", but it comes at the cost of long-term codebase heath.

I have noticed this bias towards lots of duplication eventually creates a kind of "ai code soup" that you can only really "fix" or keep working on with AI from that point out.

With the right guidance and hints you can get it to refactor and generalise - and it does it well - but the default style definitely trends to "slop" in my experience so far.

I would guess the same way humans do.

Put brain in creative mode, bang out something that works

Put brain in rules compliance mode and tidy everything up.

Then send for code review.

Get a grip.

You know as well as I do that you are simply marking and correcting your agent's work. I hope it works for you but why not be more rigorous in the first place when you do stuff?

I would expect this conf to expand on those types of concepts and strategies.

Why would they need to do that? Is that even a goal or something that this conference is addressing at all?

Really that’s the core of it

Gimme an in without calling me an idiot for what I’ve been doing up until now!

This is not at _all_ my interpretation of Casey and JBlow's views. How did you arrive at this conclusion?

> They're more concerned about user experience and efficiency than they are about correctness.

They're definitely very concerned about efficiency, but user experience? Are you referring to DevX? They definitely don't prize any kind of UX above correctness.