The New AI Bottleneck, Faster Coding, Slower Shipping

Yeah, I mean, um, you know, obviously when you're building, uh, software and, uh, a lot of the energy, um, prior to AI was, was kind of bulked into the middle, right? The build process.

And even though you would, you would spec things out and you would create kind of PRDs or-

Mm

... product requirement documents to kind of outline the scope of what you're trying to build, you know, whether you're a product manager or a developer, you, you kind of ru- you kind of jot stuff down, but you're not really, um, not to the granularity, you know, right, that you would expect, right? You, you just kind of-

Mm-hmm

... kind of high level it because you know you're gonna have plenty of time to dig into it when you get into the build process.

And then, uh, on the QA side, because you don't have as fast of a build, you know, prior to pr- pr- pre-AI, you know, the, the, the QA side was kind of part of the process too, right? You would finish out a task, it would kind of go through a process.

It was the next gate in the process, yeah.

There was no, there was no like, yeah, there wasn't any bottleneck there.

But now what you've cr- what we've created is like, um, more of an hourglass, you know, on its side, um, where a lot of the work now is on the front end, right? Uh, because in order for you to avoid AI slop or AI drift and, and, and really try to give it as much information as you can, you want to build out a really sophisticated, detailed, um, product requirements document along with, um, a testing plan ahead of time.

Mm-hmm.

And then when you plug it into the AI and let AI go, go to work, then that's where it kind of really streamlines and it starts coding very fast, um, and starts building things, you know, at record speed.

And then on the back end, you know, you have the second bottleneck, which is, okay, now we're generating so much code and, and, uh, how do we test it, right? How do we QA it? And of course, you can use AI for some of that, but, you know, human in the loop, uh, is definitely required, um, to, to make sure that you're, you're scanning code.

And then of course, when you're making, uh, you're going through that QA process and you have to make some adjustments or make some, uh, iterative changes, uh, or AI just kind of goes off a completely different direction than what you would like to see, um, you know, you have to cycle back and, and obviously, um, kind of readjust or re- re- re-prepare Um, you know, your new prompt, right? Your new, your, your new requirement.

So I think there's a back and forth that happens.

Um, everybody has their own style on how they do it.

It also dramatically depends on whether you're having a brownfield, right, an existing software product that was built pre-AI, and you're trying to push AI into it-

Yeah

... or if it's a completely new AI greenfield project, which is certainly different, um, you know, it-- So there's a lot of factors that go into play, but certainly what used to look like, you know, um, 80% of your time was spending building, now if you're kind of depending on AI to do that, you're now having to reframe your, your time, and you're actually kind of writing out the specs, right? And you're spending time testing and writing out the test scripts.

So and this is something that most developers, um, you know, really never really took time to do, and now developers almost have to relearn, uh, how to do that.

Of course, you can use AI to help, but at some level, there's a lot more coordination that has to happen.

Yeah.

There's definitely like a, an old dog new tricks kind of vibe, I think, when it comes to some of this stuff.