I think you're over generalizing. Software rot happens for a multitude of reasons
- hardware changes
- firmware changes
- operating system changes
- dependencies change
- bits flip, thanks cosmic rays!
> Right now, it feels like AI coding is ~2.5
A lot of people would argue that it's not even 1. I'd argue it sits between 0.7 and 1.2 given the task.
I don't think it's close to 2. It's hyper fixated on task implementation. It's really bad at abstractions, even when requested. You have to be pretty precise over these instructions. I've only seen it develop well, even with explicit instructions, a handful of times. But hey, that's true about a lot of people too. Though people will not claim to have solved the proper generalization, people will just say "I just care that it works", which is kinda why AI is trained that way. It's not being programmed by your Knuths, it's being programmed by your Zuckerbergs.
In the case of Flipper, we're talking about fixed hardware (granted, eventually parts may need to be substituted) and firmware as the rotting target, so no OS level above that.
Which makes it a simpler problem.
I'd say SOTA AI coding systems are better than 1 and still improving. They often build better code than junior engineers. And with prompting and feedback harnesses, they're starting to grok architectural alignment.
> I'd say SOTA AI coding systems are better than 1 and still improving.
I'm including Fable. I use agents every day. They still are very sloppy and hyper fixated on tasks and are poor at making code flexible and maintainable in the long run.
But again, I'm not surprised because that is the modus operandi of the typical Silicon Valley programmer (irrespective of seniority). They leave in their wake bugs that are avoidable and unnecessary and significant amounts of tech debt.
But then again, no one ever gets credit for squashing bugs before they exist. The agents are just as myopic as we are
I don't think it's close to 2. It's hyper fixated on task implementation. It's really bad at abstractions, even when requested. You have to be pretty precise over these instructions. I've only seen it develop well, even with explicit instructions, a handful of times. But hey, that's true about a lot of people too. Though people will not claim to have solved the proper generalization, people will just say "I just care that it works", which is kinda why AI is trained that way. It's not being programmed by your Knuths, it's being programmed by your Zuckerbergs.