Apparently everybody needs to have a take on AI and code, so here's mine.

First, background. I'm a software developer with over twenty years of experience. The vast majority of my work has been web development on the back-end.

I originally wrote this post on LinkedIn, where less than half of my connections are fellow programmers. I think I have a higher percentage on the Fediverse, so more of you will be familiar with these concepts.

Personally, I don't use any AI code generation tool in my work. Not because I think it's wrong, or anything like that. I've tried it a few times and haven't found it to be particularly useful for me personally. But if others find it useful, I think that's great.

So, the big question that everybody seems to have an opinion on is whether or not AI will replace programmers in the future (whether that's three months or thirty years).

One thing that's implied in this question, but nobody seems to be outright stating is whether or not AI will ever be able to program itself. Keep that in mind.

There's a concept in computer science known as "Garbage in, garbage out". What this means is that if you have a process or algorithm and you feed it garbage input, you're going to get garbage output. You can't expect a computer to give the right answers, if you give it the wrong information.

This is an important concept in generative AI. What something like a large language model does is it takes its input (which is a whole bunch of text), and matches it up, correlating words that are frequently found together, and then spitting them back out upon request.

So, if the underlying data is flawed, you're going to get bad results.

But generative AI actually amplifies this. Because it's not intelligent (the emphasis in "artificial intelligence" should always on the "artificial"), it doesn't actually reason about any of this. It's just matching patterns. So, with AI, instead of "garbage in, garbage out", it's more like "garbage in, really stinky garbage out").

But even with really good models, the results of generative AI are always going to be inferior to the dataset that their based on, because, despite how it might look, it's just matching patterns. It's not reasoning about anything. So, some patterns might be very relevant in some contexts, but completely irrelevant in others. And without the ability to reason about them, a computer can't make that determination.

So, when it comes to code generation, the first thing you have to consider is the source of the data, the code its trained on. As far as I can tell, these are trained on whatever code is on GitHub and StackOverflow. Some of this is wonderful. Some is horrible. So, the bad code is already mixed in with the good code.

This makes a really good coding model great at assisting code generation for common patterns. But it makes it useless when generating new algorithms. So, it's great at throwing together code quickly, even getting you to a fast MVP. But it doesn't help with more complex tasks let finding bugs, or even more, designing a new system architecture. It can only solve already-solved problems. And the very nature of it means that that is all it will ever be able to do. So, no, it can't replace programmers, and never will, but it can help them. Especially for basic tasks, and especially for programmers that are less experienced in their career.