AI is trained on data. Where does that data come from? The internet. What is it? It’s the internet. It’s Reddit, it’s Wikipedia, it’s GitHub, it’s Stack Overflow. It’s internet content, not intelligent thought.
What?
Take the array of thoughts and opinions of software engineers on the internet and lay it out in front of you. The vast majority of those thoughts and opinions are probably going to be from people who have less than the average experience. This is because the software engineering industry has been doubling in size every five years or so. Half of the workforce has 5 years of experience or less. 25% has 10 years. 12% have 15 years. Do the math. The average content of the internet is going to be the thoughts and opinions of developers with five years of experience.
AI is trained on the content generated by the mass, not the fundamentals of the most intelligent.
Argumentum ad populum: Popularity != truth — Just because an agent says it doesn’t mean it’s right.
So?
Everything we want to do with an LLM that is advanced is a fight. The most advanced models don’t have access to the private Stripe or Google software repositories. And even if they did, we don’t know if Stripe and Google’s software is actually better than any other software at any other company. I want to believe it is, but I fear it is not.
Ask it about how to write software well and it will tell you YAGNI and DRY. It won’t tell you about the main line or about how to manage dependencies. It won’t tell you that code that is highly depended on should be highly abstract. It won’t tell you that code that nothing depends on but has many dependencies can be highly concrete. It knows about those concepts, but they aren’t tied deeply into its neural network.
If you ask your LLM about the stable abstractions, it will tell you. If you ask your LLM to do TDD, it will tell you. But by default, its core is probably modeled after the average contribution on GitHub.