In your opinion, a codebase has become slop when...
@datarama because the tool usage is the slope, the second you consider its usage, you now have to deal with the reality of that choice, its now a permanent consideration, a condition to be managed.
@cyberquixote I think the tool usage being unhealthy and risky is different from (but related to) where the point is that it turns a project into slop.
@datarama agree to disagree, but it doesn’t mean im not considering your take. thanks for sharing.
@cyberquixote To illustrate what I mean:
A developer has daily chats with an LLM about a mental health issue, and otherwise doesn't use them. This is risky, but I wouldn't say that the dev's projects are slop now: The LLM isn't directly involved in the projects.
A developer has executive dysfunction and uses an LLM to pick from a list of tasks on the project to decide on which to tackle first. I still wouldn't say that the project becomes slop; while the LLM *is* now involved in the project, it might as well be "asking my hamster for advice" (although a very articulate and also non-sentient hamster, the use of whom will probably erode what little executive function the user has left).
A developer sometimes uses Google, and sometimes the AI Overview feature actually answers usefully and saves a click on a link to Stack Overflow or a manpage. Now it's more involved - but only as a syntax aid. I still wouldn't say the project is slop (though at this point I'd begin worrying about memory erosion).
A developer sometimes fires up an LLM chatbot and has a chat about technical choices when they're having trouble with indecision. The LLM has been instructed to never generate code, and when it does so anyway, the developer rolls their eyes and doesn't even read it. I'd *still* say it isn't slop, because the actual code getting written is still the result of a thousand little decisions made by the dev. But at this point, I'd worry that the dev is becoming *worse* at making decisions on their own.
...and so on. My point is, aside from the "LLM reviews the code for bugs, human uses their good judgment to fix them properly" use case that everyone trots out, I think there are a lot of ways developers can use LLMs that are, at the same time, 1) not a good idea, 2) not slop production.
(That is to say, the LLM does generate slop, since that is what LLMs do, but none of the slop ever lands in the codebase.)