![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
amaraoInteresting part of review, is when you learn something from wiser man code, or teach someone how to do it right.
AI slop is the most horrible for review.
It works, passed CI, everyone are (presumably) happy.
It subtly broke abstractions or introduce new requirements or assumptions on environment which were not here before.
Normally, if someone spend a week writing +5000 -2000 PR, you can respect this person time and spend a day or two, trying to understand deeper reasoning. If you find a flaw in the reasoning, you explain this to the person, and that's another week of work of it.
But with AI you spend a day reading +5000 -2000, find a flow, and your carefully crafted review send to AI which says 'Nicely spotted' and rewrite another +4800 -1800 PR, which someone need to read again, and it can be in complete disregard of what was said in review (if it's too abstract).
The main problem with +5000 -2000, is that if I spend 20 minutes writing an explanation for the problem (even with mvp to highlight the issue), I expect that person to learn something from it.
With LLM it's all wasted time, you get slop back, and next PR will not learn it. Partially it can be alleviated by updating AGENTS.md, but it can't be too large, so long explanation just waste tokens.
So, it's a pipe of slop and you need to review it.
What happens in reality? You skim across codebase, looking for random bits which looks odd, degrading to pattern matching machine. It's impossible to understand +5000 -2000 lines of code in 20 minutes. But they get merged. You accept them, other accept them, but codebase is slowly drifting, so you need code janitor to come and restore original semantic at the end.
Yes, we all do this, and it's terrible. Yes, it's real-politic productivity gains, but there is some kind of horrible built-in negligence in all of this.
AI slop is the most horrible for review.
It works, passed CI, everyone are (presumably) happy.
It subtly broke abstractions or introduce new requirements or assumptions on environment which were not here before.
Normally, if someone spend a week writing +5000 -2000 PR, you can respect this person time and spend a day or two, trying to understand deeper reasoning. If you find a flaw in the reasoning, you explain this to the person, and that's another week of work of it.
But with AI you spend a day reading +5000 -2000, find a flow, and your carefully crafted review send to AI which says 'Nicely spotted' and rewrite another +4800 -1800 PR, which someone need to read again, and it can be in complete disregard of what was said in review (if it's too abstract).
The main problem with +5000 -2000, is that if I spend 20 minutes writing an explanation for the problem (even with mvp to highlight the issue), I expect that person to learn something from it.
With LLM it's all wasted time, you get slop back, and next PR will not learn it. Partially it can be alleviated by updating AGENTS.md, but it can't be too large, so long explanation just waste tokens.
So, it's a pipe of slop and you need to review it.
What happens in reality? You skim across codebase, looking for random bits which looks odd, degrading to pattern matching machine. It's impossible to understand +5000 -2000 lines of code in 20 minutes. But they get merged. You accept them, other accept them, but codebase is slowly drifting, so you need code janitor to come and restore original semantic at the end.
Yes, we all do this, and it's terrible. Yes, it's real-politic productivity gains, but there is some kind of horrible built-in negligence in all of this.