well i also think there’s another bottleneck.. knowing how to structure your conversations with the LLM to not muddle things up.

when i work with these models, I try to maintain same way i would do things in real life and i tend to notice that it’s kinda taking me some time to even get these tasks done because now i have to read each lines to be sure it wrote it the way i would

We can't even review human-written code (fast enough). One of my PRs took over a year to review.

The problem is not the code review per se - it's understanding of the code and the changes by the people who need to further develop and maintain the code. Even if the so-called AI reviews the code, the review comments still need to be reviewed by the same people.

I don't know what the solution might be.

[dead]

The bottleneck is and always has been writing good code, so reviewing is as important as ever. Interestingly enough, LLMs help tremendously with writing code but severely impact writing good code, so nothing has changed.