AI-generated code should meet the same review bar

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 1:13:25

   code but maybe it's not exactly how you'd like it or maybe the end result has some non-obvious problems with it ai coding is a tool like every other tool that we use and you have to learn how to use it so i think one of the most you know karpathy obviously like he knows how to code i think a lot of people that are new to this kind of tool they tend to just ask it to do stuff that's a little bit too big, or they hold a different bar for the model's code versus their own code. And so something that I do for the team is for the quad code team, we have the same exact bar, regardless of whether the code was written by the model or by a human. And so if the code sucks, we're not going to merge it. It's the same exact bar. And you just ask the model to improve the code and make it better. I think there's also kind of different ways of wielding these tools.

2. 1:14:12

   Sometimes you want to vibe code. And this is really important for throwaway code and prototypes, code that's not in the critical path. And I do this all the time, but it's definitely not the thing you want to do all the time because you want maintainable code sometimes. You want to be very thoughtful about every line sometimes. So the way the problem, depending on the problem, you want a different approach.

3. 1:14:34

   And so there's like a set of approaches that I'll use, like sometimes I'll vibe code stuff. This is actually quite rare. It's actually mostly for like prototypes and things like that, that are throwaway code. Usually I pair with a model to make to write code and so first we align on a plan you know so this is like shift tab and quad code to get into plan mode and first the the model will make a plan then i'll see the code and then i might ask it to improve the code or clean it up or so on but it's a very involved process i'm pairing with the model we're working together to to write this code and then sometimes i'll still write the code by hand so you know there's some parts of our core query loop where i have very strong opinions about like you know things like the names of parameters or kind of like on

4. 1:15:15

   which particular line of code is some some code is and so for this i'll still write it by hand i think the the thing though is the models i think are still overall not great at coding i think there's still so much room to improve and this is the worst it's ever going to be but it's just insane to me to think a year back where the state of ai coding was you know like type ahead and it's just a completely different world and you sort of think about like where this is headed and kind of what's about to happen and what this looks like over the next few months and few years. And yeah, I think for me, what keeps me really excited about it is just having the context about this trajectory that we're on. When people hear cloud code, they think about coding.