OpenAI’s Post

Transcript

Hi, I'm Joey, I'm an engineer in the Codex team. Today I'm going to be sharing a feature we've been working on which is best to fend for. Codex Best event will run multiple versions of your task and allow you to choose the best result. In my experience this has been incredibly powerful to hunt down really tricky bugs as well as exploring different implementation strategies. I'm going to show an example of a real task I worked on a few days ago to fix a bug. In this case, I briefly described the bug. I also offer a couple of solutions to fix it. I'm going to set this to 4 versions to give the model multiple attempts. Set it off to cook. While that's running, let's open up the previous task that I was actually working on and we can see here it also had four attempts. Generally, when I'm looking at best and task, I scanned through the code to see if it has the general strategy I was thinking of. And then I dig in deeper later. This also seemed to provide a couple planning steps instead of writing code right off the bat, which if any of the coding results weren't up to par, I could also just work off this and work with the model that way. However, in this case version 4 seemed to be exactly what I was looking for, and after testing locally I was able to merge this into our code base and fix the bug. I think Best of Men has dramatically improved improved reliability for codecs and I'm very excited to see what engineers are building this.