Corey Knowles

And then you could take like arguably an even more specialized function like research and like researchers are using Codex a ton as well.

as part of how they're doing training runs.

They're doing a lot of interesting things also across research and design and other functions to analyze and understand what's going on.

Previously, it would be quite expensive to write code for a one-off analysis, but now you can, or you could write a custom viewer to evaluate how the model is doing.

We're even starting to experiment with using Codex as part of deciding how to handle training runs as they're ongoing.

I don't think this was all in response to like I don't I think like people like me, I'm thrilled to have these tools because as a generalist, it's very useful for me.

But I also think that if you're a specialist, you can actually use it to just like go way faster down your specialty as well.

Totally.

Right.

It was, you know, again, it was vibe engineered.

So a ton of the code is written using codecs, but there were humans like really like highly qualified humans in the loop the whole way through, right?

Like really great engineers who like know exactly what they're doing and you've been using AI for a while.

So I think if that is your goal,

then you should be treating Codex as a massive accelerant to your own productivity.

But ultimately, you are still treating this like an engineer.

You're being very thoughtful about architecture.

You're defining that before you get started on what you're building and telling the model what the architecture is, using tools like planning to make sure that it builds in that way, maybe being more modular with code than before so it's easier to test and easier to trust units of code without

having to make a ton of changes or even arguably review all the code super carefully.

Yeah, you definitely want to be doing things like writing in typed languages.

Actually, I really think even compiled languages are kind of almost there's a potential comeback for them, like codexes and Rust, because it gives you even more confidence that the code is going to work if it passes its validation conditions.