Development experience of AI tools is becoming the biggest blocker. And a large part of that is performance (in time, not qualitatively).

Most of us are doing relatively simple things like using with the largest/slowest models possible instead of downgrading to a smaller model. Reason: our tools lack the ability to escalate to more expensive models as needed and switching models manually is tedious. Ideally, a cheap locally running model would be first in line and respond quickly for quick things.

And then there's the whole asynchronous vs. synchronous thing. With Codex, it runs in a browser and it allows you to create a pull request whenever it is done. You can work on multiple pull requests with it even and it might work in parallel.

What's good about the Codex experience is that your input is only needed at the end. What's bad is that it takes ages. Even for simple stuff. Like a simple follow up question results in it boiling the oceans to startup the container again.

Slow AI is exactly like slow builds: frustrating and likely to distract you. If it's going to take a minute you are going to do something else. And that might not be work related. So it breaks your flow because you are sitting on your hands and filling your short term memory with garbage. Context switches are expensive (and break flow). Our brains don't do that well. And then you forget to switch back so you lose time.

I don't use reasoning models that much for this reason. It's easier and faster for me to manually patch up my code with whatever the LLM says I should fix. And on larger repositories the chance of a good PR drops sharply. So, even if it takes more context micromanagement to feed it all the detail it needs, this can be faster and more effective. And I get an answer in a few seconds instead of in a few minutes.

I'm really really enjoying the Codex UX so far, though. Having it in the cloud makes it so much safer, as it can't access my local data, and the slowness isn't so bad if you make a few changes to how you work on a project. It's not an app for vibecoding, it's essentially an intern