Fear of Coding Out

A first taste of a 24x7 AI

My mind has been blown this past week.

I’m not sure I can ever go back to the way it was.

I felt this way getting my first personal computer in the early 80’s.

Again two years ago with the launch of ChatGPT.

And now, with the launch of OpenAI Codex for the web.

Last week, I wrote about how well OpenAI executed the Codex launch. What I failed to anticipate was how transformational that product was going to be.

For all of my non-technical readers, Codex is a software programming agent from OpenAI. At one level, AI tools for software developers are now old hat. As I’ve written a lot about over the past few years, the basic idea is that these tools help developers write and understand code faster. They’ve been truly remarkable.

However, up until last week, as amazing as those tools are, they fundamentally worked only when the programmer was working.

Now, that has changed. Codex allows you to create independent agents that can run coding tasks on their own. Google shipped something very similar called Jules, and Microsoft did as well with Github Copilot Agents. I’ve been using all three over the past week.

This is what has been crazy.

I’m now coding even when I’m not “coding”.

So far I’ve managed to have up to twelve different AI agents all programming for me at the same time. And in the meantime, I’m doing something else—email, board meetings, even bike riding. Because Codex works on the web, I went on a five-hour-long bike ride last Sunday with a dozen agents running. I had my phone attached to my bike, and from time to time the different agents would alert me as to the progress they made.

Admittedly I had to stop a few times to give some guidance to the AI, so my five-hour bike ride really should have taken me no longer than four hours! It was still eye-opening—and as a side note for any readers in the Seattle area—this ride was training for the Seattle to Portland 200+ mile bike ride later this summer. Hit me up if you’re going!

This is not just an anecdotal feeling of progress. Out of curiosity, I had one of my AI agents write a program to analyze the code I’ve written over the years. I picked three times of “peak” productivity, where I was coding regularly. 2013 and 2022 were very focused years where I had few meetings and could code regularly during the workday. Then I picked this past winter and the past four days since Codex’s launch. One thing to note about these two recent periods is that my work days are filled with meetings of various kinds, and I can only code in the evenings.

Here are the results, in average lines of code per day:

For the technically astute reader, yes, I know that Lines of Code is a limiting and imprecise measure of productivity. However, it’s directionally interesting, and given the magnitude of the change, it’s an easy way to quantitatively articulate the impact of Codex!

More importantly, I know I’ve only scratched the surface of what’s possible and how far I can push this.

The productivity gains are so addictive that I don’t want to sleep, go to lunch, hop on a Zoom call, or pretty much anything else without setting up some coding agents to work for me in the background.

What happens when that process gets even more automated? What happens when I can set up not just agents to do my work, but also create agents of agents to coordinate whole, large effort tasks?

Truly, my mind is blown. There is no going back from this!