Creating checkpoints by gaslighting a Postgres database

**Ryan Donovan** (0:09)
The next breakthrough in human connection might not be human at all. Raymond Yin of The Tech Between Us welcomes Leo Chen, head of US Operations for Engineered Arts, in the conversation about humanoid robots designed to engage and connect with the human world. Listen from your favorite podcast platform or visit mouser.com/empoweringdashinnovation.
Hello everyone, and welcome to The Stack Overflow Podcast, a place to talk all things software and technology. I am your host, Ryan Donovan, and today we're talking about what happens to databases when AI is the primary creator and user of them. And my guest for that is Bryan Clark, who is the Director of Product for Lakebase over at Databricks. So welcome to the show, Bryan.

**Bryan Clark** (0:55)
Hey, thanks for having me, Ryan. Really good to be here.

**Ryan Donovan** (0:57)
So before we get into the authority subject today, how did you get involved in software and technology?

**Bryan Clark** (1:03)
I mean, I've been working in technology my whole life. I just as a kid, super interested in it. I think my first computer was a Tandy Sensation, which is one of the first PCs available to people.
And yeah, my career has been just like this awesome arc. My first job was at Red Hat. I was working as an interaction designer on a number of different projects on the GNOME desktop, things like that. From there, I went to Mozilla. I got to work on Thunderbird and Firefox. From Mozilla, I went to GitHub. That's where I did most of my kind of developer tools, because I actually have a background in computer science. So this whole time, I worked as a designer and then transitioned into product management, because I felt like I had this superpower between understanding the technology, but I actually did a lot of psych minor in school. So I understood how people think about technology, how they understand it. So I spent a lot of time dev tools at GitHub. And then from there, leaving GitHub, I basically wanted to work on Postgres. And so I had this journey out of GitHub. I was like, where can I go work on Postgres? Because that's what inside GitHub you could see. Every developer was switching to start using Postgres in their CI TV and things like that.
So yeah, I ran product at TimeScale. From TimeScale, I went over to Neon, and I was a VP of product at Neon, and then Databricks acquired Neon last year. I think I've been at Databricks almost a year now, and I run the Lakebase group at Databricks.
And yeah, it's been an awesome journey, mostly focused on developer tools and user experience.

**Ryan Donovan** (2:45)
All right. Yeah, a lot of good logos on that CV.

**Bryan Clark** (2:50)
It does help with the resume.

**Ryan Donovan** (2:51)
Yeah. Obviously, you mentioned you wanted to work on Postgres, and Postgres is pretty consistently the number one ranked database on our developer survey year over year.
But there was a stat you gave in the pitch where it was like, 80 percent of databases are created by AI agents. What happens, and I'm assuming a lot of them are Postgres instances, what happens to the database when AIs are just spinning up them autonomously?

**Bryan Clark** (3:19)
Yeah. I mean, it's a crazy stat when you first see it, and actually, it has jumped from what was nothing a year or two ago to that and even more now. It's just how much better AI has gotten at using software. The thing that we saw, and I think the thing I had told you earlier, is we learned how agents are so sloppy about using databases and using infrastructure. We nicknamed them as teenagers, because they almost refused to clean up after themselves. It's not even that they refuse, it's actually that the agents would have to spend extra time and tokens doing this work. Most people, if they're doing some kind of vibe coding, you often have a fan out strategy for agents. You pick 10 agents and say, go try and solve this problem in 10 different ways, and then I'll have a supervising agent decide who has the right solution. If you imagine 10 different databases, it's spun up in isolation. Each of these agents is trying to work on the problem and solve it in a different way. Then at the end, only one is chosen as the winner.
Just like a teenager, you could ask them all to clean up their rooms after the fact, but if that's going to cost you extra time and extra tokens, that's more money spent on a thing that you don't actually care about.