Why Google failed to make GPT-3 + why Multimodal Agents are the path to AGI — with David Luan of Adept

**Alessio** (0:00)
Hey, everyone, welcome to the Latent Space Podcast. This is Alessio, partner and CTO and resident at Decibel Partners, and I'm joined by my co-host, Swix, founder of Small AI.

**Swyx** (0:08)
Hey, and today we have David Luan, CEO, co-founder of Adept in the studio, welcome.

**David Luan** (0:12)
Yeah, thanks for having me.

**Swyx** (0:14)
Been a while in the works, I've met you socially at one of those VC events, and you said that you were interested in coming on, and glad we finally were able to make this happen.

**David Luan** (0:21)
Yeah, happy to be part of it.

**Swyx** (0:23)
So we like to introduce the speaker, and then also just have you talk a little bit about what's not on your LinkedIn, what people should just generally know about you.
You started a company in college, which was the first sort of real-time video detection classification API that was Dextro, and that was your route to getting acquired into Axon, where you were director of AI. Then you were the 30th hire at OpenAI.

**David Luan** (0:47)
Yeah, 30, 35, something around there.

**Swyx** (0:49)
Something like that. VP Avenge for two and a half years, two years and a bit, briefly served as tech lead of large models at Google, and then in 2022, started Adept.
So that's the brief CV. Is there anything else you want to fill in the blanks, or people should know more about?

**David Luan** (1:07)
I guess a broader story was I joined OpenAI fairly early, and I did that for about two and a half to three years, leading engineering there. It's really funny, I think, second or third day of my time at OpenAI, Greg and Ilya pulled me in a room and were like, you should take over our directs and we'll go mostly do IC work.
So that was fun, just coalescing a bunch of teams out of a couple of early initiatives that had already happened. The company, the Dota effort, was going pretty hard, and then more broadly trying to put bigger picture direction around what we were doing with basic research. So I spent a lot of time doing that, and then I led Google's LLM efforts, but also co-led Google Brain, was one of the brain leads more broadly.
There's been a couple of different eras of AI research. If we count everything before 2012 as prehistory, which people hate it when I say that, kind of had this like you and your three best friends write a research paper that changes the world period from like 2012 to 2017 I think the game changed in 2017, and most labs didn't realize it, but we had OpenAI really did. I think in large part helped by Ilya's constant beating of the drum that the world will be covered in data centers. It's cause of the need. Yeah.
I think we had conviction in that, but it wasn't until we started seeing results that it became clear that that was where we had to go. But also part of it as well was for OpenAI. When I first joined, I think one of the jobs that I had to do was, how do I tell a differentiated vision for who we were technically, compared to, hey, we're just smaller Google brain, or you work at OpenAI, if you live in SF and don't want to commute to Mountain View, or don't want to live in London, right? That's not enough to hang your technical identity as a company. What we really did was, I spent a lot of time pushing this, is just how do we get ourselves focused on a certain class of giant swings and bets? How do you flip the script from you just do bottom-up research to more about how do you leave some room for that, but really make it about what are the big scientific outcomes that you want to show and then you just solve them at all costs, whether or not you care about novelty and all that stuff. That became the dominant model for a couple of years. Then what's changed now is I think the number one driver of AI products over the next couple of years is going to be the deep co-design and co-evolution of product and users for feedback and actual technology.
I think labs every tool to go do that are going to do really well. That's a big part of why I started Adept.

**Alessio** (3:13)
You mentioned Dota, any memories thinking from the switch from RL to Transformers at the time and how the industry was evolving more in the LLM side and leaving behind some of the more agent simulation work?