Galen Mead is building computer use base models at a research lab targeting AGI

Apr 23, 2026 · Full transcript · This transcript is auto-generated and may contain errors.

Featuring Galen Mead

me joins us on the show. And I believe Gayen is in the waiting room, so we will bring him in to the TVPN Ultradoo as soon as he's ready. Gayen, how are you doing?

Hey, doing well.

Reintroduce yourself and the company.

Uh, yeah, we build computer use models, a general research company.

Yeah.

Towards Alji. Um, I dropped out uh I think around three years ago. didn't spend much time in university and I just wanted to do a research full-time.

Did you jump straight to computer use or was there something in the intervening years?

We at the time that we were started there wasn't really this notion of a neolab.

So we from the start wanted to be a general research company.

Yeah. Um but initially we we started with some work on audio models that I had continued from SF compute.

Interesting.

Just because you know you train some state-of-the-art model and people see that you can train models and

uh you can go from there. Um I think this is maybe one of the

worst decisions um not not worst but worse uh

yeah sure

uh strategically but um computer use models are are nice because it's a general form factor. it always wanted to there's space for for taking actions.

Yeah. Why is computer use important? I think there was maybe like a oneweek period where everyone was super everything will be a CLI uh and there's maybe like a resurgence in computer use but uh what what what what uniquely captivates you about uh computer use as a uh as a as a goal here?

So it's a training recipe that I had always wanted to do. You can see in some of the early work on games from deep mind, there's this notion of training a policy on tons and tons of supervised data from diverse uh environments and getting something that is like a base model for actions and it it occurs to us that this exists for real world work in the form of like screen recordings. Yeah. with a computer being the universal like actuator that humans use to interact in very diverse environments. Um so there's this real opportunity to pre-train base models for being agents and this is a very appealing concept from a research perspective if you want to make very competent agents. Mhm.

Do you think uh how far out do you think we are from uh using like a truly complex piece of software like Premiere Pro or Cinema 4D or AutoCAD, any of these tools? It it it feels like we're on the cusp. Is that this year? Are you seeing glimmers of Okay, the average Premiere Pro user will probably be interacting with it through a prompt pretty quickly in the same way that uh software engineers migrated from the uh from the the you know IDE to the CLI and the text box pretty quickly.

Yeah, I think it'll be this year. Um I back in like middle school did freelance like animation.

Oh yeah. uh

with blended

and I had to sort of bring that out recently for our model launch. It was a bit over a month ago

um to put together Dharma videos and it honestly felt like stone ages compared to working with like the AI assisted code editors for my day-to-day job.

Yeah.

Um so I think we'll get there by the end of this year

and I think that will be a pretty big step change for all of these industries that aren't used to working with co-pilot. Yeah. Um, is there any uh is there any difficulty with the lack of opensource tools in other computer use like uh areas because it feels like part of the reason potentially for the acceleration in in coding is that the the the idees were open source and also the programming language is open source and like GitHub is this rich repository. Uh but if you're if will there be like this battle fighting between uh you know some proprietary piece of software that doesn't want to be used by an agent and then you're trying to figure out how you can do it. Have you have you grappled with any of that?

So I think we're somewhat unique in that we are rather obstinate about pitting the input and output of a human. Exactly.

Sure.

So we take video input, we output mouse deltas and like character level keystrokes. Mhm.

Um and yeah, nobody's copyrighted the form factor of a screen and a keyboard, right? So if we interact on that level, it's fully general.

Yeah. How do you want to uh instantiate this with a customer?

We're pretty agnostic on that right now. There's a lot of ways to go on the model capabilities.

Um

we can get pretty crisp signal and like there's a lot of things where

we should be able to tell the model to do something and it should do it.

Yeah. And once we're at that point, we'll

just download the app and start paying probably.

Little bit of a little bit of a tangent, but how have you processed the SAS apocalypse? There are some there are still some SAS bowls and they will tell you we're going to have uh we might even have more seats because we're going to have agents that are using computers and using existing uh software uh just like a human would. Sure.

But uh but how have you processed it overall? I mean, I I have one friends doing uh a search product who framed it nicely that there's an opportunity for a lot more usage if you have tons and tons more agents actively using these products.

Um I I haven't been following uh this sort of stuff very closely. I'm pretty off the internet.

Locked in. Love it.

Uh did we get how you made your first dollar? That's always

Yeah. Was that was that doing blending?

Freelance animation.

Freelance animation.

Basically art artist. What uh what what was the actual project?

Was it like marketing videos or

It it was a bunch of like forum graphics stuff. It was a very niche space.

Yeah.

That's cool. Well, thank you so much for taking the time to come chat with us.

Great to get the update. Come back on for your next uh big launch.

Yeah. Congrats on the progress. We'll talk to you soon.

Cheers.

Um back to the Cultured Magazine article. Um they asked us, "What grounds you and what invigorates you?" We said, "The daily schedule grounds us. The show goes live every day, every weekday at 11, so there's no room for projects to drift or spiral. Uh, booking a fascinating guest on the same day. Everyone is discovering them is incredibly energizing. That's the Sohon Periq moment in a nutshell.

They asked, "What would you wear to meet your greatest enemy?"

What' we say?

A TBPN racing jacket covered in logos. So, at least we're getting ad impressions for our sponsors.