David Luan: evaluation

Transcript context

…So I think Adept is the least understood company in the broader space of foundational models plus agents. So I'll give some color and I'll explain what it is and I'll explain also why it's actually pretty different from what people would have guessed. So the goal for Adept is we basically want to build an AI agent that can do, that can basically help humans do anything a human does on a computer. And so what that really means is we want this thing to be super good at turning natural language like goal specifications right into the correct set of end steps and then also have all the correct sensors and actuators to go get that thing done for you across any software tool that you already use. And so the end vision of this is effectively like I think in a couple of years everyone's going to have access to like an AI teammate that they can delegate arbitrary tasks to and then also be able to, you know, use it as a sounding board and just be way, way, way more productive. Right. And just changes the shape of every job from something where you're mostly doing execution to something where you're mostly actually doing like these core liberal arts skills of what should I be doing and why. Right. And I find this like really exciting and motivating because I think it's actually a pretty different vision for how AGI will play out. I think systems like Adept are the most likely systems to be proto-AGIs. But I think the ways in which we are really counterintuitive to everybody is that we've actually been really quiet because we are not a developer company. We don't sell APIs. We don't sell open source models. We also don't sell bottom up products. We're not a thing that you go and click and download the extension and like we want more users signing up for that thing. We're actually an enterprise company. So what we do is we work with a range of different companies, some like late stage multi-thousand people startups, some fortune 500s, et cetera. And what we do for them is we basically give them an out of the box solution where big complex workflows that their employees do every day could be delegated to the model. And so we look a little different from other companies in that in order to go build this full agent thing, the most important thing you got to get right is reliability. So initially zooming way back when, one of the first things that DEP did was we released this demo called Act One, right? Act One was like pretty cool. It's like kind of become a hello world thing for people to show agent demos by going to Redfin and asking to buy a house somewhere because like we did that in the original Act One demo and like showed that, showed like Google Sheets, all this other stuff. Over the last like year since that has come out, there's been a lot of really cool demos and you go play with them and you realize they work 60% of the time. like Google Sheets, all this other stuff. Over the last like year since that has come out, there's been a lot of really cool demos and you go play with them and you realize they work 60% of the time. But since we've always been focused on how do we build an amazing enterprise product, enterprises can't use anything that isn't in the nines of reliability. And so we've actually had to go down a slightly different tech tree than what you might find in the prompt engineering sort of plays in the agent space to get that reliability. And we've decided to prioritize reliability over all else. So like one of our use cases is crazy enough that it actually ends with a physical truck being sent to a place as the result of the agent workflow. And if you're like, if that works like 60% of the time, you're just blowing money and poor truck drivers going places. Interesting. One of the, our investment teams has this idea of services as software. I'm actually giving a talk at NVIDIA GTC about this, but basically software as a service, you're wrapping user productivity in software with agents and services as software is replacing things that, you know, you would ask somebody to do and the software just does it for you. When you think about these use cases, do the users still go in and look at the agent kind of like doing the things and can intervene or like are they totally removed from them? Like the truck thing is like, does the truck just show up or are there people in the middle checking in?…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.