David Luan: recommendation

Transcript context

…Makes sense. I was going to say, I would love to talk about the interaction layer. So you train a model to see UI, but then there's the question of how do you actually act on the UI? I think there was some rumors about open app building agents that are kind of like, they manage the end point. So the whole computer, you're more at the browser level. I read in one of your papers, you have like a different representation, kind of like you don't just take the dome and act on it. You do a lot more stuff. How do you think about the best way the models will interact with the software and like how the development of products is going to change with that in mind as more and more of the work is done by agents instead of people? This is, there's so much surface area here and it's actually one of the things I'm really excited about. And it's funny because I've spent most of my time doing research stuff, but there's like a whole new ball game that I've been learning about and I find it really cool. So I would say the best analogy I have to why Adept is pursuing a path of being able to use your computer like a human, plus of course being able to call APIs and being able to call APIs is the easy part, like being able to use your computer like a human is a hard part. It's in the same way why people are excited about humanoid robotics, right? In a world where you had T equals infinity, right? You're probably going to have various different form factors that robots could just be in and like all the specialization. But the fact is that humans live in a human environment. So having a human robot lets you do things that humans do without changing everything along the way. It's the same thing for software, right? If you go itemize out the number of things you want to do on your computer for which every step has an API, those numbers of workflows add up pretty close to zero. And so then many points along the way, you need the ability to actually control your computer like a human. It also lets you learn from human usage of computers as a source of training data that you don't get if you have to somehow figure out how every particular step needs to be some particular custom private API thing. And so I think this is actually the most practical path. I think because it's the most practical path, I think a lot of success will come from going down this path. I kind of think about this early days of the agent interaction layer level is a little bit like, do you all remember Windows 3.1? Like those days? Okay, this might be, I might be, I might be too old for you guys on this. But back in the day, Windows 3.1, we had this transition period between pure command line, right? Being the default into this new world where the GUI is the default and then you drop into the command line for like programmer things, right? The old way was you booted your computer up, DOS booted, and then it would give you the C colon slash thing. And you typed Windows and you hit enter, and then you got put into Windows. And then the GUI kind of became a layer above the command line. The same thing is going to happen with agent interfaces is like today we'll be having the GUI is like the base layer. And then the agent just controls the current GUI layer plus APIs. And in the future, as more and more trust is built towards agents and more and more things can be done by agents, if more UIs for agents are actually generative in and of themselves, then that just becomes a standard interaction layer. And if that becomes a standard interaction layer, what changes for software is that a lot of software is going to be either systems or record or like certain customized workflow execution engines. er. And if that becomes a standard interaction layer, what changes for software is that a lot of software is going to be either systems or record or like certain customized workflow execution engines. And a lot of how you actually do stuff will be controlled at the agent layer.…

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