David Luan: evaluation

Transcript context

…General question. So first of all, I think like the sort of input mode conversation. I wonder if you have any analogies that you like with self-driving, because I do think like there's a little bit of how the model should perceive the world. And you know, the primary split in self-driving is LiDAR versus camera. And I feel like most agent companies that I'm tracking are all moving towards camera approach, which is like the multimodal approach, you know, multimodal vision, very heavy vision, all the Fuyu stuff that you're doing. You're focusing on that, including charts and tables. And do you find that inspiration there from like the self-driving world? That's a good question. I think sometimes the most useful inspiration I've found from self-driving is the levels analogy. I think that's awesome. But I think that our number one goal is for agents not to look like self-driving. We want to minimize the chances that agents are sort of a thing that you just have to bang your head at for a long time to get to like two discontinuous milestones, which is basically what's happened in self-driving. We want to be living in a world where you have the data flywheel immediately, and that takes you all the way up to the top. But similarly, I mean, compared to self-driving, like two things that people really undervalue is like really easy to driving a car down highway 101 in a sunny day demo. That actually doesn't prove anything anymore. And I think the second thing is that as a non-self-driving expert, I think one of the things that we believe really strongly is that everyone undervalues the importance of really good sensors and actuators. And actually a lot of what's helped us get a lot of reliability is a really strong focus on actually why does the model not do this thing? And the non-trivial amount of time, the time the model doesn't actually do the thing is because if you're a wizard of ozzing it yourself, or if you have unreliable actuators, you can't do the thing. And so we've had to fix a lot of those problems. I was slightly surprised just because I do generally consider the way most that we see all around San Francisco as the most, I guess, real case of agents that we have in very material ways.…

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