Astrid Wilde: A Bright AI & Robotics Future

Brandon Beylo (0:00)

Astrid Wilde, this podcast has been a long time in the making. I appreciate you being flexible with me and having my schedule with two young kids. And I'm in my daughter's nursery right now because my office was in the basement, and now we're trying to finish the basement. So I had to move upstairs, I don't have my mic set up, but I wanted to get this conversation in the books because there's so much happening in AI and robotics. And the release of Astra, the latest GPT model, I just had this feeling like maybe I need to really talk to somebody that is at a much deeper level, working on this basically 24-7 and looking at things, and has an inside baseball look at it. And anytime I see your tweets, like I was saying earlier, I was like, anytime I see your tweets, I just feel like I'm not bullish enough on AI.

And so I think to kind of frame this conversation, let's start with who you are, what you do for a living, how you got to where you are. And I think your company right now is Instinct HQ or Inheritance HQ or something like that. So we can even touch on that and then dive into everywhere we want to go next.

Astrid Wilde (1:16)

Sure. Yeah. Well, I guess most of my time, my day-to-day life is people coordinator, slash owner operator at Inheritance AI.

We primarily, the most material aspect of our business is robot data. And so taking what we call the, forgive my jargon, but it's human priors, which is like all of the embedded information and knowledge that human has about actions or doing stuff in the worlds. How can we take that innate or learned behaviors from humans and move that into an embodied intelligence or which we now think robots or physical AI would be the bigger category. But I think that robots is like two different, entirely different design architectures. One kind of robot would be literally your dishwasher, which doesn't necessarily need to know much of anything, but it does a job and it does it really well. And I think that that's the perfect robot.

But there are so many, so many, so many cases where the task or the action is more complicated than just doing one thing with a deterministic set of rules. And we need more quote unquote intelligence baked in to be able to handle edge cases. And that world is the world in which we deal with. And so we're trying to take the things that humans do and package that up into a nice form such that it's easy to train on for these large AI companies that I'm sure you and the listeners are all familiar with.

Brandon Beylo (3:05)

Yeah. And I think I've seen, I mean, you've seen examples of that with companies that basically are paying people to wear cameras or to wear hand, gloves with sensors so that they can train robots. And then you've got the Nvidia simulators, right? Where they can run things.

And you mentioned it, you mentioned kind of an interesting point where there's things that are very easy for robots to do. And you can kind of think about it like auto manufacturing, like where you're seeing a lot of the initial robots being deployed. It's like, hey, you're an assembly line. Your only job is to do this thing. And then when it comes to something like folding laundry or cleaning the house, like that's a multivariate, super complex problem. So how do you even begin to like solve that and then create the environment and create the data and then train the model?

Astrid Wilde (3:57)

For solving the problem, the answer is we don't solve the problem.

The bitter lesson, so to speak, is that we don't actually have to solve our problems anymore. We just have to provide the model with enough data of high quality and varied enough.

And the magic of the attention mechanism or this miracle of deep learning kind of takes care of the rest. So when we're capturing data, sometimes the data that we're capturing, we would refer to as primitives, where let's say in tennis, for an individual player who's playing tennis, there's a lot going on at any particular point to keep track of. You have to know about the score, and that might determine your strategy. Moment to moment, you have to know where's the other person relative to this ball that I'm going to hit.

But all of the things that make up the play space that is tennis are very simple, and they involve standing on two feet, that involves swinging racket, forehand, backhand, above, below, account understanding the spin of the ball, because that's important for both when you make contact of where do you need to make contact with the ball, and where's the return target on the other side of the court. But those are the kind of learned behaviors, so to speak. The primitives are just moving around, tracking a ball, and swinging. And there are so many tasks that it's economically inefficient for us to capture the actual full task end-to-end, but where it's very easy for us to capture the primitives. And so sometimes we're just dealing with primitives, sometimes we're dealing with full tasks. And you've seen there's a whole ecosystem around both. And so in the capturing full tasks, like you said the example of maybe you give somebody a camera on their hat or tape to their forehead, or some other hacky solution, and they go about just doing their job. And the problem with that approach is people are very messy. So you might have somebody that's on an assembly line, and they just turn on the camera, and then there's an eight hour unedited segment of video that includes bathroom breaks, that includes people pulling up something obscene and watching it while they work, that includes them stopping to talk to their co-worker for five minutes. And these are not the behaviors that we care about that are generating economic value. And so there's a whole pipeline of people that comes after the people that capture that data to split this up into meaningful segments of where's the important information happening from when to when, what is that important information? And ultimately, that's the real grunt work of this that takes the really, really messy real world people data and turns that into something that we can then train on.