The People's AI: The Decentralized AI Podcast
00:11:51 · Jeff Wilser, Bayley Wang
[00:11:51] S0: one North Star for an addressable universe. Yeah. Yeah. Okay. So you're just like, alright. This is a fairly hyper niche thing. We can probably find a... Or hopefully find a a more scalable, better use case and and and better kind of, like, type of robot that does more than just, like, freaking backflips for the hyper wealthy. [00:12:16] S1: Yeah. Yeah. Yeah. So after... Right. After the after that, we, just sort of went off and did did our own things. I went and did a... Did, like, worked on a company doing displays and optics. But robotics has always been top of my mind and things that I'm interested in and I... Things that I think are really high impact. So, you know, a couple years ago, g... GPT three showed the world that large scale unsupervised training on huge datasets was incredibly powerful, and the resulting emergent behavior could do crazy things. So everyone everyone started thinking about robotics again and how the same types of software and same types of workflow could make robotics better. But when you come and do robotics, you immediately run into a data problem, like a a data problem, a scaling problem, where if you wanna train a model which handles text, you go download the whole Internet and you go do your training. When you wanna train a robotics model, step one is to collect data and then you like stop because it's really annoying to collect data. It requires hiring hundreds or thousands of people to do the data collection to really do anything meaningful.
00:13:21 · Jeff Wilser
[00:13:21] S0: Famously, Google had to... Like, when originally to try and get Google Maps to work. Right? It had to... And Google Street View, they had to get, like, cameras on cars and drive thousands or probably millions of miles around the world, in the physical world to get to see, okay, here is the real world. It's one it's one... Prior to that, Google was like, cool. We can just... On web searches, we will own this. But then you try and get the real world. It's a lot more physical infrastructure you have to to stand up to do that. And so that's relatively simple compared to the idea of streets and highways is a relatively simple thing, I'm guessing, compared to the messiness of your robot and kitchen example earlier. Imagining all of the granular... If if if a countertop is sloped a bit. Right? Or if there's winds, air friction, how fast the robot's arm moves. That sounds in a way much more complicated than certainly than, like, predicting the next word in a sentence and arguably more complicated than cool. I'm on Main Street, and next to me is 1st Street, and here's how the... Here's what the the the neighborhood looks like. I mean, I warm as far as my idea of how complicated this is in a physical sense and where I was to struggle to figure out how do you, hoover up enough data to be able to make meaningful predictions.
00:14:51 · Bayley Wang
[00:14:51] S1: Yeah. So robotics is a little more complicated in in terms of the sensor modalities that are really needed relative to some of the other tasks. But the way you think about it is this. Right? Like, say you wanted to learn how to make a make a French omelet. Omelet. You would probably go watch a video of how to make a French omelet. And because you have additional knowledge about how your fingers interact with the world and how to pick things up by just watching that video, you can learn how to complete the task. And I think with robotics, it's going to be a similar situation. We're going to have low level models that are trained for that particular hardware probably on data from simulations in order to learn how to, like, hold objects without crushing them and understand things like friction. But the higher level models, the ones that require a huge amount of data are the ones that then teach the robot how to complete tasks. Mhmm. It's it's... It it also circles back to this question of like, hey. Do you need force data to be able to train a robot? And my answer to this is probably not because when humans want to learn how to complete new tasks, you... I don't have to, like, pick up your arms and, carry you through the motion of making that omelet in order to teach you how to make an omelet. Right? Like, you're smart enough about the world from your interaction, like, basic interactions with it to learn how to do this task. And sure, for very complex tasks, like like like if I ask you to, like, complete a advanced manufacturing task, I would probably have to also, like, give you a few trials. But I think that parallel also draws the robotics where maybe you do large scale training on, like, visual data. Then for the final run, you go operate the robot to go interact with the environment a bit in order to learn a little more more about the the forces and the friction in the materials.
00:16:40 · Jeff Wilser
[00:16:40] S0: So tell us about what brought you to and what kinda led to PrismaX your and light bulb of, okay, given the challenges of gathering data in the real world, we need some other solution.
00:16:56 · Bayley Wang
[00:16:56] S1: Yeah. So, fundamentally, what inspired PrismaX is this idea that diverse large datasets and diverse large operator, like, environments are what powers these models and makes them robust. If you if you... Like, a lot of the decentralized data projects are saying like, oh, we can provide data for cheaper. But for robotics, it's about providing the data that works. And I think that's a unique opportunity where the decentralized project has very strong potential to be the state of the art and, like, and beat out all of the centralized projects. Right? Because if if I'm a centralized project and even if I had the money to go hire a million people to collect data for me, I I still don't know who to hire. Like, I need to select the, like, split of million people that I think best represents the world. And I would argue that you will always wind up with, managers bias in a situation like this. It's it's like very subconscious in your mind that people will plan for the things that they're familiar with and the things that they find important. Right? And this is a problem. You wind up with a model and an AI and a robot which is tailored to the preferences of like a committee of 20, 30 people who maybe they tried their best to plan, but in the end it was not not successful. If you let the community involve organically, you wind up with a sampling of the world which much more closely corresponds to what people care about in the real world, and that gives you robustness. It's emergent behavior. Right? Decentralized networks have emergent behavior just like models trained on large scale data, And I I I think that's important. It gets rid of as many rules as possible and lets it lets it evolve according to like a very simple economic target of like this task has value or this task is one that people do more frequently. And that's what led that's what led to this idea that by using decentralization and by using protocols to coordinate millions of people, you could scale this... You could scale robotics up much more quickly than you would be able to with any sort of centralized project. And the rest of it just fits together.
00:19:00 · Jeff Wilser
[00:19:00] S0: That makes complete sense in theory. I love that that framework. Talk about what it looks like in kind of... In reality or kinda be concrete. So it's... As I understand it, it's a robotics data DAO. What does it look like? What are some examples of how people are are feeding, data to the system and getting compensated for it?