This content is locked. Please login or become a member.
One of the things that people often misperceive about AI is how different the deployment of AI is in a purely virtual environment versus something that touches the physical world. So something that is software versus real wear or bits versus atoms. And there’s several things that really make that a very significantly different thing.
One is: in the virtual world, we are blessed with a proliferation of data that we can train on. People produce massive amounts of text and images and stuff and put it on the web. And so we have this unbelievably large collection of training data — which, when you’re dealing with the physical world, you need to collect often by interacting with the physical world. So when you have a self-driving car and you want to train a self-driving car, it actually needs to drive on a real road. And we do not have gazillions of hours of cars driving on the road collecting data up until quite recently. So that’s number one.
Number two is the tightness of the feedback loop. When you’re doing stuff on, you know, in a virtual world, things often happen instantaneously or on very short turnaround times. You display an ad, people click on the ad. Or you play a virtual chess game, and you collect that, you know, that information in a very tight feedback loop. In the real world, we’re limited by the rate of speed that that the world acts in. So if you are in a self-driving car and you’re driving on Broadway in rush hour traffic at three miles an hour, you are going at three miles an hour. And if you are, as you are in my industry, doing an experiment on cells and the cells take two weeks to respond to your drug, you can AI them all day long, but they’re still going to take two weeks. And so you’re limited at the rate of which you can experiment with the world.
And then the third is you can do a heck of a lot more damage in the real world than you can do in the virtual world. I’m not saying you can’t do damage in the virtual world. Of course, you can. But when you’re in a self-driving car, you can kill people. When you’re putting a drug into people, you could kill a lot of people. And so you have to be careful in those environments in ways where you can be a little bit more loose in the virtual world.
And so all of that makes AI in the physical world, where the bits are touching atoms, something that just takes longer. And so when people say like, “Oh, why does drug discovery take so long?” It’s because there are these rate limiting constraints that are just there, and you can’t do much about them. And taking a different analogy, which is self-driving cars, people have been talking about self-driving cars for probably twenty years, maybe twenty five. And every time it’s like, “Oh, in two years, we’re going to have self-driving cars.” And now finally, for the first time, you have robo taxis that are actually self-sufficient and functional on the roads of a number of cities. Not every city, not every road, but in a number of cities. And so I think those are challenging, but at the same time, I think those are also some of our biggest opportunities. Because at the end of the day, we as human beings live in the physical world, for a large part of our existence. And so when we make AI that meaningfully intervenes and makes better, hopefully, the physical world, the potential for impact there is much, much larger.