The global conversation about robots and the workforce has shifted substantially in recent years, from widespread concerns about robots taking jobs to growing questions about how quickly they can fill gaps in the labor market.
One of the ventures at the forefront of this quest is Sanctuary AI. It’s a Vancouver, B.C.-based company that has raised more than $100 million Canadian dollars to pursue its vision for labor as a service. Sanctuary makes a 5-foot, 7-inch general-purpose humanoid robot called Phoenix, powered by an AI system called Carbon.
“The economic opportunity here, in terms of being able to both provide the labor that people want and being able to profit from it by developing a technology like this, is an opportunity unlike any that I think has ever existed,” says Geordie Rose, Sanctuary’s CEO and co-founder. “It’s way bigger than the steam engine, for example.”
Rose joins us on this episode of the GeekWire Podcast to talk about those long-term trends, the capabilities and limitations of Phoenix in its current form, the company’s overall approach, and the challenges ahead.
Listen below, and continue reading for edited excerpts.
Trends in population and the labor market: “The zeitgeist around automation has changed dramatically over the last five years. When we were starting on this journey of trying to build increasingly sophisticated machines for doing work, the dominant narrative was, ‘automation takes jobs.’ And somewhere along the way, that flipped to, ‘we need automation,’ because if we don’t have it, we won’t have anybody doing any of the things that we need done. …
“There’s this dramatic, alarming shortage of people to do the things that are necessary to keep our economy going. Right now there’s more than 10 million unfilled jobs in the United States. So the answer has to be technology. …
“In the future, all work will be done by machines. It’s just a matter of when, not if. This is clear. Robots get cheaper and better. AI gets better. At some point, there’s going to be an easy calculus: if I want something built or something done, why would I have a person do it? Because machines are going to be so much better at it.”
Specialized robots vs. general purpose robots: “What we decided to do was a different thing, which is, try to automate the person in the warehouse, or the whatever, by building a machine that can do all of the small things that you wouldn’t automate on your own but in sum total, makes it a very valuable thing to have.
“So it can walk around, it can press buttons, it can pull levers, it can operate other machines, it can collect the trash, it can do security, it can do all these different things, none of which you would actually build a special-purpose solution for. But now I’m doing 40 of them. And when I add up the value of all 40, this is a huge deal.
“We want to solve all of the in-between problems that these special-purpose machines are not suited for.”
The potential of new generative AI models for robots: “The output you get from these types of models tends to be very quick, but very unreliable. So in some ways, it’s the opposite regime of logical reasoning.
“Let’s say you’ve got an angel and a devil on your shoulder. The devil, the large language model, is this impulsive, brilliant mind that has imbibed the entire internet and is very quick to just spew out an answer. Whereas the other guy on your shoulder is the old sage, the guy who’s spent his entire life thinking about things, and is very deliberative, and has a lot of wisdom. And that takes a long time.
“So you’re sitting in the mouth of the cave, waiting for the declaration from the oracle for, I don’t know, a couple months, but when when it comes, then you know that you can trust it.
“These two ways of being are both complementary. Neither of them by themselves is a solution to the problem, but maybe they can be put together somehow. A lot of people are now working on trying to fuse these two ideas — that is, logic and reasoning and knowledge on one hand, and the statistical, next-token prediction, large language models … but so far, no one’s been able to do that.”
Subscribe to GeekWire in Apple Podcasts, Spotify, or wherever you listen.
Audio edited by Curt Milton.
