Vicarious is on a journey to achieve what some call the ‘Holy Grail’ of artificial intelligence: giving robots the capability to think, act and learn like human beings. Robots today can do many tasks, but they still cannot be as adaptable and flexible in different situations as people.

In this endeavor, Vicarious co-founder and CEO Scott Phoenix has the backing of some of the most famous people in technology: Facebook CEO Mark Zuckerberg, Amazon CEO Jeff Bezos and Tesla CEO Elon Musk, among other luminaries. In 2015, the World Economic Forum named Vicarious among its tech pioneers; it was also a Goldman Sachs pick the same year.

Phoenix recently spoke to Knowledge at Wharton about what’s ahead for AI. An edited transcript of the conversation follows.

Knowledge at Wharton: Vicarious is backed by some of the most famous names in tech, such as Mark Zuckerberg, Jeff Bezos, Elon Musk, Jerry Yang, Peter Thiel, and many others.

Scott Phoenix: Many of my personal heroes.

Knowledge at Wharton: These are really some big hitters supporting you. What gets them excited about Vicarious?

Phoenix: Well, I think that it’s getting harder and harder to argue that artificial intelligence is not the next big thing. I think that AI has already come to touch so many different parts of all of our lives. And I think that’s only going to get more significant as the future approaches.

Knowledge at Wharton: Let’s start by explaining what artificial intelligence is, before we get into more detail about what your company’s actually doing. What exactly is it, and what are some misconceptions people have about it?

Phoenix: The first thing to do is introduce a little bit of a distinction because there are a lot of misconceptions about artificial intelligence, and I think a lot of them stem from not making this distinction. So, I’d like to put artificial intelligence into two different bins. One is artificial narrow intelligence, and the second is artificial general intelligence.

Artificial narrow intelligence — you can think of it as a decrease in the price of prediction. Right now, AI is used to predict all kinds of things, like what ad to show you when you’re on Facebook or Twitter or Google, or what person might be in a photograph. So our advances in AI in the last two decades, and our advances in computer power and storage and data sets, have led us to a situation where those narrow predictions have gotten cheaper and cheaper and cheaper. And now it’s so cheap that you can have systems that predict what word you’re saying when you talk to your phone. And that’s what gives us Siri.

This also includes things like Google DeepMind, AlphaGo, Go player. That’s a narrow intelligence that predicts whether a Go board is a winning Go board or a losing Go board. And that’s all it does. It’s a very narrow intelligence.

Now, there’s a second kind of intelligence I would call “artificial general intelligence.” And that’s intelligence that, instead of just making predictions about a narrow data set, is designed so that if you give it the same kinds of experiences that a human has from birth to adulthood; it learns the same kinds of concepts and gets the same kinds of abilities that a human has.

Knowledge at Wharton: Basically, for narrow intelligence, it’s a much more limited functionality. Whereas for general intelligence, it functions more like the human brain, and it is adaptable and flexible.

Phoenix: Yes. And I wouldn’t say one is strictly better or worse than the other, in the sense that narrow intelligences can do all kinds of different things. Things that humans can’t do that well. Like, for example, most humans can’t play Go very well. But you can build a narrow intelligence to play Go better than any person. And you can build a narrow intelligence that’s better at regulating temperatures inside of a data center than any human could. And so that’s the advantage of a narrow intelligence: You can design it to be better than any human at tasks that humans aren’t necessarily that great at. Or even some tasks that humans are good at, like predicting who’s in a photograph.

But general intelligences are ones that are limited, in the sense that they focus on doing what humans already do well — controlling a body, understanding human-like concepts, and accomplishing long-term goals. … Our research focuses on artificial general intelligence. And specifically, how do you make a robot do all the kinds of things that a human can do?

Knowledge at Wharton: And why is that important?

Phoenix: It’s really important, because right now the world is full of cheap robot parts, like motors and sensors and electricity, and nobody owns any robots. You know, we live in Bizarro land, where all of the work that’s being done around the world is generally done by human hands.

Right now, there’s 800,000 humans assembling iPhones, not 800,000 robots. And that’s despite the fact that robots are physically capable of doing a lot of these kinds of manual tasks. A robot is perfectly capable of cooking a meal, for example. But no one uses robots to cook meals, because the artificial intelligence isn’t general enough to allow the robot to do that.

So, the reason why Vicarious is focused on artificial general intelligence for robots is that this is a technology that would enable us to live in a world that’s a lot more like the Jetsons’ world. It would let us make labor much more affordable, which would then let us all rise in society.

“The world is full of cheap robot parts … and nobody owns any robots.”

Knowledge at Wharton: And how exactly do you do that? I’m sure that maybe the Holy Grail for AI is human-level intelligence. What strategy or technology or algorithms are you using to accomplish that?

Phoenix: You can think of what we’re doing as taking what we’ve learned and what neuroscience has learned about the brain in the last 30 years, and applying that to building models that work much more like the brain than contemporary artificial intelligence approaches. There’s some brain inspiration in today’s AI algorithms, but not very much, compared to how much structure we know to be inside the brain.

Knowledge at Wharton: You and Dileep George, your co-founder, have mentioned in the past that there’s something called “the old brain” and something called “the new brain.” And Vicarious is focused on the new brain. Can you explain that a little bit more?

Phoenix: You can think of the old brain as a lot like these narrow intelligences. You have frogs that have super-human ability to catch flies. They can catch flies far better than I can. But they’re not generally intelligent. And you can think of AlphaGo almost like a frog that [only plays] Go or something. It’s not smarter than that. And it’s following a very narrow set of programming. Whereas you and I are capable of learning all kinds of new skills on the fly, and adapting our behavior.

Knowledge at Wharton: I understand that because of the artificial general intelligence that Vicarious is developing, you’re able to solve some pretty tough programs that narrow AI has not been able to solve, such as correctly identifying CAPTCHAs (Completely Automated Public Turing test to tell Computers and Humans Apart). Can you talk about that a little bit more?

Phoenix: We published some research in the journal Science this year about how we were able to use our more general vision systems to understand the contents of text-based CAPTCHAs, and circumvent them. And so that was one of the research papers that we recently published. And really, what it’s all about is not necessarily solving CAPTCHAs. It’s about showing that we can build a vision system that has the same kinds of properties that the human neocortex has, and apply that not just to CAPTCHAs, but also to a wide range of different problems, such as recognizing objects in an industrial environment on a robot.

Knowledge at Wharton: One of the examples that you like to highlight was that a regular AI could be fooled by such things as making a screen brighter, and suddenly they get confused. Whereas the kind of AI you’re developing will actually see through that, and continue working.

Phoenix: One of the misconceptions, I think, that I see regularly applied to AIs of today is that when people see a system do something that is intelligent in its behavior, like play Atari or play Go, they assume that it’s intelligent in the way a human is intelligent. So, they assume that that system that can play Go can also play checkers, or play chess. And really, in order to go from playing Go to playing chess … it needs to be completely retrained from scratch.

… The human equivalence is about three to five thousand years of continuous, 24/7 playing of chess, in order to learn how to play chess. And after learning how to play chess, it can’t play Go any more. So it’s a very, very narrow intelligence. And a common misconception I see is that people will read a story in the news, and they assume that means that the AI is really smart. And it doesn’t really mean that. It just means that the AI is able to do a narrow task in a very particular narrow circumstance.

Knowledge at Wharton: What are the practical applications of your technology?

Phoenix: Right now, in order to get a robot to do something you have to program it using a series of joint configurations — basically, tell it to move its arm through a sequence of way points. And as you can imagine, that’s very error-prone. If I asked you to make a sandwich, but instead of just asking you to make a sandwich using bread and peanut butter and jelly, I told you where to move your arms in 3D space as a list of commands, it probably wouldn’t work very well, and it would probably take a long time.

And if someone bumps the bread a couple inches to the left, all of those joint configurations that I gave to you to follow are going to now be misaligned, and you’ll be putting peanut butter on the table. And so it’s those kinds of problems that you really need a more general AI in order to solve. And that’s what we’re working on solving at Vicarious.

Knowledge at Wharton: What products do you see coming out of your technology?

“This is a technology that would enable us to live in a world that’s a lot more like the Jetsons’ world.”

Phoenix: Our first product is an intelligence layer for robots that helps robots to do tasks that currently humans have to do inside of warehouses and factories. So it’s things like assembly and packaging that robots for a long time have been physically capable of doing, but they’ve been too expensive or difficult to program to do.

Knowledge at Wharton: Are you primarily targeting the enterprise sector, or the consumer sector, or both?

Phoenix: We’re focused entirely on the enterprise sector for now. And I think that’s the way a new technology like this needs to be adopted. The nice thing about a factory or a warehouse is that it’s very consistent. There’s a lot of regularities that we can take advantage of, in terms of the kinds of objects that are being manipulated, and the types of operations that are done to those objects, and how well-controlled the environment is.

In a person’s home or an office or an unconstrained outdoor environment, there’s just a lot more variation. And so it’s a much tougher task for an AI system to solve. So I think the order will be first solving it for the enterprise, and then eventually solving it for the consumer.

Knowledge at Wharton: When are you going to market with your first product?

Phoenix: We should be doing some pilot tests this year. And then depending on how those go, we’ll look at moving to full-scale sometime after that.

Knowledge at Wharton: Tell me about the competition in this market. Are there many competitors?

Phoenix: There’s a collection of older robotics companies. One of them, for example, ABB, is the largest robotics company. It’s one of our investors. And so I don’t know that they’re competition so much as they’re a potential partner or a sales channel for us. And then there are also companies that are startups thinking about the space, that are, generally-speaking, a lot smaller than Vicarious, and haven’t been working on it for quite as long.

And then lastly, there are large companies like Google or Apple or something. And I think of those as being less competitors insofar as their primary business isn’t really working on building an intelligence layer for industrial robots. It’s much more about either consumer products or web services, or things of that nature. And so they’re not really exactly traditional competitors, in that sense.

Knowledge at Wharton: How much money have you raised?

Phoenix:  We’ve raised about $130 million so far that we’ve announced.

Knowledge at Wharton: And that puts you at what valuation?

Phoenix: We haven’t announced a valuation.

Knowledge at Wharton: The who’s who among your list of investors — what kind of impact have they had on the company? And are they giving you strategic advice?

“Going from there to robots that can take over the world … is a pretty big leap.”

Phoenix: I’ve had great conversations with all of our investors so far. And like I said, they’re some of my personal heroes. So it’s been great to learn from them and to hear some of the war stories of their path through building their companies. They’re people who I think all of us respect a great deal, and that’s been a really rewarding part of building Vicarious, to have a chance to get some advice and some expertise from these really incredible people.

Knowledge at Wharton: Elon Musk has spoken out about the dangers of AI, and how society could perhaps lose control of these self-learning robots and systems. Where do you stand on this issue?

Phoenix: Elon is a really long-term thinker. For him, a reasonable near-term goal is to build a permanent human colony on Mars. And so when you have that long term of a perspective, I think that a lot of his comments start to make sense. But for the average person who might be listening to this podcast or just following the news, I don’t think they should be spending a lot of their time thinking about whether the robots are going to take over.

There’s just too much work that needs to be done just to build the most basic of functionality. Like, right now, you can’t have a robot that makes you a sandwich. And going from there to robots that can take over the world somehow and outsmart the smartest humans, is a pretty big leap.

That’s where there’s a difference between Elon’s perspective and the average person’s perspective. And I think that everyone who listens to what Elon has to say about this should really keep that in mind.

Knowledge at Wharton: What about AI’s impact on jobs, though?

Phoenix: AI’s a really interesting technology, in that people, I think, have been more concerned now than they have been about past technologies. But if you zoom out and look at the broad lens of human history, in the last 3000 years of human development, what you see is that that is a 3000-year story of people building technology — taking a task that used to take 10 humans to do, and turning it into a task that takes two humans to do. Now, every technology from the wheel onwards is that kind of task.

And the result of 3000 years of automating is that there are more jobs now than there have ever been, and there are more people employed now than there have ever been. And so I think that it seems counter-intuitive to me to think that adding yet another automation technology is going to somehow drive out all the jobs.

It’s not a new fear. There have been so many points in history in the past, where new automation is introduced, and then people have felt very anxious that it’s going to disrupt all the jobs, and that no one’s going to be employed anymore. I think that is a very pessimistic view of human ingenuity, and of where our society could go. Because I think that we’re so far from living in a Jetsons-like world. And to get us from here to there is going to require a lot of hard work by a lot of people, and all kinds of new jobs that don’t even exist today.

The U.S. Department of Labor projects that something like 60% of jobs that our children will have don’t even exist yet. And if I were to tell you 20 years ago that I’m a social media expert, it’d be very unclear what I was even meaning. Or an Uber driver. And so I think that we should keep that in mind when we worry about what’s going to happen from an employment perspective with new technologies.

“It was really important for us to be building this technology in a way that disseminates its benefits to humanity broadly.”

Knowledge at Wharton: Is it true that Vicarious is a flexible purpose corporation?

Phoenix: Yes. When we incorporated ourselves, it was really important for us to be building this technology in a way that disseminates its benefits to humanity broadly. And so we set ourselves up as a social purpose corporation, in order to ensure that we were held accountable by our board, and by our shareholders, to achieve that purpose.

Knowledge at Wharton: Does that mean that you are a nonprofit?

Phoenix: We’re still a for-profit company. But the difference between being a for-profit company and being a social purpose corporation is that at a for-profit company, the only thing that you’re held accountable for is to maximize shareholder wealth. And you can see that go wrong in a lot of cases, where companies will take shortcuts on the environment, or on their community or their obligations to their employees. And the world will be worse off for it, in exchange for a temporary lift in their stock price that only really benefits the executive suite.

More companies should be social purpose corporations, because if they were, they would take into account a more diverse group of stakeholders, and the world would be better for it.

Mastering Innovation: From Idea to Value Creation
Become the catalyst for company-wide change when you learn how to construct the architecture that drives innovation in an organization.
Learn more.