Our impact outweighs --- too many development hurdles means our impact happens before AI could be existential

BROOKS: People always want us to fight, but we don't really. I think, although I agree with the general themes that Andy talks about, I think it's very easy for people who are not deep in the technology itself to make generalizations, which may be a little dangerous. And we've certainly seen that recently with Elon Musk, Bill Gates, Stephen Hawking, all saying AI is just taking off and it's going to take over the world very quickly. And the thing that they share is none of them work in this technological field. So let me explain why—and they're all smart people, but I think they're making a fundamental error and it gets to NEIL, actually. THOMPSON: After taking down Bill Gates, Elon Musk and Stephen Hawking, he's going to take down the Dalai Lama. Please continue, Rodney. BROOKS: So let's go back to an example from the '90s, when IBM's Deep Blue beat Kasparov, beat the world chess champion. And Kasparov got up and said, well, at least it didn't enjoy beating me. That was his—holding on to his humanity. And now, today, you can get programs that run on—and that was on a supercomputer and now you can get programs that run on laptops. There's about twelve of them that have a better chess rating than any human being has ever had. So people see that -- MCAFEE: It's so bad now—let me underscore what Rod is saying. It's so bad now that they asked human grand master a couple years ago how he would prepare for a match against a computer and he said, I'd bring a hammer. BROOKS: So they can play chess really, really well. And I think people generalize that in the way that if a person can do some task really, really well, they can do adjacent tasks quite well. But none of those chess programs can play tic-tac-toe. Imagine a chess grand master who couldn't play tic-tac-toe. It doesn't make sense. None of those chess programs can give advice to an aspiring human on how to play better. All they can be is a sparring partner. MCAFEE: That program couldn't play tic-tac-toe without being substantially redirected, right? BROOKS: Right. So people, I think, are seeing some of the image labeling that's going on, for instance. Google came out with image labeling, which is a great commercial problem for them. They want to be able to label images. And one of the examples was, that Jeff Hinton shows, one of the chief scientists, is, it's a picture and it says there's a baby holding a teddy bear or doll in there. You look at it, it's a baby holding a teddy—a doll. But then if you ask the program, where is the baby? All it can say is, well, this pixel has 10 percent probability of being a baby, this pixel has 80 percent. And people have done experiments. You have a mashup of, you know, a grotesque mashup of baby parts and it says it's a baby. It's a baby. It's got all the parts. But a person says, no, that's a grotesque mashup of baby parts in the image. THOMPSON: But Abhinav, you've solved this, right? GUPTA: No, no, no. BROOKS: Well, he's working towards-- GUPTA: So, can I-- BROOKS: He's working towards it because it's such a hard problem. GUPTA: Yeah, OK. Thank you. So since we are talking about images, I think I should chime in a little bit and tell you that—so what Andrew is talking about, that we have made big advances, again, they're very, very specific tasks. Given an image, tell me what label can you put on that image? We have gotten really good at this task. Some people claim even better than humans. I don't buy that, but let's assume that even better than humans. But that doesn't mean we can do anything else apart from that exact task. And that's what Rodney was talking about. They have no idea that—what does baby mean? What does having a baby mean in those images? No idea. You, as a human, would know, OK, if I'm saying there's a baby, it has a lot of meaning inside. You get a lot of meaning out of that thing. So while we have made significant advances in the last two years, I want to boil it down a little bit and say we still are a long, long way to go but Elon Musk or Bill Gates, everyone is talking about, we still have a long way to go. But there's hope, and that's what I think we have to see here. Two years ago, if you asked me can computers take an image and solve this problem, I would say I have invested seven years of my life but if you give me a random image, it will not work. And now, given a random image, it will work. So all it -- MCAFEE: This was the guy who was doing this for a living and if you asked him two years ago would this happen, he would say, no. This progress is weirdly fast and is surprising insiders in the field. GUPTA: Yeah. I agree. And I'm an insider in the field and I'm very surprised, I have to say. Now, at this point of time, I am like living in an awe of myself, in some sense, that—but it was like Rodney told—Rodney told us that in thirty-five years he never thought he'd do this all. I also thought some of it, like that, for thirty-five years, I will not solve—see this kind of classification performance. But as I'm saying this, still a long, long way to go and much harder way to go. What—where all the kind of gains have come from is the data. And I think—so technology, this like deep-learning technology has been there from '70s and '80s. Don't misunderstand that this technology came two years ago and everything's changed. This has been there for thirty years. It's just that for the first time in our technological advances we have data for this deep-learning technology to learn. MCAFEE: Let me jump on this because I think the three of us are really agreeing, instead of disagreeing. I chose my adjective pretty carefully. I said these advances are going to be economically significant. I completely agree with Rod that they're not going to be existentially significant on any timeframe that we really need to worry about, for exactly the reasons that you're bringing up. One way to think about this, the way I try to get my mind around it, is there are, from what I've been able to take in, there are something like between ten and twenty really fundamental challenges that these guys and their discipline have been working on. Common sense is a really great example of that. As I've looked around, these breakthroughs that we're seeing seem to be—kind of indicate that we're making real progress on one of those challenges, the challenges of learning in a pretty unstructured environment. That's a big deal. There are lots of other fundamental challenges in the discipline where the progress has not been as fast, and these are the ones that you're working on.