Tech Brew Ride Home - (Bonus) Building AI We Can Trust With Gary Marcus

Episode Date: September 15, 2019

As I mentioned in the Weekend Longreads segment, Gary Marcus and Ernest Davis had an op ed in the times that has changed the way I think about the state of AI. But they’re also the authors of a great new book, Rebooting AI: Building Artificial Intelligence We Can Trust. There’s a reason people remain fearful about AI… it hasn’t earned our trust yet. In all sorts of ways that we get into on this episode. And also, the interesting ways AI development needs to change to take the state of the art to the next level. Sponsor: MintMobile.com/ride Learn more about your ad choices. Visit megaphone.fm/adchoices

On April 4th, 2023, around 2 in the morning, a man was found stabbed multiple times on a sidewalk in downtown San Francisco. Hey, who did this to you? What happened next turned the story into a political firestorm. Reports have identified the victim as Bob Lee, the founder of Cash App. From Bloomberg Podcasts, this is Foundering, the Killing of Bob Lee, beginning April 16. Welcome to another weekend bonus episode of the Tech Meme Right Home. I'm Brian McCullough. As I mentioned in the weekend long reads segment on Friday, Gary Marcus and Ernest Davis had an op-ed in the times that has changed the way I think about the state of AI right now. They're both also the authors of a great new book, Rebooting AI, building artificial intelligence we can trust. There's a reason, they argue, that people remain fearful of AI. It hasn't earned. our trust yet. In all sorts of ways that we get into in this episode, AI needs to change. So a lot of this episode is discussing the interesting ways that AI does need to evolve in order to take

the state of the art to the next level. So as we talked about off air, I recommended your op-ed that was in the Times this past week on the weekend, long-read suggestions. Um, you, Basically, you start out by saying that AI has a trust problem that we're relying on it more and more, but generally the public doesn't have a lot of confidence in it yet. And I thought that that was an interesting way to frame a discussion about AI skepticism. So first of all, what actually said, to be clear, we, Ernie Davis and I, what we actually said, we didn't say that the public hasn't earned it yet. Well, go on, continue. So talk about why that is and why humans haven't been convinced by AI and why AI has yet to earn our trust. So I think humans sometimes are convinced by somebody trusted at Tesla to drive them.

The Tesla worked pretty well under ordinary circumstances, but the weakness of current AI is that it doesn't work that well when circumstances aren't a semi-taking a left onto a highway, which is encountered something that was unusual, didn't know what to do with it. at 20 or 50 hours and it's been fine. Good at dealing with unusual circumstances. And it actually sometimes gives AI. A lot of people kind of treat AI as if it's magic. And if you actually work on, that there are actually lots of limitations.

The systems are not good at things that are outside the set of examples that they've been trained on. I was going to say, and let me just stipulate that I'm a complete layman when it comes to this stuff. But my sense of it and my skepticism around it is, is like, okay, AI has, we've proven that you give us a hundred thousand, given AI, 100,000 photos, it can maybe pick out a face out of that 100,000 photos in a few seconds.

Although there's, sometimes it's not good at doing that as well, but that's another conversation. But my point being that we've always understood that computers can do wrote stuff quickly, faster than we can, you know, finding the square root of a large number. But what's missing is that, and you're talking about the self-driving cars as the example, It AI hasn't hasn't demonstrated any judgment yet, right?

It's a funny question to ask. I think that's right. AI doesn't really have judgment. It can do decision-making. A machine can decide all day long. Is there more money in this account or less money? Do you have a simple decision that's something?

But they don't really have an understanding of the world so they can't offset in considerations that are kind of complex. There's no machine that can do that. Or even, okay, using again, the Tesla example, So, you know, we're training these autonomous driving AIs to – it's just pattern matching. Is that a baby carriage? Is that semi-switching lanes?

So if it accumulates enough knowledge and enough examples, it should be a fine 99% of the time. But if it encounters something, it's never encountered before, at least as of this point, it can't intuit a solution to the problem on its own. That's right. In that sense, there's examples that they have labels for. So they have seen something that looks like this scene before and what worked before was to go lab. If you have a scene that isn't like the system, it doesn't really understand what people are trying to accomplish when they're driving or, you know, that semis might have different, have any depth of understanding. The way we put it in the book is that what we have right now is a technique called deep learning.

And deep is really a misnomer. But there's no, there's deep learning, but there's not deep understanding. There's no comprehension in these systems. They can memorize the statistics of the English language and make up streams of words that sound like English, but those words of English don't actually mean anything because there's no understanding of moving to another and there being a conflict and attention and attention being resolved. They don't have that level of comprehension that we would expect of high school students. Right. And that's another interesting point that you made. I don't remember if it was in this piece or the wired one, that computers can't actually read in the sense that we understand reading. Again, we think they can because Google indexes the entire internet and can scan all the words and suss out relational context. But it's not actually reading these pages that it's putting into its index.

It's not understanding what is written. And like you said, even a four-year-old or five-year-old can read far better and comprehend far better than any computer can right now. Sometimes I like to distinguish between text processing, which is like matching keywords. That's what Google does. So you can find all the pages that have these words. You don't have to understand those pages. You just have to do some matching, some kind of basic pattern recognition. But to do what, let's say, my five-year-old does when she reads a story,

is to reconstruct in your mind what is happening and to be able to ask questions. Like, I wonder what would happen if the character did this. And current systems can't do that. So the point of the wired piece, and we have a longer version of it in the book rebooting AI, is to go through a single example, which is reading a children's story by Laura Ingalls' Wack fame. We just go through a few paragraphs of it. We show all the inferences that people have to make in order to understand what's going on, because most interesting writing doesn't spell out everything.

It assumes you can figure things out. So if somebody reaches their back pocket for their wallet, you know that that's because the wallet is not. There's all this background knowledge we have as human beings, and we use it to parse the things that we're reading. And the machines don't have it. They don't really have a lot of knowledge about how the world work. And so the kind of comprehension that they get is really, really thin. It's physical matching, but they can. can't reconstruct in their mind, what you might call a cognitive model of what's going on.

And that greatly restricts their ability to do anything like reading. And people are like, oh, AI is, you know, magic is here. It's going to take all their jobs. They can't even read. Ultimately, AI can succeed and will be able to do things like read the medical literature, which nobody can keep up with anymore and invent new treatments and things. Right. And let me mention the book specifically since we haven't so far.

It's rebooting AI, building artificial intelligence we can trust. I get the sense, maybe I'm being too harsh or putting words in your mouth, but I get the sense that maybe you feel like current technology has gone down a bit of a cul-de-sac because, again, you said it just there that you want new approaches that are inspired by human cognition and psychology as opposed to these sort of, like, you know, pattern-matching, as we're describing. Do you feel that way that like maybe either... Maybe called a sec is too harsh, so maybe it's just we've pushed what we can do as far as they can go. I don't think it's too harsh, although I think that some folks like Jeff Hinton and

Yashua Banjo and Jan Lacoon, it's a really great hammer. But it's tempting when you have a hammer to think that every new technology of deep learning. So recognizing pictures to them automatically identifies that other kinds of things that people do that are not about something in this category like language. So every sentence is different. You can't put things in categories of sentences that mean the same thing because Every sentence is different and it's different in context and so forth. And it's not that useful.

It probably in the final way in which we put all this stuff together. But we need lots of tools. We need hybrid solutions that bring together the best of that approach to AI, which is very information and abstraction and so forth. So we actually need to bridge different. Right now, the data-driven one is clearly dominant. It's getting better results. But the answers don't all lie there.

A plane and a drill and so forth. And we need to invent these other tools too. You just, you said something that this is entirely an aside, but the books that I've read on AI development and the history of the field is it does feel like that you have all of these sort of warring camps. Like there's this discipline of doing it this way and this discipline of doing it that way. And either one at some point becomes more prominent over the others and then years down the road, flip-flops and things like that. Why am I right in intuiting that? Or why does it seem like everybody, you're either this type of AI or you're that type of AI. Why is it so siloed like that?

And they fought over grant money from the beginning. And we stridently believe their own views. I think there's not been a lot of work to synthesize the views. People are people defending their turf. This is true. People are the same thing as them in their academic. It's a kind of similar tension. And now there's a lot of money at stake.

And people are defending their turf. And it can be good for the individual. because they get some money, they get some recognition, it's not really good for the field. You know, ultimately, the right answer to AI has to be synthetic. It has to bring different things together because the problem it's trying to solve. Intelligence brings different things together.

There's perceptual classification, but there's also language and there's attention and there's reasoning. And we're going to need all these things to work together. And that's not necessarily in the interest of the individual players at the moment, but it is in the interest of them. So the solutions that you guys write about are essentially that synthesizes, a hybrid approach that pulls from all the different disciplines. But also, you talk about, like, building in some sort of, you know, basic background assumptions

into these machines. Like, you speak about that AI needs to understand time, space, and causality. Like, what are you describing there when you say we need to move beyond, like, these basic frameworks of numbers and things like that, to these other basic background understandings? We give a simple illustration of this, which is if you ask whether George Washington owned a computer, a search engine can't figure that out. Unless it happens to be a web page that it's actually given the answer, it's something you need to infer. It's not hard to infer. You can figure out, well, George Washington must have been alive in the 1700s, and you could look it up and figure out exactly when he died.

And if computers came out and you realize that's more like in the 1900s, and so, you know, so inside, there's no way that George, Washington could have owned a computer. But an AI system doesn't know what it needs to be alive, and so it can't do the temporal reasoning in order to analyze this. It doesn't really understand some causality. So let's make them. An average human being knows a lot about a lot of different things, and they can use that in the course of really have a way of incorporating that knowledge.

I think these are fairly challenging problems, and it's just trying to pinpoint the problems that need to be solved. Now, the reason that we wrote, I know the answers to all of them. some of them I think are going to hear exactly is where the problem. Moving to a slightly different topic, one thing that has come up on the show a lot recently is sort of the space race, as it were, for AI research between the West and China. I'm wondering if you could give us any sort of your take on that. There's all these fears that China is far ahead, that is pouring more money into it.

How do you see that playing out at the moment? I don't think China is far ahead yet, restricting visas for much less appealing for some of the best talent in the world to come here, which means certainly like China has a lot of resources and they see the value of doing this. The U.S. is finally starting to think about an AI policy, but, you know, it's a clear... Well, yeah, the real Cassandra's I've read are worried about this notion that is a really meaningful breakthrough in AI, they fear whoever may be. it would immediately be so far ahead that not only would it be like generationally ahead against other competing nation states or whatever, but also that that breakthrough would allow them to essentially always hold back everyone else, so that it's a zero-sum game in AI. Do you think that that would be true?

It actually makes me think of a great book called, I think, War Made New by Max Booth's technology, is that somebody does get a huge short-term advantage, but other people eventually hold that advantage. You could have something where somebody invented technology that they keep secret for 100 years that nobody else finds, or come up with a technology that nobody borrows, steals, or replicates for 100 years. I kind of doubt it. Is there any recent work, research that you're really excited about that maybe you feel like could be a huge breakthrough in the next five years, say? There are some technical things. that I think trends that I think are good. One is that MIT, for example, it's not stuff that's

going to immediately change the world, but it's pointing in the right direction, and it's good because it's overcoming this long history of people being in silos, and you have people like the leader, one of the leaders of deep learning, Jeff Hinton, and then similar models, and this is actually not, traditional deep learning models, you take something from a memory location and you do some operation on it. Computer programming works, and deep learning hasn't had any of that. Some people are trying to build some of that in the context. Kind of early days for that, but people didn't even want to touch that before. And I think there's a realization that you can't solve everything with massive amounts of data. You

can solve chess and go that way, but you can't solve reading that way. And so there's a new openness, I think, just in the last year or so, to trying these things. I wrote a paper 2018 in January for free. And I think... All right. Final question. I was... at a conference this past weekend where it was a tech conference. Every other startup was like, well, here's a new dating app that leverages AI to do X. We're a new company that uses machine learning to do whatever in oil and gas discovery. And it's like, you're like, is that really AI or do you just have a bigger database and a better algorithm? I'm just curious what your take is on

almost the buzz wordification of AI and ML for startups. Some of it's real and some of it isn't. And some of it's stuff that you could be doing with more simple statistical techniques like regression or simpler machine learning techniques like decision trees and so forth. Some of it's real. It's not, but it's a lot. Right now, I think a lot of people think, and the reality is AI does some things well and some not. It's just a set of techniques. And some of those techniques are better for some things than others, and it depends on the problem. So right now, if you have a massive amount of data and your situation that you're trying to,

to analyze is very stable, it works pretty well. But if your situation is complicated enough like driving, it doesn't work well enough. And it depends on what you're trying to do. If you're trying to recommend ads, the alternative is not having any kind of statistical analysis. Well, it's certainly better than that. If you were trying to build an elder care robot and it worked 95% of the time wasn't perfect, that might not be so good, right? You know, 5% error in an healthcare robot might be a dangerous thing. So it depends on the nature of the problem, how open-ended it is, the more open-ended, the less our current techniques are good at it, how much accuracy is required. So if you need just moderate accuracy, we might be able to get that now.

So you can think like phototagging. You have a lot of data, if you're Google, and you don't have to be. So that's a good application. On the other hand, if you're doing medical diagnosis, able to read people's charts, and they're not accurate enough. And so part of why Watson, I think, is not. Again, the book is rebooting AI, building artificial intelligence we can trust. Thank you so much, Gary.

Terrific, thanks for having me.