[0:00:04] Hello. So cleverly, you'll notice that this panel is titled How to Build the Most Responsible AI Future, which begs the question of, is there a responsible AI future? So let me introduce our panelists. We've got Jack Clark from OpenAI. And I used to have a ton of fun introducing Jack by his former title, which was the world's only neural network reporter.
[0:00:27] And so I think now you're still one of a kind. You're a snowflake, don't worry. But it's now the world's only strategic and communications director of a massive nonprofit AI backed by Sam Altman and Elon Musk. It's easy to be number one if there's one of you. He did it. Doesn't roll off the tongue quite as well, but still a unique position. We've got Tim, who's led AI policy at Google for the last year. And among the many crazy, crazy awesome things he's done includes co-founding the awesome project, which...
[0:00:54] Give a No Strings Attached $1,000 grant to anyone who's trying to unapologetically promote awesome in the world, which is kind of cool. And so, can I have one of those? Sure. Yeah. So the reason for this panel, why am I so excited about it, why did we want to have it here? So there's, on the one hand, you've got...
[0:01:16] You're doing the bleeding edge research. On the other hand, you've got everyone else in society who's opining from sort of this sci-fi scared perspective of the future of AI. And you so rarely get this intersection between those two worlds. Yet we've got two of the biggest research institutions and the two people at those institutions that are supposed to interact with the rest of the world. And so we're going to pretend we're sitting around having a conversation here because we want to really get into it, not necessarily toe the party line.
[0:01:45] Tim, why don't you start and tell us a little bit about not just Google, because we know what that is. If anyone hasn't heard of Google, you can leave. But the eight options of Google that we should be thinking about. Sure, definitely. So Google's sort of an interesting beast because it has both kind of a research side and then also a product side. And those two sides actually live very closely with one another. So Google Brain, which is our kind of deep learning research group, works very closely with...
[0:02:12] Products to get basically basic research deployed out in the field very quickly. And it's a model that we think has worked really well and is actually a really exciting way of kind of like getting these insights and then quickly testing them to see actually if they play out the way we think in the wild. And I can intuit that AI has something to do with openness and artificial intelligence, but why do you exist? So we're in this honeymoon period of AI at the moment where there's an almost unparalleled level of openness. You know, Google publishes research.
[0:02:42] Microsoft does, all of these companies do, and we want that to continue. So the goal of OpenAI is to be a force to encourage everyone to be open by publishing our own research and producing open tools and helping to further pull the community in that direction because in the same way the computer security industry has benefited from the fundamental technologies being mostly open, and that's allowed us to build incredible things on top of it,
[0:03:09] We think that AI substrate needs to be similarly open, or you end up with strange vulnerabilities, strange problems, and these ivory towers of research which actually slows the rate of overall progress, which would be a shame. And so both of your positions are specifically to deal with the outside world, and so we hear a lot of far-fetched concerns, but...
[0:03:32] You guys are literally tasked with dealing with those. So what do you think? I promise not to use the S word on this panel, either of the S words, in fact. But what do you guys? I was going to use an S word. Were you going to use an S word? Yeah, me too. Sexism, that's a bad one. Safety, that's another S word. I was thinking super intelligence and singularity and all those things. So we've got a lot of S words that are rolling around on the panel here. But what are the practical concerns?
[0:03:55] What should we be thinking about? Sure, definitely. I can jump into that. And I think this is a fun part of the panel because it's how to build a responsible AI future, which begs the question of what responsible means. And at least for us, I think responsible is all about how do we get the benefits of this technology to be as broadly distributed as possible. And really, in my mind, that has three big pillars in how we're pursuing it at Google. The first one is open tools, which Jack has talked about. We're doing that through TensorFlow, which basically makes an entire
[0:04:25] Training Pipeline. I'm just curious. Who's heard of TensorFlow? Raise your hand. Oh, cool. No, you're doing your job. You should definitely check it out.
[0:04:33] The second thing I think is open publication, open research. There's been a norm, and some people may have talked about it today, is that the deep learning community was really small and suddenly there's a lot of attention on it. But there's a lot of good norms that were kind of cemented in the early kind of phase of the community, which we think are really beneficial to kind of pushing these benefits out to as many people as possible. And one of that is open publication.
[0:04:56] So we definitely try to do as much publication as we can in the wild, and we have a lot of researchers participating in things like NIPS and ICML and other conferences. And then I think the final one, which maybe kind of gets really close to your question of like, what are the issues we should be worried about? I think there's a real need to encourage sort of concrete, tangible machine learning research on kind of the social impact of AI.
[0:05:20] Two of the papers that I'll point out that we've been working on recently, one of them is taking this question of AI safety. And we try to reduce it to a problem which is very tangible, which is like, if you're building a cleaning robot, how do you make sure the cleaning robot doesn't destroy things in the room as it goes about cleaning? Which is actually a kind of profound CS problem and actually gets us to a point where we can frame these problems as engineering problems and develop solutions to them. And so there's a paper released called Concrete Problems on AI Safety in collaboration with OpenAI Allied.
[0:05:49] That kind of looks at that. Another issue that we've been really interested in recently is the question of fairness.
[0:05:55] So how do you detect that a system has gone awry because of biases that it's learned in the data that it's trained on? And for certain types of training models, it actually turns out to be kind of an interesting, challenging problem. And so we've got a new paper that came out called Equality of Opportunity in Machine Learning, which kind of proposes one way of looking at this problem and one way of sort of fixing classifiers so that they sort of don't sort of unduly bias or unfairly bias sensitive groups that might be in the system. Excellent. One quick thing I want to
[0:06:25] I want to mention, I'm curious to get your guys' take on it. Pretty much every other panel has talked about the importance of data and data sets. And that's the one thing you guys haven't talked about yet in terms of openness. You guys are promoting openness of all other things. Where does data come in? So I think for classifying stuff, data still matters a lot. If you're doing self-driving cars or speech or images, you need data and you need to augment the public data sets with large stores of your own. But for some of the newer territories like
[0:06:54] Creating robotic systems or agents that can play a variety of games and then transfer that learning into other domains and other bits of software. It's more about compute, and you're limited by your ability to simulate a world or to run a game for a certain number of cycles. So I think we're going to move from an era where data is the limit to where compute and your ability to wield large amounts of computers on quite complex problems is going to be what holds back research, at least in certain cutting-edge areas.
[0:07:23] And so how do we combat that? So data is something that you can make freely accessible if you choose to for a fairly cheap price, whereas compute is not cheap yet. Yeah, Amazon Web Services made $3.3 billion this quarter, I read today, so it does well. But the way that you can make it cheap is by actually creating tools that let people run machine learning in a large way on clouds, create large amounts of demand and lower prices, and that will also put some money into it.
[0:07:52] Which I'm sure he's pleased about. Well, I mean, I think APIs are a really powerful approach, right? As Frank brought up on the last panel, one of the really exciting things is that you don't necessarily have to build your own computer vision model anymore, right? Like the fact that you can pick one up means that you can immediately start investing all these other energies in finding a good product market fit, for instance. And I think that's actually a really interesting part of the ecosystem.
[0:08:14] But I think the other thing that's worth challenging, too, is that I think it's easy to get caught up or obsessed with data and the quantity of data, because I think it's easy to focus on. It feels tangible in some ways. But we actually, with TensorFlow, we're starting to hear all these really interesting cases about how a relatively small amount of data can actually get you a really long way. So there's this funny story that we got a few months back. There's this kid who is the son of a family cucumber farm in Japan.
[0:08:41] And he retired from his tech job and basically went back to help out his family. And it turns out that one common problem in cucumber farming, how many people are familiar with cucumber farming? This is a little different than TensorFlow. Oh, there you go. One big problem is that you have to sort cucumbers. And so his mom was basically spending 8, 10 hours a day during harvest hours sorting them manually. And so he basically said, this is ridiculous.
[0:09:02] Design a model that will recognize cucumbers and then sort them according to size and color and so on and so forth. And he basically trained it on 3,000, 4,000 photos that he took just manually with a digital camera taking photos of cucumbers. And he basically, with a Raspberry Pi, was able to hook up basically a rudimentary cucumber sorting robot. And that's actually a clear benefit on a small business being able to actually run with AI and actually see tangible benefits. It's funny, I've been calling this hipster AI. So I heard a story about a...
[0:09:32] Independent chocolate maker in Seattle. And I guess the secret to creating the best chocolate is just the best beans. Only pick the best beans. And he's just like, I'm a really smart computer scientist. Why am I going through these beans individually? I'm going to use TensorFlow to train this up to do that. So we've got more hipster AI. Like artisanal classifiers. Exactly. Artisanal AI. Back to Ariel's point about more artistic things. So I want to share with you guys my concern. And I want you to tell me if I'm crazy or not or what we should do about it. Which is the mechanism.
[0:10:01] That power AI, so it's basically data plus compute, really lend themselves to monopolies and oligopolies, I think, more than other technology trends, right? Matt Turk, who you just heard from on the last panel, has done some great writing on data network effects. And essentially, once you have that data, and once you have the interactions of that data going back to your users, you kind of have this...
[0:10:21] Virtuous cycle or vicious cycle, depending on how you want to think about it in our context. And are we going to live in a world where, and I'm not questioning Google's motives, but Google, Facebook, Microsoft, Amazon get bigger and bigger and bigger and there's nothing else that exists. That's currently the worry that the community faces. And the issue here is you end up with this kind of feudal system where you have companies which have, to some extent, locked themselves into a proprietary IP.
[0:10:51] And they're able to access amazing technology to classify, say, a cucumber. But as companies like Google are for-profit entities, they're not necessarily going to release the full model, and nor should they, because they have a fiduciary duty to their shareholders to not do that, because that's the IP they worked hard for. But I think as an AI community, we need to think about if there are ways we can get these large companies to collaborate or to create some general standards which truly allow switching.
[0:11:21] Because otherwise you end up with these locked costs which will slow down people. And I think the onus is on Google to work with their peers at Microsoft and Amazon to ensure that there's always going to be easy switching between them. Right. I think one response here is I think like things in technology are actually always much more competitive than you think they will be. Right. Like Google wasn't the first large tech company. And, you know, large companies routinely don't see a lot of opportunities or don't choose to chase after opportunities. There's like a known pattern.
[0:11:49] So I think some of the concern is maybe unfounded in that case. But at the same time, I also don't think that competitors are without any tools to go and compete. I think, for example, the investment infrastructure that we've all talked about today, VCs, can very well invest in valuable APIs that may be competitive. And I think we should expect to see that happen, particularly because the rents that you would get from being a company that has the only model of a certain type
[0:12:18] Would invite a lot of other people to come and compete. And I think we see that happening, actually. I think we see a lot of investment in those types of things. And sort of on the same point, one question. Do we have a long, short-term problem? And I don't mean that in a technical sense. So we keep hearing a lot about, you know, because the opportunity is so great here, all of, well, not all, but the vast majority of the professors have been
[0:12:38] Pulled out of academia, that next wave of AI talent is not going to go through the university system. And the question is, does that impact the long-term research of AI? Do we have a problem here? Well, we actually have, interestingly, a lot of examples of researchers leaving to go back into academia. And there's also a fair amount of circulation between the companies as well, which I think actually is a super healthy thing. It reminds me of the old analysis of why
[0:13:03] Boston's VC world fell behind while Silicon Valley stormed ahead. And one of them was that you had a lot of shared expertise constantly shared. And I think particularly in terms of research norms, that's one of the reasons we're such a big fan of open publication is that at least the knowledge is shared out. What are some bad things you guys have seen that you don't like that we should be fixing? In what? What bad things? Bad actions, bad actors, bad trends. What do you want to change? I think that...
[0:13:32] We need to be very careful about the data sets we're using. And if you don't inspect the underlying data, you'll end up reflecting biases in it. And this can cause huge problems in terms of, I mentioned sexism earlier, but every bad trait of society can be reflected in the data which you grab. And so you need to not be kind of indiscriminate in how you hoover it up. And we need to invest more as a community into that research and also do more.
[0:14:01] Outreach about it. Because if I have a popular photo app and it misclassifies me and I don't understand why, I'm going to lose trust in that company. And so the owners, people in the community to educate people about how these algorithms really work and make them less of a black box. I think there's two things. I mean, one of them is around challenges and interoperability. The worry is that we might launch systems that we can't fully investigate. And there's ways of kind of...
[0:14:28] Constructing systems so they have certain control windows. But there's a worry that maybe you don't deploy the systems as well as you might. So I think that's one big concern. Actually, I have an additional concern. I do too. Back and forth, back and forth. This is going to sound atrocious coming from me, but there's sort of a very severe gender equality issue in AI. There are way more men than women.
[0:14:54] There's two-thirds men on this panel, and I think that needs to change, and we need to figure out either funding mechanisms or ways to reach for pipeline earlier so that we have not just sort of just basically a sea of dudes is the phrase that some female researcher at Microsoft... That was an article you wrote. Yeah, yeah, but it was represented to me, and I'm on the other side now. I'm allowed to do recursive referencing. I love the title.
[0:15:20] It is a real problem, and we do need to fix that as well. The best analogy I heard here was just around a lot of medicine works well for white men between 18 and 35 because historically they've been doctors. And so that helped contextualize for me exactly what that meant in terms of shaping our software. So we've got to wrap up here. The one thing I wanted to ask you guys, we're going to get a little bit creative, is number one thing on your wish list. If you could just wave a wand and change anything, and this could include asking your comrade to do something.
[0:15:52] I'd like it if we could simulate reality perfectly. Most problems of how we unlock loads of value in AI are about how you get robots to do stuff which is useful. And Geordie Rose talked about this earlier, how it's a huge problem how you transfer from this sort of abstract world into the real world. And I truly think if we solve that, we can reap huge rewards as a society.
[0:16:21] Great. And I have one, which I guess is maybe more immediate, but I think would make a really big difference, which is I think it'd be great if we had good, accessible materials for people to understand the emerging concepts in deep learning. Because it's still a little bit of a special sauce. And I think both the public at large, but also from an educational perspective. Yeah, exactly. I mean, I think textbooks exist, but I think there's still a kind of missing gap, I think.
[0:16:47] In terms of really kind of articulating some of this stuff. There's a really cool project that we're starting to experiment. One of the researchers at Google, Chris Olah, is working on a project called Distill, Distill.pub, which is attempting to do this through visualization. And I think more of those efforts, I think, would be really great. Wonderful. Well, thank you both so much. And here's to the most responsible AI future. Thank you.