#131 Andrew Ng: Exploring Artificial Intelligence’s Potential & Threats

0:00Do AI systems actually understand the world or are they call them stochastic parrots? I think AI systems do understand the world and they're also kind of getting better at explaining what they understand and what they don't understand. In fact, you know, there's this term stochastic parrot, this is just parroting words, mimicking intelligence. But there was actually one study that really influenced my thinking called Othello GPT, in which a model was trained to predict the next move on a game of Othello. When we think about existential risks, I think there are existential risks to humanity. I think high on the list would be, you know, maybe the next pandemic, fingers crossed.

0:43Or I think that global climate change is a risk to certainly a large fraction of humanity. What an asteroid did to the dinosaurs, that was maybe tens of millions of years ago. So it's less urgent risk, probably not going to happen in our lifetimes, but who knows. Very unlikely to happen in our lifetimes, I think. But to me, when you look at the actual things that could be an extinction of the rest of humanity, I think AI would be part of the solution. Hi, I'm Craig Smith, and this is Eye on AI. In this episode, I'm going to talk to Andrew Ng, a well-known researcher and pioneer in machine learning.

1:21Andrew has taught courses at Stanford and trained many of the AI luminaries working today. For a time, he was the chief scientist at Baidu, China's AI giant, and he co-founded Coursera, the online education company. Currently, he has an MLOps company called Landing AI, which has a suite of tools to make it easy for businesses to build, deploy, and manage machine learning models. I wanted to talk to Andrew because he's on the optimist side of the so-called threat debate, arguing that AI does not pose an imminent threat to humanity and that there's plenty of time to work out safety protocols, guardrails, and international agreements to prevent catastrophe.

2:11Now here's Andrew. To begin with, do you believe that AI poses an existential threat even in the long term? No, I don't. I'm trying to keep an open mind, but I'm having a really hard time seeing how AI creates. I know Jeff Hinton and Yasha Banjo have both spoken about the extinction risk to humanity from AI. I spoke with both Yasha and Jeff and I have the deepest respect for them. I'm trying to follow the reasoning. I'm struggling myself to see how AI, which is a fantastic technology, making society much better off, I struggle to see how that creates any meaningful extinction risk for humanity.

2:55Yeah. And just for people that haven't been following the debate but see the headlines, there's no evidence, no scientific evidence to support the notion that AI poses an existential threat. Right? it's it's it's it's theory it's uh speculation yeah so the you know i feel like um the two arguments about ai potentially causing an existential rest of humanity is the two arguments i heard is one is um what if ai is so powerful you know not gpt4 but gvt 25 or something um allows some really angry person or maybe someone that you know to create a bioweapon that then causes humanity to go extinct i think the risk of that seems small because it's actually pretty difficult to wipe out humanity no matter how angry one person may be i do think this is a risk of you know meaningful harm someone could create a weapon maybe but extinction is a scale of risk that seems very unlikely, very implausible to me.

4:09And the second argument that I've heard is sometimes it's been called the paperclip argument, not to diminish it with that terminology, but what if someone misspecifies an AI objective, like build a profitable company, but it figures out the way to make a lot of money, accidentally wipes out humanity. I think if an AI is smart enough to do that, it will probably be smart enough to understand our intentions uh and you know so many people work on ai safety so ai smart enough to accidentally wipe us out would probably be smart enough to know if we said make a lot of paper clips we did not mean make a lot of paper clips and along the way it's fine to wipe our humanity so that seems you know unlikely to me as well yeah is is there a way that uh lay people looking at this debate can differentiate between real risks and speculative fears?

5:07So, I don't know. It's a good question. It's challenging. So AI, just to acknowledge it, does have short-term harms today. There have been documented cases of bias, unfairness. You know, we've seen that poorly programmed self-driving car systems can lead to car accidents, these kinetic events that lead to human death right now. So I feel like there are those harms that AI researchers are working to address. One nice thing about large language models, like CHATGPT bought various models, is I do see many researchers working to make them safer. So they're getting safer every month right now. So I feel good about that.

5:45You know, I sometimes think about the aviation industry. With the rise of aviation, airplanes were really dangerous. They crashed, they killed people. Completely tragic, no excuse for really airplane crashes. Having said that, humanity, we learn from the airplane crashes and over time, aircraft got much safer. Today I think there's fear about some aspects of AI because people say you can never control a large language model. And you know what else you can't control? You can't control an airplane either. You can't perfectly control where an airplane points because winds will blow it around. And sadly, even today, even in society today, there are occasional airplane crashes which are very tragic and leads to really catastrophic loss of human life.

6:29Having said that, I think society seems much better off to me with airplanes than not. And even though we can't perfectly control an airplane, we can control them pretty well. So when I got in the airplane just a few days ago, I felt fine, you know, working on my laptop and not really worrying about a plane crash. And I think that we're on that trajectory as well, that large language models, even the last six months have gotten much safer, much harder even for someone malicious to make them do something bad, although I think it is still possible for someone malicious and determined. But I think that over time, the risk of either accidental or malicious harm, it seems to be going down rapidly.

7:10And I don't think we'll ever fully control them, but we can't fully control airplanes either. And I feel okay getting one and entrusting my life to one. Yeah. A couple of questions. One, And this concern about AI centers really on intelligence, how intelligent AI systems are becoming. And there's the debate that, well, you won't know when AI reaches superintelligence because it won't necessarily tell you. can we objectively measure the understanding of ai systems do you think so there's this debate about you know do ai systems actually understand the world or are they um sometimes call them stochastic paris i think ai systems do understand the world and they're also kind of getting better at explaining what they understand what they don't understand In fact, you know, there's this term stochastic parrot, this is just parroting words, mimicking intelligence.

8:23But there was actually one study that really influenced my thinking called Othello GPT, in which a model was trained to predict the next move on the game of Othello, the game of reverse Othello. And the authors of this study, which was published in the iClear conference, I think, demonstrated that the way to predict the next move in this game is to actually build a world model. And specifically, the author's found that if you probe the state of the neural network after being fed a set of moves, it seems to be learning an underlying representation of the status of the board. So it's not just mimicking surface level, parroting out what's the next move.

9:02It's actually figured out what the game of Othello is. So from a list of moves, it figures out what's the status the board and therefore what are the possible next moves so to me that was a fascinating study you know small scale it's just the game of othello but that really convinced me that i think today's large language models they are building a model of the world and um to me i believe they do understand the world um and also if if a large language model says something and you say no you know that isn't right? Or you added that math wrong in step five, the fact that you can say, you know what, I apologize my error, I did make a miscalculation in step three.

9:42And now let me fix that fee. And this is my new math result for some math puzzle. I think that analogies between machines and humans are always dangerous, you know, because but I feel like, I don't know that that's very different than how a child learning to solve math word problems for the first time they interpreted where so I feel like the language models are getting better at explaining to us, you know, how they're seeing the world. And just like when, when my daughter explains to me how she's thinking about the math problem, I believe her, you know, I think she's telling the truth. So I think when the large language model was telling us how it's reasoning about the world, for the most part, I know it doesn't always tell me the truth.

10:27But I think it says something close enough to the truth, it feels okay to believe it at least some of the time when it says that's how it's thinking about the problem. Yeah. So on the threat, your conversation with Jeff, sort of the takeaway was that the research community needs to come to some consensus about the level of the threat and how to counter the threat? Because right now with the divergent opinions or the split in the community, the public's confused and certainly policymakers, regulators are confused. So how can the research community reach a consensus on the threat? Yeah, more conversations.

11:23I feel like I'm continuing to have conversations with very different views, including those actively worried about extinction to those worried about other risks. Also, many people are just busy building valuable applications. but not just me. I think if everyone in the community can engage in, you know, polite, constructive discussion, I think that we can maybe slowly, probably will be slower than any of us want, gradually come to narrow a range of views. You know, I remember six, maybe seven, eight years ago, with the rise of deep learning, kind of the wave of deep learning was started maybe 10 years ago, a few individuals, including most notably Elon Musk's Bill Gates, warned of some of the risks of AI.

12:12And I think there was a spurt of worry at that time that then candidly died out. I think I heard much less about that after a while. I think Elon Musk said something like, with AI, we may be summoning the demon. I may be misquoting him, but I would respectfully submit that in the last decade, we summon a lot more angels than demons with AI. I mean, not to dismiss the harms, there have been harms, but net net AI has been a massive contributor to society. And so I think that wave of worries, frankly, died out. This new wave of worries with buy-in from prominent scientists with Jeff and Yasha talking about extinction risk, that was a surprise to me.

12:53But I think it'd be worthwhile to keep having conversations to see where we end up with this as well. I'm really trying to keep an open mind, even though I think I said I just don't get it. I don't see how AI could lead to human extinction. Yeah. Yeah. The lack of that consensus, is that a danger in itself that either regulators listen to the people who are pressing the existential threat narrative and overregulate, or they They listen to those who say, look, those threats are far, far in the future if they exist at all and under-regulate. Yeah. So one of the things Jeff Hinton and I said in that video that I posted on social media, Twitter and LinkedIn, is one of the reasons why climate scientists have been an effective force is because they are aligned, right?

13:56So there are economic interests to ignore climate science and just keep on generating carbon emissions. But the fact that climate scientists are more or less unified on the science of climate change, that's made it much harder for economic interests. They would find it more convenient to just not be regulated, to lobby regulators to ignore climate science. So I think while the AI community is splintered, it actually makes it unfortunately much easier for anyone with an agenda to lobby regulators to whichever argument is more convenient for the business. So the AI community is unified, then I think like climate scientists have, we collectively do a much better job, helping regulators unleash AI to treat all the value it can while also mitigating against realistic risks.

14:55And just to be clear, when we think about existential risks, I think there are existential risks to humanity. see. I think high on the list would be, you know, maybe the next pandemic, fingers crossed. Or I think that global climate change is a risk to certainly a large fraction of humanity. And, you know, what the asteroids did, what an asteroid did to the dinosaurs, that was maybe tens of millions of years ago. So it's less urgent risk, probably not going to happen in our lifetimes, but who knows, very unlikely happen in the lifetimes, I think. But to me, when you look at the actual things that could be an extinction of humanity, I think AI would be part of the solution.

15:35So AI, I don't think the world responded that well to the last pandemic, but I think AI enabled drug discovery and monitoring of disease conditions. That seems important to me. As climate change becomes worse, I think AI enabled solutions to mitigation, to the smart to accelerate the electrification of society. Or, candidly, I often think about climate geoengineering. Do we need solutions like using high-altitude aerosol sprays to slow down cooling of the planet? Or if not doing it now, because it's a dangerous technology, certainly to advance the science of climate geoengineering with AI-enabled modeling of the climate.

16:18So we better understand if this is even an option, we should seriously consider. So if we look at the real extinction risk to humankind, at least the ones I could foresee more clearly, AI seems like an important piece of the solution. So I would say if you want humanity to survive and thrive for the next thousand years, I would rather make AI go faster rather than slower. I want to give a shout out to this episode's sponsor, Masterworks, an art investing platform that buys contemporary masters outright. Works by Picasso and Vansky and Warhol and others, then qualify it with the SEC to offer it as an investment.

17:01Net proceeds from the sale are distributed to its investors. Last year, most investors lost money even with a diversified stock portfolio, and that's why some of the biggest money managers in the U.S. have been looking outside the stock market for assets with low correlation to traditional investments, things like fine art. The opportunity was exclusive to professional investors until now. Masterwork makes it available to ordinary investors, like me. Since the inception, they've sold over$45 million worth of artwork, and so far, all of Masterworks' exits have delivered positive net returns for their investors.

17:48Of course, historical returns are not a guarantee for future returns. I'm not a financial advisor, so do your own due diligence. There's always a risk of loss. Masterworks has over 750 ,000 users, and their art offerings usually sell out within hours, which is why they've had to create a wait list. But Eye on AI listeners can skip the line and get priority access right now by signing up at www.masterworks.art backslash I on AI, E-Y-E-O-N-A-I, all run together. Take a look and see what you think. Yeah, yeah, I agree. I mean, that's an argument I've given around the dinner table that the real existential threat is, the confirmed existential threat is climate change.

18:48And AI can do a lot to help us adapt or mitigate that. In fact, a few years ago, I think Joshua Benger and I were both authors on this, but David Rowanick wrote a really nice survey paper

19:08on many of the ways that AI can play a role in climate change. So that paper, which I still... Really, others, David Roenig and others did all the work, but I'm quite proud of having participated in that work to think through how AI can help climate change. Yeah, that was the foundational paper for climate change AI. I remember that. Yeah. And there's a lot of concern on the threat side, and not only the existential threat, but immediate harm about open source AI systems that allow people to build beyond the purview of regulators or, you know, corporate management and that sort of thing. I was surprised to find recently, I mean, I know you were, it was quite a while ago, but you were at Baidu.

20:11I was surprised to find that there's a very robust open source ecosystem growing up around large language models or large models in China. Do you think that open source is a danger? and specifically on China, have you used any of the open source stuff they've put on GitHub? Yeah, I think the open source community is a global one. There's a lot of great work in the United States. There's a lot of great work in Europe and there's also a lot of great work in China. So what I'm seeing is that open source work is a fantastic force that is giving access to a lot of people around the world, to large language models.

21:02I remember when Stable Diffusion released their image generation model, right, the trained model open source in late 2022, I think. There were justifiably reasonable fears of, wow, what if this is used for harm or this open source one, anyone can generate images now, what if that goes really poorly? And yes, indeed, that has to be acknowledged, there were some problematic use cases. But I would say, compared to the harm from a handful of problematic use cases to the massive amount of innovation with people incorporating the models into all sorts of creative video and image editing software, new UIs, I think the benefits vastly outweigh the harms.

21:46I'm not dismissing the harms at all. I think that the models do exhibit bias. you also to generate people of certain professions only generates pictures of one gender or narrow range of ethnicities those are real problems that we need to work to solve and there's certainly negative harmful use cases of you know deep fake imagery having said that what we've seen so far is that releasing open source seems to create much more benefit when when weight against the harms so i'm nervous about um proposals to require licensing of models open source models. That seems like an absolutely terrible idea that will stifle innovation and lead to regulatory capture, where it's much easier for big tech companies to satisfy regulatory requirements, thus concentrating power in large tech companies and disadvantaging academic research groups and smaller businesses that don't have time or resources that deal with overly onerous regulatory requirements that really are not, in my opinion, effective at really protecting people.

22:51Yeah. You know, and I don't want to get off on a dogleg on China, but immediately after GBT-3 came out, there was a lot of talk in the press about how China is going to be left behind because their speech laws, their laws that control speech, restrict political speech, will limit the efficacy of large language models. And I argued against that to a lot of people because our, you know, OpenAI or Microsoft or Google's large language models also have restrictions on speech. So I thought that was a spacious argument. Nonetheless, I was surprised to see open source thriving in China. Does that surprise you at all, just from what you know of China?

23:50Gosh, I don't know. I can't say I'm an expert on that. But it has been fantastic to see open source models thriving all around the world. And it's been interesting how global the research community is. A few weeks ago, I was in Canada. for CVPR, where, you know, speaking about landing AI's visual prompting technology. But it was nice to see researchers from all around the world come together and just share ideas freely and collectively, right, work hard to advance the science of AI, computer vision at that conference, but really AI more broadly. Yeah. Actually, and I'm sorry, I've left this till the end, but can you tell us about the visual prompting developments that you've been pursuing at Landing AI?

24:42Yeah, so I think the text prompting revolution, which I've already seen through ChaiGPD and BART and so on, is coming to vision. And so Landing AI has been developing visual prompting technology to make it easy for anyone to build and deploy computer vision applications. So maybe, yeah, I'm going to flash my nerd license and get technical for a second. But the text prompting revolution was really enabled by text transformers, right? The team I started and once led, Google Brain, published the text transformer paper in 2017. And since then, there was a wave of innovation by many, many research groups on scaling up text transformers.

25:25transformers and that led to GPT-2, GPT-3, you know, Megatron, sorry, many models along the way until we got InstructGP and then ChatGPT and BART and BingChat and so on. Not many people know, but the vision transformer paper came out about three years later in 2020. And if you go to computer vision conferences, I think since 2020, there's been this wave of innovation and scaling up and exploration of how to get vision transformers to understand images. And the underlying trends for both text and vision, but vision coming a bit later is these are models trained on very large amounts of data, text or images.

26:10The training is increasingly on unlabeled data. So this lets you have access to a lot more data or unlabeled or technically self-supervision. And what was already seen in text is with pre-training a neural network on tons of unlabeled text data, you can feed it a very simple text prompt and have it make inferences. What we're starting to see in computer vision, which I think is like maybe a couple years, one and a half years, two years behind text, is when you train a very large vision transformer on a lot of images on the internet, then you can give it a simple visual prompt and then have it start to make inferences in just seconds.

26:53So Lending AI will release our take on visual prompting where anyone can go and use it for free to label a few pixels and have it then automatically figure out what to do after that. As a developer too, this is exciting because just as we've seen with text prompting, applications that used to take me six months to build, now anyone can build in maybe a day. There are applications that used to take me six months to build in computer vision that with our tools, it's now getting to be possible to build in maybe a day or a few days. So the innovation is exciting, not just from my team landing AI, but from many groups working on computer vision.

27:38Yeah, that's fascinating. And just on visual prompting, large image models, if that's what they are, they're vectorizing an image by converting each line of pixels to a vector. Is that right? Patches of the image rather than... Catches, yeah. But they tokenize that in the same way that for a sentence,

28:17you convert that to an embedding, and then you're trying to predict the next token, right? Yes. And just like, yes, so I think in text-based transform is the core task, sometimes called the pretext task, is to predict the next word or technically predict the next token. So because text has that linear order, but they think the next word, the next token is a nice way to use unlabeled data. One of the reasons why there's so much buzz in exploration and computer vision is the process for taking an image and converting that to a sequence of tokens. There are a few different options for that. So do you convert to the patches?

29:00Do you hide or mask some patches along the way? But so different researchers are exploring different ways to do that. But there's multiple ways to see it through working and that scaling up of vision transformers, we're already seeing very exciting results, right from that. And then meta SAM segment anything model was another exciting breakthrough I'm just seeing. Actually, when I was at CVPR, you know, earlier this year in 2023, I felt that vision transformers have become a solid alternative to convolutional neural networks and the buzz and innovation. I remember a couple of years before chat GPT in the NLP on the text processing community, everyone knew something was in the air about transformers.

29:47No one knew exactly what's going to happen, but everyone knew something was, well, people knew something was up. And today at CVPR, I feel like something's in the air. People know something is up, even though the exact applications the exact ways this will get used is still being worked out no and and and your visual prompting the landing ai tool is specifically for uh i mean its primary use cases for uh labeling images is that right oh no um visual prompting is for building computer vision applications so for example, we have users using it for cell counting, quite a lot of light signs as users, but you know, given a picture of cells in a petri dish, in seconds, you can build an application to detect the cells and then post process it to count the number of cells.

30:37Or I don't know, I've used it to handle some satellite imagery, where she had a lot of geospatial error imagery applications as well. Easy might be, you know, finding tree cover, or you can now do that in seconds. And then of course, a lot of landing AI's users have been in manufacturing, industrial automation. And so for many of the manufacturing defect detection tasks, especially ones based on texture rather than shape, our visual prompting is letting people build and deploy systems in seconds. But I've been surprised also, even though I mentioned some of what we are seeing as our most common use cases, industrial manufacturing, life sizes, and geospatial error imagery, I'm seeing a very long tail of applications as well, where many people, many of our users are coming with all sorts of applications in computer vision that would not have imagined and now able to build and deploy them.

31:30For example, there was one group that was doing cosmology, you know, and since I'm not an astronomer, I would never have thought of doing that with a computer vision visual prompting application. Facilities, creativity, once people have this tool, is bewildering. Just like I think when people with access to chat GPT, you know, the creativity of things people did with it was incredible. And we're definitely not yet there at Computer Vision, but I think collectively the field is getting closer and closer. Yeah. At Landing AI, where do you, you personally, where are you on the spectrum between peer research and building applications?