Andrew Ng: Opportunities in AI - 2023

1. 0:00 – 0:05

   Intro

   1. SPSpeaker0:00

      [on-hold music]

2. 0:05 – 2:14

   AI as a General-Purpose Technology

   1. SPSpeaker0:05

      It is my pleasure to welcome Dr. Andrew Ng tonight. Andrew is the managing general partner of AI Fund, founder of Deep Learning AI and Learn, uh, and Landing AI, chairman and co-founder of Coursera, and an adjunct prof-professor of computer science here at Stanford. Previously, he had st-um, started and led the Google Brain team, which he had helped Google adopt modern AI, and he was also director of the Stanford AI Lab. About eight million people, one in one thousand persons on the planet, have taken an AI class from him. And through both his education and his AI work, he has changed numerous lives. Please welcome [clapping] Dr. Andrew Ng. [clapping]

   2. ANAndrew Ng1:07

      Thank you, Lisa, and it's good to see everyone. So what I want to do today is, um, chat to you about some opportunities in AI. So I've been saying AI is the new electricity. One of the difficult things to understand about AI is that it is a general-purpose technology, meaning that it's not useful only for one thing, but it's useful for lots of different applications, kind of like electricity. If I was to ask you, what is electricity good for? You know, it's not any one thing, it's a lot of things. So what I'd like to do is start off sharing with you how I view the technology landscape, and this will lead into, uh, the set of opportunities. So a lot of hype, a lot of excitement about AI, and I think a good way to think about AI is as a collection of tools. So this includes a technique called supervised learning, which is very good at recognizing things or labeling things, and generative AI, which is a relatively new, exciting development. If you're familiar with AI, you may have heard of other tools, but I'm gonna talk less about these additional tools and I'll focus today on what I think are currently the two most important tools, which are supervised learning and generative AI.

3. 2:14 – 5:40

   Supervised Learning: Labeling Things and the Scale Recipe

   1. ANAndrew Ng2:14

      So supervised learning is very good at labeling things or very good at computing input to outputs or A to B mappings. Given an input A, give me an output B. For example, given an email, we can use supervised learning to label it as spam or not spam. The most lucrative application of this that I've ever worked on is probably online advertising, where given an ad, we can label if a user's likely to click on it and therefore show more relevant ads. For self-driving cars, given the sensor readings of a car, we can label it with where are the other cars. One project that my team at AI Fund worked on was ship route optimization, where given a route that a ship is taking or considering taking, we can label that with how much fuel we think those will consume and use this to make ships, uh, more fuel efficient. Did a lot of work in automated visual inspection in factories, so you can take a picture of a smartphone that was just manufactured and label, is there a scratch or any other defect in it? Or if you want to build a restaurant review reputation monitoring system, you can have a little piece of software that looks at online restaurant reviews and labels that as positive or negative sentiment. So one nice thing, one cool thing about supervised learning is that it's not useful for one thing. It's useful for all of these different applications and many more besides. Let me just walk through, you know, concretely the workflow of one example of a supervised learning labeling things kind of project. If you want to build a system to label restaurant reviews, you then collect a few data points or collect a dataset where it say, you know, "The pastrami sandwich, great." It'll say that is positive. "Service is slow," that's negative. Uh, "My favorite chicken curry," that's positive. And, um, here I've shown three data points, but you're building this, you may get thousands of data points like this or thousands of training examples, we call it. And the workflow of a machine learning project or an AI project is you get labeled data, maybe thousands of, you know, data points. Then you have an AI engineering team train an AI model to learn from this data. And then finally, you would find maybe a cloud service to run the trained AI model, and then you can feed it, you know, "Best smoothie I've ever had," and that's positive sentiment. And so I think the last decade was maybe the decade of large-scale supervised learning. What we found starting about ten, fifteen years ago was if you were to train a small AI model, so train a small neural network, a small deep learning algorithm, basically a small AI model, maybe not on a very powerful computer, then as you fed it more data, its performance would get better for a little bit, but then it would flatten out. It would plateau, and it would stop being able to use the data to get better and better. But if you were to train a very large AI model, lots of compute on maybe powerful, you know, GPUs, then as we scaled up the amount of data we gave the machine learning model, its performance would kind of keep on getting better and better. So this is why when I started and led the Google Brain team, the primary mission that I directed the team to solve at the time was, let's just build really, really large neural networks that we then fed a lot of data to. And that recipe fortunately worked. And I think the idea of driving large compute and large scale data, that recipe's, uh, really helped us driven a lot of AI progress over the last decade. So if that was the last decade of AI, I think this decade, um, is turning out to be also doing everything we had in supervised learning, but adding to it, um, the exciting tool of generative AI. So

4. 5:40 – 7:40

   How Generative AI and Large Language Models Work

   1. ANAndrew Ng5:40

      many of you, maybe all of you, will have played with ChatGPT and Bard and so on, uh, but just, you know, given a piece of text, which you call a prompt, like, "I love eating," if you run this multiple times, maybe you get, "Bagels with cream cheese," or, "My mother's meatloaf," or, um, "Out with friends," and the AI system can generate output like that. Um, given the amount of buzz and excitement about generative AI, I thought I'd take just half a slide to, you know, say a little bit about how this works. So it turns out that [clears throat] Generative AI, at least this type of text generation, the core of it is using supervised learning that inputs output mappings to repeatedly predict the next word. Um, and so if your system reads on the internet a sentence like, "My favorite food is a bagel with cream cheese and lox," then, um, this is translated into a few data points where if it sees, "My favorite food is a," in this case, try to guess that the right next word was bagel. Or, "My favorite food is a bagel," try to guess next word is with. Um, and similarly, if it sees that, you know, in this case, [clears throat] the right guess for the next word would have been cream. So by taking text that you find on the internet or other sources, and by using this input-output supervised learning to try to repeatedly predict the next word, if you train a very large AI system, uh, on hundreds of billions of words, or in the case of the largest models, now more than a trillion words, then you get a large language model like ChatGPT. And, you know, there are additional other important technical details. I talked about predicting the next word. Technically, these systems predict the next subword or part of word called a token. And then there are other, uh, techniques like, um, RHF for further tuning the AI output to be more, uh, helpful, honest and harmless. But at the heart of it is this using supervised learning to repeatedly predict the next word that, that's really what's enabling, um, the exciting, you know, really fantastic progress on large language models.

5. 7:40 – 11:25

   LLMs as a Developer Tool and a Live Coding Demo

   1. ANAndrew Ng7:40

      So [clears throat] um, while many people have seen large language models as a fantastic consumer tool, you can go to a website like ChatGPT's website or Bard's or other large language models and use it as a fantastic tool. There's one other trend I think is still underappreciated, which is [clears throat] the power of, um, large language models not as a cons- not as a, not just as a consumer tool, but as a developer tool. So it turns out that, um, there are applications that used to take me months to build [clicks tongue] that a lot of people can now build much faster by using a large language model. So specifically, the workflow for supervised learning, building the restaurant review system, say, would be that you need to get a bunch of labeled data and, you know, maybe that takes a month to get a few thousand data points. And then [clears throat] have an AI team train and tune and really get, you know, optimized performance on your AI model. Maybe that'll take three months. Then find a cloud service to run it, make sure it's running robustly, make sure it's recognized. Maybe that'll take another three months. So a pretty realistic timeline for building a commercial-grade machine learning system is like six to twelve months, right? And so teams I've led will often take, you know, roughly six to twelve months to build and deploy these systems. And, and they-- some of them turned out to be really valuable, uh, but this is a realistic timeline for building and deploying a commercial-grade AI system. In contrast with prompt-based AI, where you write a prompt, uh, this is what the workflow looks like. [clicks tongue] You can specify a prompt that takes maybe minutes or hours, and then you can deploy it to the cloud, and that takes maybe hours or days. So there are now certain AI applications that used to take me, you know, literally six months, maybe a year to build, that many teams around the world can now build in maybe a week. And I think this is already starting, but the best is still yet to come. This is starting to open up a flood of a lot more AI applications that can be built by a lot more people. So I think many people still underestimate the magnitude of the flood of custom AI applications that I think is going to come down the pipe. Now, [clears throat] I know you probably were not expecting me to write code in this presentation, but that's what I'm going to do. Um, so it turns out this is all the code I need, um, in order to write a sentiment classifier. So I'm gonna, you know, some of you will know Python, I guess. Import some tools from OpenAI, and then add this prompt that says, "Classify the text below, delimited by three dashes, as having either a positive or negative sentiment." I don't know. I'm-- I don't know what I... "Had a fantastic time at Stanford GSB. Not only [mumbling] learned a lot and also made great new friends." All right. So that's my prompt. And now I'm just gonna run it. And I've never run it before, so I really hope it-- Oh, thank goodness, it got the right answer. [laughs] [audience applauding] [clears throat] And this is literally all the code it takes to build a sentiment classifier. Um, and so today, you know, developers around the world can take literally maybe like ten minutes to build a system like this. Um, and that's a very exciting development. So one of the things I've been working on was, um, trying to teach, you know, online classes, um, about how to use prompting, not just as a consumer tool, but as a developer tool. So

6. 11:25 – 15:51

   Where AI Value Sits Today and Where It Is Heading

   1. ANAndrew Ng11:25

      to start off the technology landscape, let me now share my thoughts on what are some of the AI opportunities I see. This shows what I think is the value of different AI technologies today, um, and I'll talk about three years from now. But the vast majority of financial value from AI today is, I think, supervised learning, where for a single company like Google can be worth more than one hundred billion US dollars a year. Um, and also there are millions of developers building supervised learning applications, so it's already massively valuable and also with tremendous momentum behind it, just because of the sheer effort in, you know, finding applications and building applications. And then Generative AI is the really exciting new entrant, uh, which is much smaller right now. And then there are the other tools I'm including for completeness. We can, you know... If the sizes of these circles represent the value today, this is what I think it might grow to in three years. So supervised learning, already really massive, may double, say, in the next three years, from truly massive to even more massive. And Generative AI, which is much smaller today, I think will much more than double in the next three years because of the nu- number of amount of developer interest, the amount of, um, venture capital investments, the number of large corporates exploring applications. And I also just point out, three years is a very short time horizon. If it continues to compound at anything near this rate, then in six years, you know, it'll be, it'll be even vastly larger. But, um, this light shaded region in green or orange, that light shaded region is where there are opportunities for either new startups, um, or for large companies, incumbents, to create and to, uh, enjoy value capture. But one thing I hope you take away from this slide is that all of these technologies are general purpose technologies. So in the case of supervised learning, a lot of the work that had to be done over the last decade, but is continuing for the next decade, is to identify and to execute on the concrete use cases. Um, and that process is also kicking off for, um, Generative AI. So for this part of the presentation, I hope you take away from it that general purpose technologies are useful for many different tasks. A lot of value remains to be created using supervised learning. And, and even though we're nowhere near finishing figuring out the exciting use cases of, uh, supervised learning, we have this other, you know, fantastic tool of Generative AI, which further expands the set of things we can now do using AI. Um, but one caveat, which is that there will be short-term fads along the way. So I don't know if, um, some of you might remember the app called Lensa. This is the app that would let you upload pictures of yourself, and it'll render a cool picture of you as an astronaut or a scientist or something. Um, and it was a good idea, and people liked it. And it's relatively just took off like crazy like that through last December, and then it did that. And [clears throat] that's because Lensa was-- It was a good idea. People liked it. But it was a relatively thin software layer on top of someone else's really powerful APIs. Um, and so even though it was a, it was a useful product, it wasn't a defensible business. Um, and when I, when I-- One thing about, you know, apps like Lensa, um, I'm actually reminded of when Steve Jobs gave us the iPhone. Shortly after, someone wrote an app that I paid one ninety-nine for, um, to do this, to turn on the LED, to turn the phone into a flashlight. And that was also a good idea, uh, to write an app to turn on the LED light, but it also wasn't a defensible long-term-- it also didn't create very long-term value because it was, uh, easily replicated and underpriced and then eventually incorporated into iOS. But with the rise of iOS, with the rise of iPhone, someone also figured out how to build things like Uber and Airbnb and Tinder, the very long-term, very defensible businesses that created, you know, sustaining value. And I think with, um, the rise of Generative AI or the rise of new AI tools, I think what really, what excites me is the opportunity to create those really deep, really hard applications that hopefully can create very long-term value. So the first trend I want to share is AI as a general purpose technology. And a lot of the work that lies ahead of us is to find the very diverse use cases, um, and to build them.

7. 15:51 – 20:24

   The Long Tail: Low-Code Tools Bring AI to Other Industries

   1. ANAndrew Ng15:51

      There's a second trend I want to [clears throat] share with you, which relates to why AI isn't more widely adopted yet. It feels like a bunch of us have been talking about AI for like fifteen years or something. But if you look at where the value of AI is today, a lot of it is still very concentrated in consumer software Internet, right? Once you go outside, you know, tech or consumer software Internet, there's some AI adoption, but it, but it all feels very early. So why is that? It turns out if you were to take all current and potential AI projects and sort them in decreasing order of value, then to the left of this curve, the head of this curve, are the multi-billion dollar projects like advertising or web search or, um, for e-commerce, you know, product recommendations or a company like Amazon. And it turns out that about ten, fifteen years ago, you know, various of my friends and I, we figured out a recipe for how to hire, say, a hundred engineers to write one piece of software to serve more relevant ads and apply that one piece of software to a billion users and generate massive financial value. So that works. Um, but once you go outside consumer software Internet, hardly anyone has a hundred million or a billion users that you can write and apply one piece of software to. So once you go to other industries, as we go from the head of this curve on the left over to the long tail, these are some of the projects I see and am excited about. I was working with a pizza maker that was taking pictures of the pizza they were making because they needed to do things like make sure that the cheese is spread evenly. So this is about a five million dollar project. Um, but that recipe of hiring hund- a hundred engineers or dozens of engineers to work on a five million dollar project, that doesn't make sense. Um, [clears throat] or as another example, working with an agriculture company that, um, with them, we figured out that if we use cameras to find out how tall is the wheat, and wheat is often bent over because of wind or rain or something, and we can chop off the wheat at the right height, then that results in more food for the farmer to sell, and it's also better for the environment. But this is another, you know, five million dollar project that that old recipe of hiring a large group of high school engineers to work on this one project, that doesn't make sense. Um, and similarly, materials grading, cloth grading, sheet metal grading, many projects like this. So whereas to the left, in the head of this curve, there's a small number of, let's say, multi-billion dollar projects, and we know how to execute those, you know, delivering value. In other industries, I'm seeing a very long tail of tens of thousands of, let's call them five million dollar projects that until now have been very difficult to execute on because of the high cost of customization. The trend that I think is exciting is that the AI community has been building better tools that lets us aggregate these use cases and make it easy for the end user to do the customization. So specifically, I'm seeing a lot of, um, exciting l- low-code and no-code tools that enable the user to customize the AI system. What this means is, instead of me needing to worry that much about pictures of pizza, um, we have tools, we can-- we're starting to see tools that can enable the IT department of the pizza-making factory to train an AI system on their own pictures of pizza to realize this five million dollars worth of value. And by the way, the pictures of pizza, you know, they don't exist on the Internet, so Google and Bing don't have access to these pictures. Uh, we need tools that can be used by really the pizza factory themselves to build and deploy and maintain their own custom AI system that works on their own pictures of pizza. Um, [clears throat] and broadly, the technology for, for enabling this, um, some of it is prompting, um, s- text prompting, visual prompting, by really large language models and, and then similar tools like that, um, or a technology called data-centric AI, whereby instead of asking the pizza factory to write a lot of code, you know, which is, which is, which, which is challenging, we can ask them to provide data, which turns out to be more feasible. And I think the second trend is important because I think this is a key part of the recipe for taking the value of AI, which so far still feels very concentrated in the tech world and the consumer software Internet world, and pushing this out to, you know, all industries, really to the rest of the economy, which, which, you know, sometimes it's easy to forget. The rest of the economy is much bigger than the tech world.

8. 20:24 – 24:56

   The AI Stack: Where the Opportunities Are

   1. ANAndrew Ng20:24

      So the two trends I shared, AI is a general-purpose technology, lots of concrete use cases to be realized, as well as, um, low-code, no-code, easy-to-use tools enabling AI to be deployed in more industries. W- How, how do we go after these opportunities? So about five years ago, there was a puzzle I wanted to solve, which is I felt that many valuable AI projects are now possible, and I was thinking, "How do we get them done?" And having led AI teams in, you know, Google and Baidu in big tech companies, I had a hard time s- figuring out how I could operate a team in a big tech company to go after a very diverse set of opportunities in everything from maritime shipping to education to financial services to healthcare and on and on. It's just very diverse use cases, very diverse go-to-markets and very diverse really, you know, customer bases and then applications. Um, and I felt that the most efficient way to do this would be if we can start a lot of different companies to pursue these very diverse opportunities. So that's why I ended up starting AI Fund, uh, which is a venture studio that builds startups to pursue a diverse set of AI opportunities. Um, and of course, in addition to lots of startups, incumbent companies also have a lot of opportunities to integrate AI into existing businesses. In fact, one pattern I'm seeing for incumbent businesses is, um, uh, distribution, uh, is, is often one of the significant advantages of incumbent companies, is they play their cards right, can allow them to integrate, uh, AI into, into their products quite efficiently. But just to be concrete, where are the opportunities? So I think of this as a-- this is what I think of as the AI stack. Right. At the bottom level is the, uh, hardware semiconductor layer. Fantastic opportunities there, but very capital-intensive, very concentrated, so it needs a lot of resources, relatively few winners. So some people can and should play there. I personally don't like to play there myself. There's also the infrastructure layer. Also fantastic opportunities, but very capital-intensive, very concentrated, so I tend not to play there myself either. And then there's the developer tool layer. Um, what I showed you just now was I was actually using OpenAI's API as a developer tool. And then I think the developer tool sector is, uh, hypercompetitive. Look at all the startups chasing OpenAI right now. But there will be some mega winners. And so I sometimes play here, but primarily when I think I have a meaningful technology advantage, because I think that earns you the right or earns you a better shot at being one of the mega winners. And then lastly, even though a lot of the media attention and the buzz is in the infrastructure and developer tooling layer, it turns out that that layer can be successful only if the application layer is even more successful. And we saw this with the rise of SaaS as well. A lot of the buzz and excitement is on the technology, the tooling layer, which is fine, nothing wrong with that. But the only way for that to be successful is if the application layer is even more successful so that, frankly, they can generate enough revenue to pay the infrastructure and the tooling layer. So actually, let me mention one example. Um, Amourai, uh, I was actually just texting the CEO yesterday. But Amourai is a, a company that we built that uses AI for romantic relationship coaching, right? Um, and just to point out, I'm an AI guy, and I feel like I know nothing really about romance. Um, and if you don't believe me, you can ask my wife. She will confirm that I know nothing about romance. But when we want to build this, we wound up getting together with the former CEO of Tinder, uh, Renata Nyborg, and with my team's expertise in AI and her expertise in relationships, because she ran Tinder, she knows more about, you know, relationships than I think anyone I know. We're able to build something pretty unique using AI for, you know, kind of romantic relationship mentoring. Um, and the interesting thing about applications like these is when we look around, um- You know, how many teams in the world are simultaneously expert in AI and in relationships? And so at the application layer, I'm seeing a lot of exciting opportunities that seem to have a very large market, but where the competition set is very light relative to the magnitude of the opportunity. It's not that there are no competitors, but it's just much less intense compared to, uh, the developer tool or the infrastructure layer, say.

9. 24:56 – 29:05

   The Startup Recipe: Baring AI Case Study

   1. ANAndrew Ng24:56

      And so because I've spent a lot of time iterating on a process of building startups, what, what, what I'm gonna do is just, you know, very transparently tell you the recipe we've developed for building startups. And so after many years of iteration and improvement, this is how, you know, we now build startups. My teams always had access to a lot of different ideas, you know, internally generated, ideas from partners. And I wanna walk through this with one example of something we did, which is a company, Baring AI, which uses AI to make ships more fuel efficient. So this idea came to me when, a few years ago, a large Japanese conglomerate called Mitsui, that is a major shareholder and s- operates, uh, major shipping lines, uh, they came to me and they said, "Hey, Andrew, you should build a business to use AI to make ships more fuel efficient." Um, and the specific idea was, you know, think of it as a Google Maps for ships. We can suggest a ship or tell a ship how to steer so that you still get to your destination on time, but using, it turns out, about ten percent less fuel. Um, and so what we now do is we spend about a month validating the idea. So double-check, is this idea even technically feasible? And then talk to prospective customers to make sure there's a market need. So we spend up to about a month doing that. And if it passes this stage, then we will, um, go and recruit a CEO to work with us on the project. When I was starting out, I used to spend a long time working on a project myself before bringing on a CEO. But after iterating, we realized that bringing on a leader at the very beginning to work with us, it reduces a lot of the burden of having to transfer knowledge or having a CEO come in and reval-- having to revalidate what we discovered. So the process is we've learned much more efficient, we just bring in the leader at the very start. Um, and so in the case of Baring AI, we found a fantastic CEO, Dylan Kyle, who's a repeat entrepreneur, uh, one successful exit before. And then we spent three months, um, six two-week sprints to work with them to build a prototype as well as do, do deep customer validation. If it survives this stage, and we've, we have about a two-thirds, sixty-six percent survival rate. We then write a first check-in, which then gives the company resources to hire an executive team, you know, build the key team, get an MVP working, minimum viable product working, and get some real customers. And then after that, you know, hopefully then successfully raises ad- additional external rounds of funding and can, and can keep on growing and scaling. So I'm really proud of the work that my team was able to do, um, to support Mitsui's idea and Dylan Kyle as CEO. And today, there are, um, hundreds of ships on the high seas right now that are steering themselves differently because of Baring AI. And, um, ten percent fuel savings translates to rough order amount to maybe four hundred and fifty thousand dollars in savings in fuel per ship per year. And of course, it's also, frankly, quite a bit better for the environment. Um, and I think this startup, I think, would not have existed if not for Dylan's fantastic work, um, and then also, you know, Mitsui bringing this idea to me. And I like this example because this is another one. It's like, you know, this is a startup idea that, just to point out, I would never have come up with myself, right? Because, you know, I've been on a boat, but what do I know about maritime shipping? Um, but it's the deep subject matter expertise of Mitsui, um, that had this insight together with Dylan and then my team's expertise in AI that, that made this possible. And so as I operate in AI, one thing I've learned is my swim lane is AI, um, and that's it. Because I don't have time. Uh, it's very difficult for me to be expert in maritime shipping and romantic relationships and healthcare and financial services and on and on and on. And so I've learned that if I can just help get a accurate technical validation, um, and then use, you know, AI resources to make sure the AI tech is built quickly and well. And I think we've always managed to help the companies build a strong technical team, then partnering with subject matter experts, um, often results in, in exciting new opportunities.

10. 29:05 – 31:18

    Staying in the AI Swim Lane and Starting from Concrete Ideas

    1. ANAndrew Ng29:05

       And I wanna share with you one other weird aspect of, uh, ano- one other weird lesson I've learned about, you know, building startups, which is, um, I like to engage only when there's a concrete idea. And this runs counter to a lot of the advice you hear from the design thinking methodology, which often says, "Don't rush to solutioning," right? Explore a lot of alternatives before you do a solution. Honestly, we tried that. It was very slow. Um, but what we've learned is that at the ideation stage, if someone comes to me and says, "Hey, Andrew, you should apply AI to financial services." Because I'm not a subject matter expert in financial services, it's very slow for me to go and learn enough about financial services to even figure out what to do. I mean, eventually you could get to a good outcome, but it's a very labor intensive, very slow, very expensive process for me to try to learn industry after industry. In contrast, uh, one of my partners wrote this idea as a tongue-in-cheek, not, not really seriously, but you know, let's say a concrete idea is buy GPT. Let's eliminate commercials by automatically buying every product advertised in exchange for not having to see any ads. It's not a good idea, but it is a concrete idea. And it turns out concrete ideas can be validated or falsified efficiently. Um, they also give a team a clear direction to execute. And I've learned that in today's world, especially with, you know, the excitement and buzz and exposure to AI of a lot of people It turns out that there are a lot of subject matter experts in today's world that have deeply thought about a problem for months, sometimes even, you know, one or two years, uh, but they've not yet had a build partner. And when we get together with them and, um, share, and they share the idea with us, it allows us to work with them to very quickly go into validation and building. And I find that this works because there are a lot of people that have already done the, you know, design thinking thing of exploring a lot of ideas and winning down to really good ideas. And there are-- I find that there are so many good ideas sitting out there that no one is working on, um, that finding those good ideas that someone al- has already had and wants to share with us and wants to build partner for, that turns out to be a much more efficient, um, engine.

11. 31:18 – 33:37

    Risks, Bias, and Job Disruption

    1. ANAndrew Ng31:18

       So, um, [clears throat] before I wrap up, we'll go to questions in a second. Just a few slides to talk about risks and social impact. Um, so AI is a very powerful technology. To state something you probably guess, my teams and I, we only work on projects that move humanity forward. And, you know, we have multiple times killed projects that we assess to be financially sound, uh, based on ethical grounds. It turns out I, I've been surprised and sometimes dismayed at the creativity of people to come up with good ideas... Sorry, to come up with really bad ideas that, that seem profitable but really should not be built. Um, we've killed a few projects on those, on those grounds. And then I think it has to be acknowledged that AI today does have problems with bias, fairness, accuracy, um, but also, you know, the technology is improving quickly. So I see that AI systems today are less biased than six months ago and more fair than six months ago, which is not to dismiss the importance of these problems. They are problems, and we should continue to work on them. But I'm also gratified at the number of AI teams working hard on these issues to make them much better. When I think of the biggest risk of AI, I think that the biggest risk, uh, one of the biggest risks is the disruption to jobs. Um, this is a diagram from a paper by our friend at the University of Pennsylvania and some folks at OpenAI analyzing the exposure of different jobs to AI automation. And it turns out that whereas the previous wave of automation, mainly, um, the most exposed jobs were often the lower wage jobs, such as when, you know, we put robots into factories. Uh, with this current wave of automation, it's actually the higher wage jobs furthest to the right of this axis that seems to have more of their tasks exposed to AI automation. So, [clears throat] um, even as we create tremendous value using AI, I feel like as citizens, um, and, and our corporations and the governments and really our society, I feel a strong obligation to make sure that, um, people, especially people whose livelihoods are disrupted, are still well taken care of, are still treated well. And then lastly, um, there's also been... It, it feels like every time there's a big wave of progress in AI, you know, there's a, there's a

12. 33:37 – 36:49

    AGI Hype, Extinction Risk, and Closing Remarks

    1. ANAndrew Ng33:37

       big wave of hype about artificial general intelligence as well. When deep learning started to work really well ten years ago, there was a lot of hype about AGI. Now that generative AI is working really well, there's another wave of hype about AGI. Um, but I think that ar- artificial gene- general intelligence, AI didn't do anything a human can do, is still decades away, you know, maybe thirty to fifty years, maybe even longer. I hope we'll see it in our lifetimes, but I don't think it's anytime soon. One of the challenges is that the biological path to intelligence, like humans, and the digital path to intelligence, you know, AI, they've taken very different paths. And the funny thing about the definition of AGI is you are benchmarking this very different digital path to intelligence with really the biological path to intelligence. So I think, you know, large language models are smarter than any of us in certain key dimensions, but much dumber than any of us in other dimensions. And so forcing it to do everything a human can do is like a funny comparison. Um, but I hope we'll get there, maybe, may- hopefully within our lifetimes. And then [clears throat] there's also a lot of, I think, overblown hype about AI creating extinction risks to humanity. Um, candidly, I don't see it. I just don't see how AI creates any meaningful extinction risk to humanity. Um, I think that people worry we can't control AI, but we have lots of, uh, AI will be more powerful than any person. But we've lots of experience steering very powerful entities, such as corporations or nation-states, that are far more powerful than any single person, um, and making sure they, for the most part, benefit humanity. And also, technology develops gradually. The so-called hard takeoff scenario, where it's not really working today, and then suddenly one day overnight, it works brilliantly, and we achieve super intelligence, takes over the world, th- that's just not realistic. And I think AI technology would develop slowly, like all the tech, you know, like... And, and then it gives us plenty of time to make sure that, um, we provide oversight and can manage it to be safe. And lastly, if you look at the real extinction risk to humanity, um, such as, fingers crossed, the next pandemic, um, or climate change leading to massive depopulation of some parts of the planets, or much lower odds, but maybe someday an asteroid doing to us what it had done to the dinosaurs. I think if you look at the actual real extinction risk to humanity, AI having more intelligence, even artificial intelligence in the world, will be a key part of the solution. So I feel like if you want humanity to survive and thrive for the next thousand years, rather than slowing AI down, which some people propose, I would rather, I would rather make AI go as fast as possible. Um, some of that, just to summarize, this is my last slide. I think that AI, as a general purpose technology, creates a lot of new opportunities for everyone. And a lot of the exciting and important work that lies ahead of us all is to go and build those concrete use cases. Um, and hopefully in the future, hopefully I'll have opportunities to maybe engage with, uh, more of you on those opportunities as well. So with that, let me just say thank you all very much. [audience applauding]

Episode duration: 36:54