We are so, so honored to have with us today Andrew Ng. He's, I guess, still my colleague at Stanford, and he's one of the still most highly cited machine learning AI researchers, was a real pioneer. He is also the founder of DeepLearning.AI. He's the managing general partner of AI Fund. So he has probably better visibility to the AI startup landscape than anybody in the world. Before that, many of you know him best as founder of Coursera. Those of us in education certainly have seen and appreciated so much of his work and benefited from his efforts to get all of our research and teaching in front of the world. And he has taught AI to more than 8 million people. So we are so honored to have you here today, Andrew. So please join me on the stage. (applause)
So it's just really hard to pick what to focus on with Andrew since he knows so much about so many things. But we're going to have a pretty wide-ranging conversation today, and we'll try to extract as much of his wisdom as we can in 45 minutes. So I thought we could start out with something really fun, just to make sure everybody's still awake at the end of a day of conferencing. The headlines in the last two weeks have basically been about how we're all gonna die. I have young adult children. They've been reading about it too. They come home every night and want to know why I'm not doing something to save us from how we're all going to die. And I think at a time like this, policymakers really do need to understand how seriously to take things and to focus their efforts on the right risks and not waste scarce effort on the wrong risks. So, Andrew, you have a visibility to everything from the science to the implementation, so I would love to get your take on whether we're all going to die.
So it's really funny, you know. I wake up in the morning, put on my clothes, go to work, and I don't seem to worry about AI killing me off. A few years ago, with the rise of modern AI, generative AI, I was taken aback by the intensity of the lobbying efforts from a handful of businesses that had often spent hundreds of millions or billions of dollars training up AI models that wanted to stifle what they call open-source and open-weight AI models, which is when someone—maybe a company, maybe a research lab—trains an AI model and posts it on the internet free for anyone to use. And it turns out if you spent a billion dollars developing some AI model, it's actually kind of annoying if someone else wants to give something similar away for free. Really degrades the value of your investment. So at that time, there started to be this intense wave of lobbying effort about these supposed dangers of AI. And AI is not completely risk-free. There are some harmful applications. But back then there were irresponsible analogies made between AI and nuclear weapons with no justification. I mean, one helps us get work done, the other blows up cities. I see no relationship between the two. But all this was part of a very intense campaign to try to convince regulators to pass stifling regulations to favor the incumbents and to shut out new entrants, especially open-source, open-weight competitors. Because of this, a number of us flew to DC to testify, did a lot of communications work, and I think at that time, this wave of attacks on AI, they lost credibility. But just in the last couple of weeks, with a new campaign that seems quite well orchestrated, I think AI has not actually had any new material advance leading to dangerous capabilities that could cause human extinction. I mean, AI, frankly, doesn't work that well. There has been improvement in AI cyber capabilities that we should take seriously, cyber attacks, cyber defense. But the one thing in the news that has shifted perception was when a team at OpenAI launched a swarm of agents that hacked into Hugging Face, and this was widely reported. So this was a step up in cyber offense, and also cyber defense, it turns out, capabilities. But this change in perception has been parlayed into, well, maybe these AI agents are out of control and can kill us all. It turns out that OpenAI's, I think, unintended hack on Hugging Face was enabled by OpenAI having poor internal security operations practices. Specifically, the sandboxing to try to contain the AI had bugs in it. And OpenAI at the time had poor monitoring processes to try to figure out what this agent swarm was actually doing. And a lot of PR headlines said things like a swarm of 1,200 agents or something like that. And this makes people picture, I don't know what picture you have. It's like a team of 1,200 ninjas running around with people running around trying to attack. But on my personal laptop, I often have more than a thousand compute processes running. My laptop, for years can do more than a thousand things at the same time. And yes, it is a step up in capability to get 1,200 AI agents to work together, but it's not the end of the world, and it's not a dramatic, shocking capability, although it's not none of an advance either. But unfortunately, well, fortunately, AI keeps on advancing. I think that's a really good thing. But unfortunately, because AI keeps on advancing and the technology is poorly understood, every time the technology advances and parts of it are poorly understood, it creates opportunities for those wanting to hype up AI or stoke AI fears to create elements that are not well understood by the public or by regulators and present that in a way to stoke up fears in order to try to get regulatory capture stifling laws passed or to get PR or to convince people they deserve more funding. One thing that was new was to try to disclaim responsibility. But I think it's actually very unfortunate that here in America, the fear on AI has spiked up again. I think this is actually very damaging for America and for the countries that feed this fear because it will slow down progress. One of the most important things for the future, for example, is frankly to build more data centers. Data centers are a key enabler to get societies to do a lot more work done. And it turns out that data centers are incredibly efficient for the work they do. So I know there's a lot of hype about data centers, water usage, which was actually a lot of misinformation promulgated by a book that unfortunately had a major fact error. But I think that, turns out one of the best things we can do for the environment is to build more data centers and to move workloads from traditional dirty on-prem workloads into data centers where they're frankly way more efficient. And so this type of PR, if it results in stifling regulations, will be anti-competitive and bad for America and bad for the developing world. And if it slows down data center construction and slows down AI adoption, I think this would be a huge missed opportunity for us and our democratic allies.
So it's funny, what's old is new again. The last time I sat on the stage with you, Andrew, was when you were so generous to come to a conference that I organized in 2024 about competition in AI. And at that time, this whole debate was kind of at its peak. And then when we invited you here for this, I wasn't sure that would be a big topic, but then it's popped back up again. But certainly that conference was about competition. But from a competition perspective, a product that's not just free now, but free and you can keep forever, is a pretty unique product, and it's got other attributes as well from the perspective of the firms and governments adopting open source. And if you're buying a cheap subscription now, the price could go up later, whereas if you've downloaded the model, you know you have it. In addition, if you're trying to run some kind of critical infrastructure on AI that you're importing from another country, that could be very risky, and it could give the other country leverage over you, versus if you're operating it locally, there's no kill switch for your economy. So I'm sort of curious, what is your take beyond this sort of competition angle for what open models really mean for the developing world?
Yeah. So today, a lot of the leading frontier AI models are still relatively expensive, or maybe not that expensive, but relatively expensive, especially relative to the resources of many individuals in the developing world. And so open-source, open-weight models is an important way to give individuals and businesses and governments and research organizations, without a lot of resources, access to some of these AI models. And so in many developing countries, open-source, open-weight models have tremendous adoption and are used by people in a daily basis. And to your point, Susan, about sovereign AI, with the complicated geopolitical situation we're in, if a country wants to maintain control over its own AI, it turns out for a country to develop its own AI system from scratch is difficult, expensive, takes a lot of resources. But if a country or if a community of nations collectively supports open-source, open-weight models, then maybe no one needs to own that open-source, open-weight model. But the advantage is, as Susan said, no one else can stop you from using it either. And I think this gives, in today's complicated world where nations and businesses are worried about who could cut off their access to AI, open source and open weight is one of the most economic ways to ensure continued access to AI that no one can deny you of. I think one thing that's underestimated as well is the importance of open-weight models in academic research. So a lot of the work, a lot of really good work today is done in frontier AI labs, which in the US and have become quite closed and proprietary in the sense of not releasing a lot of details of their training recipes, what data they're training on, how to deliver these techniques. And it turns out that open-weight models are ones that any research group and any university can just download and study. And this has been a huge force for enabling the free academic realm of things, to use these models and also to develop technologies or also develop methods to adapt these models to a specific country's language, use case, regional dialects, to train the models to understand that nation in a way that maybe American frontier lab may not have had the time to get the data that relates to that country to make an AI model smart yet about that country's language and the facts about the country. So, to me, one great thing about open-weight models is it gives everyone in the developing world and in developed economies more choice. And to be clear, I hope the American frontier labs do well. I think that there's room for the leading American firms to be great businesses, and at the same time, for there to be also open-weight options that everyone can choose from.
Well, from an academic perspective, we've certainly gone crazy using the open weight models here, and it is amazing how much you can do for relatively low cost, whether we're using cloud services, when our lab is using things like Together.ai, and you can get a whole bunch of stuff classified for $25. Just put in your credit card. And so it's really quite amazing as just a research tool across the social sciences. And then just from a user perspective, I like to give the example of my son who had to commute twice a week to an activity through a dead zone. And so he just downloaded an open model so he could keep coding on his laptop as he went through the dead zone and didn't have good internet access, and then switch back to the cloud-based models when he got through the dead zone. And that just kind of shows that although we all get focused in the newspaper on who's winning the math competition, there's a large volume of work now that I see big enterprises and even rich people in developing countries using just because it's most convenient for their circumstance. And then at the WDR, we traveled around Africa, and we met professors from different universities who were training models for local dialects with a small amount of compute and using it in low-resource environments. So the promise really seemed to be coming to fruition in those kind of pilot cases. I'm sort of curious if you could say more about what you're seeing in terms of applications that can help meet people where they are in low-resource environments, whether with open-weight models or in other ways?
One of the wonderful things about AI is it has lowered the barrier to building custom software. In fact, one of the most important changes in work has been software engineering, because AI turns out to be remarkable at writing code. And what this means is it gives nations as well as businesses without the resources of having a lot of software engineers more opportunities to build a lot of exciting, cool, custom software. And then one other interesting thing that's happened is, there was a lot of hype about AI taking people's jobs. And again, we seem to be more worried about that in the developed economies than developing economies are worried about AI job displacement. But nonetheless, in software engineering, even though AI can write code, postings for software engineers are up, right? And maybe it's not surprising that if AI makes software engineers much more productive, frankly, we wanna hire a lot more of them. And just for myself and a lot of my friends, we just can't seem to find enough highly qualified AI engineers. And to be clear, if someone writes code like it's still 2022 before ChatGPT, I will not hire them. It's just because the nature of software engineering has changed, and only engineers that have learned the new skill set, frankly, know how to enjoy the massive productivity gains and do the great work that can now be done. But AI-native or people skilled in AI, we just can't find enough of them. And for the developing world as well that did not have the resources to do a lot of this work, the barrier to learning how to build software applications has lowered dramatically. And so I think there'll be a lot more that these countries and businesses will be able to do. One other piece of, I think, hopefully good news. One thing I'm actually pretty convinced by is, AI has entered software engineering so strongly because as developers, we write software, it's kind of human nature to write tools to help our own work. I think AI will slowly enter other fields as well. It's making inroads into finance, into marketing, into other sectors as well. And to the extent that what's happened in software engineering will be a harbinger or predictor of what will happen in other job categories as well, I actually feel optimistic about the global job market going forward, with one caveat, which is that, for a lot of job roles, doing things the way that we did five years ago is just not interesting or particularly useful. But if countries or businesses can upskill people to do things in a more AI-native way, then the amount of productive work and exciting, fun, and productive work that can be done seems huge. And for the developing world, I think there should be many interesting opportunities to not just upskill, but also potentially leapfrog. So that would be exciting to see it play out in the next few years.
So I'd love to talk a little bit more about AI applications. And so one observation that I have is that the narrative we have, both in the popular discourse as well as from a lot of social scientists, is really focused around the chatbot, the ChatGPT, the Claude. And of course, it's not surprising because that's what most humans have interacted with so far. And to the extent that there are scientists out there, a lot of them are coming from the big labs because they're the ones who are building the products and are most familiar with it. And there's less discourse about a future where a large language model is just an input, and there's a whole application layer that's built using these software development tools, but also, as you say, is customized software and meeting people where they are. That may be more important in the developing world, where, as the WDR report shows, the frontier lab performance falls way off on dialects, on local dialects, even like medium-sized languages already aren't well served. But in even a country like India, which is a big country, there's big swaths of the country where the language skills are just nowhere near what they need to be for really production uses. And then another trend recently has been that the labs have realized that in the US, even in developed countries with rich customers, rich business customers, they still need to go out and send humans in to implement because the implementation, it's not just, you know, use this product that's put up on the web, but actually there's a lot of work in implementation. So I'm curious, putting on your investor hat and your observation of what's getting built out there, can you fill in more of the gap about the application layer and what you see coming and what you think could be particularly valuable?
So most consumers have been using chatbots like ChatGPT, Claude, Gemini, Meta AI, Grok, and so on. And that's exciting and very valuable. But when my teams, AI Aspire and so on, go into many businesses to look at business applications, there's a lot of customization work to be done. So for example, when we work with financial institutions, we'll often work with, say, a large bank to take a look at the internal business processes and the workflows and figure out which ones can be automated or partially automated with AI workflow or AI agentic workflow. So, take loan approvals or underwriting. There are many steps to get a loan application, figure out, do the due diligence, maybe KYC to verify the customer's identity, a bunch of steps in order to maybe make a loan approval or rejection decision. And it turns out that it's in practice often a lot of work to look at the many ideas for what a single business or what a government could implement, and then to sort through based on value and technical feasibility, which ones we should tackle. And then also to do the work, to implement it, make sure it's reliable, test it, deploy it, make sure it's compliant, secure, private, and so on. I think that for the next decade, quite possibly much longer than a decade, all of us will have a lot of work to do to identify workflows to fully or partially automate this. I know some of the hype in Silicon Valley is, in two quarters, AI will be AGI or superintelligent or whatever, and that solves all problems. And frankly, I wish it were that easy. AI intelligence is very jagged. It's amazing at some things, and it's frankly not very good at a lot of other things, which is why the practical business implementation of AI requires going in and understanding use cases, and then frankly doing quite a lot of work to implement AI systems to do some of that work for you. And there will be a lot of opportunities, both in the developing world and in developed economies, in order to identify and implement those opportunities.
So let's take this conversation into another area in which you're the world expert, which is education. And so obviously it was incredibly powerful to democratize access to education, to teaching, through your work, through Khan Academy and others. But then we also saw that not everybody takes it up. So just because the information is out there to be gotten doesn't mean that people know what's the right fit for them, and for a variety of reasons, they don't go and get it. Now, AI allows another level of personalization. It could help people navigate, it can help curate educational materials. So we can take that and then put it against the backdrop that in the World Development Report, we describe the large gaps between rich and poor countries in terms of teachers per student. There are also large gaps in the education and qualification of the teachers. A lot of science teachers don't have science education and so on. The use case remains the same use case that you originally went after. So I'm curious. Now, what are the biggest new opportunities with this latest round of AI innovation for getting education to really be globally available?
Yeah. So the educational teams I work with, Coursera, DeepLearning.AI, LearnVector, I feel like we still have a long ways to go to help more people gain new skills. But I want to share something that, to me, is clearly not working in education right now. And I want to say something that may sound controversial, but I think it's just very obviously true, which is that chat is terrible for learning. At universities, all of us professors, this is actually very clear, but the rigorous academic studies are also coming out to demonstrate this, that when students use chatbots for homeworks, their homework scores go up. But their final exam scores or measures of long-term learning go down. And I think this has just been obvious to teachers on campuses for a long time. But it's actually no surprise that if a student engages in cognitive offloading to a chatbot, so have a chatbot do the work for you, the work gets done. So you get great homework scores, but learning suffers. So I feel like there is a little bit of delusion in some circles that chat is helpful for learning. And yes, there are some ways that people could use chatbots that are helpful for learning, but the fact is the vast majority of ways that the vast majority of people are using chatbots is very harmful for learning. Mind you, I'm very pro-AI. They're wonderful at getting stuff done. I use them to get stuff done all the time, but let's not fool ourselves that they're a great tool for learning, either at the formal education levels, kind of higher ed, K-12, or even for adult learners. For adults, use it to get stuff done, but let's not fool ourselves into thinking this is really good at helping us to learn. So that's what's happening now and where AI is not working, at least in its current instantiations. Having said that, we know in tech, what works today is not the same as what we'll manage to get to work in the future. And I am very optimistic that building on top of AI technology, we will be able to build learning experiences that are based on sound pedagogy, that teach, they're engaging, and that also allow us to go from much more of a one-to-many to a one-to-one experience. So even today, on a lot of online education, everyone kind of watches the same video, and it works. There's really nothing wrong with that. It does work. But with AI technology, there is now a much more clear path to me to turn learning from this one-to-many model, one instructor, many learners, to a much more customized one-to-one model. And we still have work to get there, but when we do start to roll out things like that, then that gives us a path which is different to the current chatbot path to actually help learning. And this ties back to something I mentioned just now, which is there's going to be a lot of work for humans. It's interesting, by the way, one funny thing about AI culture, and Susan does this too, for some reason, AI people, we don't call people people, we call people humans. I'm still not quite sure why we do that. But so it's gonna be a lot of work for humans, but the skill mix needed is very different than what it was a few years ago. And maybe one thing I worry about, I think I love the higher education system. Universities are a tremendous force for the good. And part of me, I am worried that AI is coming so fast that most universities struggle to adapt curricula at the pace that the technology is demanding. Because frankly, I don't wanna train people for the jobs of 2026. I wanna train people for the jobs of 2028 and 2029 and beyond. But sadly, many universities that aren't as privileged as Stanford, to be in Silicon Valley, highly connected. Many universities are still training people for the jobs of 2022, and that just doesn't work. But so how to help have AI solutions to bring the new skills to the students and support universities, support individuals to gain those skills that lets them use AI and do this exciting work. I think that still remains a huge challenge that we're working on.
