Andrew Ng on winning the AI race

Note: Transcripts are automatically generated from episode audio, and are not fully corrected for spelling, grammar, and formatting.

RANA EL KALIOUBY: Hi there, listeners! Today, we are sharing another conversation from this year’s Masters of Scale Summit. Andrew Ng is a true AI pioneer. He’s co-founder of Coursera and DeepLearning.AI, and managing partner at AI Fund — a studio that incubates new AI companies. He was joined on stage by DJ Patil, former Chief Data Scientist under the Obama administration and General Partner at GreatPoint Ventures.

It’s a dynamic discussion about the current state of AI, how future generations should approach the technology, and how America can stay competitive in the global AI race. I can’t wait to share it with you.

Let’s jump in.

DJ PATIL: Alright, everyone. Well, I’m just gonna ask Andrew to come on out here because, well, please give a warm hand to Andrew Ng. All right.

ANDREW NG: Thank you, Jay.

PATIL: Okay.

NG: I was told there was a lot of heart in this audience.

PATIL: Yeah. There’s heart in this audience. And now we go to AI, ’cause no one’s heard any of this, right? About AI. But here’s the important thing — I’ve known Andrew for a long time. He’s been talking and working on AI long before it was cool.

I remember actually sitting down with Andrew when we were coming up with these ideas around data science and these things. And he was talking about AI. I was like, hey, we’re still in the middle of winter for AI. But he’s been on this for the incredible journey. But to just give you some of the highlights of Andrew, he was one of the first people to advocate for using GPUs for deep learning.

He wrote —

NG: Should have bought Nvidia.

PATIL: I was gonna say Jensen, how many jets does Jensen owe you?

NG: Yeah. I don’t know.

PATIL: Like, what size, right? Did he ever give you anything?

NG: I think he gave me a few GPUs. That was nice.

PATIL: GPUs. You got some GPUs out of it. All right. You wrote the first major online course on machine learning and AI, which led to Coursera.

NG: At the time.

PATIL: It was radical thinking to actually teach an online course, and it has helped over 10 million students. Is that correct?

NG: 10 million. Yeah. Thank you.

PATIL: Incredible accomplishment. Along with Jeff Dean, Greg Corrado, and Quoc Le, you started the Google Brain project.

NG: Google Brain.

PATIL: Google Brain, thank you.

Much more. But some of the things that you’re doing right now — you have a fund and a studio investing and building on AI, and you also have some of the most seminal cited papers in AI today.

NG: I have to admit, I probably have highly cited papers. I don’t track my citation count as often as I used to.

PATIL: Exactly. You leave that for everybody else or the AI systems, but there’s also something that you don’t know about.

How many people have heard about agentic systems these days?

Maybe I should ask it the other way around, just to highlight: guess where “agentic” came from?

It’s a little-known fact that Andrew was actually the guy who really came up with “agentic.”

Copy LinkHow agentic AI became the language for a major shift

Actually, let’s start with that. What’s the story behind agentic?

NG: So almost two years ago, I saw this rising trend in AI that a lot of people were excited about, but within the tech community, there was all this debate — with some people writing software saying it’s an agent, others saying, no, that’s not an agent.

It’s an agent, not an agent. I thought, this is a waste of time. Why don’t we, instead of having a binary agent-or-not-an-agent debate, just call it all agentic and stop arguing, and get on with the work. And so I actually kind of ran a campaign — I didn’t publicize it, but I did it anyway — to try to get more people to just adopt the word “agentic.”

What I didn’t realize was that a few months later, a bunch of marketers would get hold of this word and slap it as a sticker on everything in sight. And that helped the movement take off. But even though the hype has gone like that, I think the real value is really growing rapidly too. So that’s been exciting.

Copy LinkWhat agentic workflows can actually do better than simple prompting

PATIL: Well, let’s stay on that. I wanna do this in four stages. Let’s first talk about today — talking about today in the state of AI through the lens of one of the OGs. What is your honest take on what AI can and can’t do, especially through this lens of agentic?

Because I think we’re all struggling with all the marketing and buzz out there of what’s real and what’s not.

NG: A lot of the work that lies ahead is to take these amazing agentic AI capabilities and map them to real business workflows. I think we’d be doing that, but what’s actually —

PATIL: What is agentic, like? Let’s ground us there.

NG: So a lot of us use AI — large language models — by prompting it and asking it to write an output.

That’s a bit like going to a human and saying, please write an essay by just typing it out from the first word to the last word, all in one go, without stopping to think, without ever using backspace. Humans don’t do our best writing like that, and neither does AI. With an agentic workflow, the idea is that we can ask AI to take a more iterative approach and say: first write an outline, then do some word research, then write the first draft, then critique it.

And so the iterative workflow takes much longer, but for a lot of tasks — from medical advice, legal advice, tariff compliance, writing code — for a lot of different things, these agentic workflows work much better. But there’s still a lot of work ahead of us. I know that some people say, oh, don’t worry about that, wait for AGI — they’ll solve all the problems. I’m not a fan of this “wait for AGI” thinking; that feels like hype to me. A lot of the work that’s very valuable right now is to take the technology and what may be possible in the next six to 12 months, and just go do valuable stuff with it.

Copy LinkHow to judge when AI is reliable enough for real work

PATIL: What’s the analogy — sometimes the analogy I think of is I’m using these systems because I’m trying to build, I’m trying to deploy AI to really help senior citizens in their healthcare journey, among other things. And sometimes I think: am I on thick ice? Am I on thin ice with what the systems can and can’t do?

I suspect a lot of people out there are wondering, does it work for this problem or is it fragile? Sometimes we just get to an 80% solution. Other times it knocks it out of the park. Sometimes we’re incredibly disappointed. How do you, as Andrew — who’s building companies, advising so many people — think about this?

NG: Yeah, it is tough. The closer a task is to only text processing, and if you have the plumbing to get all the information — hopefully in text — that people need to do a task, the easier it is for AI to do it. If you need to feed in images, voice conversations, it’s not impossible, but it gets harder.

And then one question I often ask is whether we have the data plumbing to give the AI system similar context as the human would need to do that task. And then for a lot of multi-step processes, if you can write a standard operating procedure — an SOP — it’s worth seeing if we can codify the SOP in a multi-step agentic workflow.

So it’s hard to determine what can and cannot be done, but I think these are maybe some suggestions for rating what’s more or less likely to succeed.

EL KALIOUBY: We’ll be back with more from DJ and Andrew, after a quick break.

[AD BREAK]

Copy LinkWhy learning to code matters more in the age of AI

PATIL: Well, let’s switch gears and go to education, because you’ve really changed and transformed academia through Coursera and your current online courses. I can’t even keep track of how many views and how many people are taking your courses.

And I wanna start with — I’m sure every parent is asking you what their kid should do to be prepared for AI. Should kids still learn to code? Is CS a thing? Is data science a thing, or is that a bad idea?

NG: One of the most important skills for the future is the ability to tell a computer exactly what you want so it can do it for you. For the foreseeable future, people who know the language of computers — people who understand coding — do that much more effectively than people who don’t. So I know that earlier this year, there were some leaders who were advising others not to learn to code on the grounds that AI will automate it. I think we’ll look back on that as some of the worst career advice ever given.

I’m already seeing on my teams, and on a lot of Silicon Valley teams, not just the software engineers but the marketers, HR professionals, analysts, finance professionals — the ones that know how to code are starting to run circles around the ones that don’t. So if your kid intends to be a software engineer, have them learn to code with AI.

And even if they don’t, it is becoming clearer that in the future, we need a lot more not just users of software, but creators of software. So rather than growing up and asking, is there an app for that? — I want them to say, I built an app for that. And with AI assistance, coding is much easier than it used to be.

So don’t code by hand — get AI to do it for you. People who do that will be more powerful and more effective than people who don’t. There’s this new skill, just like today I can’t imagine going through college without learning how to do a web search — that’s kind of weird, it limits your job prospects. I think if you go through college without knowing how to create software, that’s kind of weird. It will limit the prospects of what someone can do.

Copy LinkWhen children should start using AI and where parents need guardrails

PATIL: When we think about when a kid should get access to this technology, what does that look like? And specifically through the lens of some of the lessons I think we’ve really started to struggle with around social media — Surgeon General Vivek Murthy really highlighted the challenges that have been happening around that. When is it too early in your view? When is it the appropriate time for someone to really start having access to AI, to make sure they’re truly AI-native and get the maximum benefits of this technology?

NG: I think it’s difficult. When kids get access to books, we think really, really young. But there are also some books that are clearly inappropriate for a 2-year-old. And I think one of the challenges with technology is that there are apps that are just fine for a very young child to use, but there’s also a lot of stuff we would not let a young child use.

My kids are four and six. They do use tablets occasionally, but I’m there with them when they’re using it — not using it as a babysitter, but having them do educational things, or doing things together that we talk about. So I think the medium has changed, but the challenge is whether the business incentives for companies push them to create or not create certain experiences for kids.

And as parents, how can we have guardrails to curate the helpful things? Just like I don’t let my kids read certain highly inappropriate books for their age, I don’t let them do certain highly inappropriate things for their age. But it is a challenge, given the incentives of certain types of companies to do things that we as parents may not want them to do ever.

PATIL: Have you talked to some of those companies or groups, and said, hey, knock it off, that’s not gonna be helpful? What’s that conversation like? Because there are a lot of people who’ve taken your classes that are actually doing some of the behaviors that I think, as parents, we find really problematic.

NG: The vast majority — 99% — of engineers and business people in Silicon Valley want to do the right thing.

The people doing these things are, frankly, our friends, maybe some people in this room right now. I think everyone kind of wants to do the right thing. And I wish we could find a way, when there are billions of dollars at stake, to still always do the right thing. It is a real problem. We do see a small number of people that will sometimes do not quite the right thing when the financial incentives are big enough. I wish I knew how to solve the problem of human incentives.

Copy LinkWhat US policy must fix to stay competitive in AI

PATIL: Well, let’s switch to something easier. US policy. Speaking of children — statements from a former US Chief Data Scientist. You were early in advocating for GPU policy, and you were one of the first people I saw that were really working with international technology companies. Given some of the places where you sit and get to see across the landscape —

We’ve seen this whipsawing on GPUs from federal policy. We’ve seen executive orders talking about what some describe as “woke systems.” We’ve also seen executive orders and policies trying to accelerate the adoption of AI. So if you had five minutes with the president, what advice would you give them to both a) responsibly unleash the power of AI to benefit all Americans, and b) ensure national competitiveness?

NG: I’m really worried about US national competitiveness in AI. Some things that the current administration has done well — I think the previous administration had some AI safety thinking that was really safety theater driven by lobbyists, fear-mongering to try to create regulatory capture, anti-open-source regulations. If you don’t wanna compete with open source, make up a bunch of stuff about the dangers of AI to try to get stifling licensing requirements. So I think the current administration seems to have very low patience for that. That’s good.

Things that worry me: both of us are immigrants.

PATIL: We have been immigrants.

NG: A lot of our students are immigrants. I really worry about American competitiveness if we make it harder for high-skill immigrants. When I came to the United States as an undergrad student, I was like 17 years old.

I was frankly pretty clueless. I don’t think I was particularly exceptional at all when the US let me in. So letting students come to the US so they can grow up and hopefully become higher-skill — I really worry about restricting that. I think defunding of science, decreased investments in science, AI, and other things — universities do have issues, let’s be candid, there are things we could fix — but diminishing the ability of this country to execute science, I really worry about that.

And then in terms of national policy, I also worry about our reliance on TSMC. It’s been interesting to see — China recently banned certain imports of Nvidia chips, a strong signal that China is moving toward independence from TSMC in Taiwan, at a moment when the US is becoming still heavily reliant on Taiwan manufacturing. One of the implications of this is if anything were to happen in Taiwan — either a natural disaster or a manmade event — disruption to the Taiwan semiconductor ecosystem could end up hurting the US much more than it hurts China, if China becomes more independent of Taiwan manufacturing than the US. So semiconductors are one concern.

And then lastly, AI semiconductors are one bottleneck, but the other big bottleneck — which is totally true and widely reported — is energy. When you build a data center, that’s a machine to turn electricity into intelligence, or turn electricity into output tokens.

And so the constraint I hear from so many friends is being hung up in permitting. Can we build a power plant here? There are permitting issues, local objections — which may be valid. But I think energy capacity is the other bottleneck that everyone is talking about.

Copy LinkWhy open models and global competition will shape the AI landscape

PATIL: Let’s think about China specifically, and let’s take China and Europe — two very radically different approaches to AI regulation. And then you sort of watch the US trying to navigate this for competitiveness. What I’m curious about from your lens is this idea of both the open source models and whether people build off of the main trunk of AI — which is US-based primarily right now, but increasingly there are new models out of China, a little bit out of Europe. What’s the right strategy here, or what’s a strategy you believe is best and optimal for the world for AI? Is it a central trunk of AI with branches, or is it many different trees in the forest — a federation of different models and approaches?

NG: I think we need multiple branches, because otherwise — to make an analogy to mobile phones — the mobile ecosystem is kind of uninteresting. It’s because there are two gatekeepers, Android and iOS. Unless they let you do certain things, you’re just not allowed to experiment. So I hope AI will not end up with a small number of gatekeepers that can limit innovation.

What has happened is that over the past year or two, China has really pulled ahead of the US in releasing open-weight models. These are models that can be downloaded and used for free, and I saw a stat showing that the cumulative adoption of Chinese open-weight models is about to surpass — or may already have surpassed — the cumulative adoption of US open-weight models.

The US closed models are still better, but open-weight models are a key part of the AI supply chain, and people are using them. I think there’s a problem that we aren’t investing enough as a nation in that. And then you asked about Europe — honestly, I love Europe. I wish Europe would wake up and get going faster.

For a while over the last few years, visiting Chinese and European regulators, I heard things like, “we wanna be leaders in regulating AI.” I don’t think that’s how you gain a competitive advantage.

PATIL: More brakes, more brakes, less gas — we win the race.

EL KALIOUBY: More in a minute. Stay with us.

[AD BREAK]

Copy LinkHow AI can earn public trust by making more people more capable

PATIL: Let’s turn to the future, in these remaining couple of minutes, and really talk about the future of AI — but I wanna talk about it through the lens of the cutting-edge next generation of students that you see, the entrepreneurs. What are the problems that you see them gravitating to?

What are their hopes, their dreams, and specifically what does that tell you about how the next 24 months look?

NG: We’re in Silicon Valley, where most of us love AI. I love AI. I love what I do. I think of it as making the world better. I think many of us may underestimate the distrust that a lot of people across the nation have for AI, and I feel urgency to act together to make sure that we can craft a compelling narrative — explain why AI is actually really good for the world.

It turns out we all get excited about productivity improvements, but when a contact center worker is scared of losing their job, when a fast food worker hears a politician say, “yep, guess what? These AI people, they’re gonna make your job go away” — that creates a lot of fear and distrust of AI that we don’t really see here in Silicon Valley. So I think to win people over, we need to make sure technology genuinely benefits everyone broadly.

And I think there is a path to that. AI can make individuals much more effective and much more productive. But to get the tools available to everyone, to teach people how to use them — that’s the upskilling, improve the tools. With the concept of a 10x engineer, I think with AI we can have 10x marketers, 10x analysts, 10x finance professionals. But to actually make that happen, there’s a lot of work ahead of us, and I worry that we have not yet won the trust of a lot of people in this country.

Copy LinkUsing AI as a thinking partner and why builders should act now

PATIL: What’s your favorite way that you use AI today?

NG: Gosh, maybe I’ll share one that is not widely known. I use AI as a —

PATIL: Where everyone’s like, go on.

NG: I use AI as a brainstorming companion, much more than even my friends know. And the trick is — it turns out AI —

PATIL: One model? Do you use multiple?

NG: Multiple models. Multiple models.

PATIL: For a friend.

NG: Actually, for coding I love Claude Code, and I’m increasingly using OpenAI Codex as well. But for brainstorming, I use multiple models. It turns out the trick is that AI is very smart, but getting context in is difficult.

And so when brainstorming, I find a lot of it is not just, let me say some stuff and give me ideas — it’s making sure you have an extended conversation. Voice or text are good; either one. When I’m driving, voice. And when I’m driving, I talk to AI quite a lot, and then I’ll say, summarize it for me, and I’ll send it to my team and just get work done.

PATIL: In the final 10 seconds. What’s a problem that you would wish people would focus on more using AI?

NG: Actually, go and build stuff. I think every one of you — this is a wonderful time to build.

So if there’s one thing you take away from what I believe in, just go and build stuff. There’s so much cool stuff you can now build that just was not possible before. So build.

PATIL: I think that is a perfect way to end on “build.” Andrew Ng, ladies and gentlemen, thank you for your work. Andrew, thank you for your research. Thank you for being here.

NG: Thank you.

EL KALIOUBY: Andrew and DJ’s conversation shows how AI is an ecosystem. For the US to win the AI race, it will take a multi-faceted approach. It’s chips and AI infrastructure, yes. But it’s also talent. We must continue to attract top global talent — and offer opportunities to scholars and innovators from around the world to build a successful life and career here.

You can find the full video of this and more from the Summit Stage at the Masters of Scale YouTube channel.