Sana’s founder and CEO Joel Hellermark sits down with Stanford Adjunct Professor and AI Fund Managing General Partner Andrew Ng for a candid, wide-ranging conversation on the future of artificial intelligence. They explore why “AI coding is the new literacy,” the shift from AI’s “high priests” to democratized creation, and the untapped creative potential in the application layer. Andrew shares hard-won lessons from building Google Brain and Coursera, reflects on the economics and competition behind the language model boom, and offers research-driven insights on productivity, risk, and the path to accessible, responsible AI for all.
Watch the full episode above or read the transcript of their conversation below.
Historical analogies—AI as the new electricity
So excited to be here with you, Andrew. I was just telling you—it feels like I’ve spent way more time with you than you have with me. Growing up, taking all of your courses, you’ve really had a massive impact. I think over 8 million folks now have taken your courses. Is that right?
Yeah, thank you. I’m really thrilled and honored to hear that.
The past few years, it’s been such a massive shift. If you look at historical analogies, what do you think are the ones that best represent what’s going on right now?
I think AI is the new electricity, and one thing that’s still not widely understood is how much it is a general purpose technology, like electricity a hundred years ago. Whether it’s the last generation of supervised machine learning—labeling things, which worked really well in the last decade—or generative AI, which is really taking off in this decade, these are general purpose technologies. That means they’re not useful just for one thing—they’re useful for a lot of things.
So a lot of work lies ahead of us to figure out all those use cases and go and execute on them.
One hundred percent.
Layers of the AI stack: hardware to application
What are you most excited about?
If I look at generative AI, which is widely talked about, there’s activity at the hardware layer—Jensen Huang at NVIDIA is deeply thoughtful on that. There’s also the infrastructure layer, with companies like large language model companies and API providers doing fantastic work. But I think the amount of creativity and diversity in the application layer is really exciting. My teams have enjoyed taking AI and applying it to maritime shipping, or helping entrepreneurs build relationship coaching apps, or working in healthcare and education. I find that all corners of the economy, affecting people in all walks of life, are being touched by this technology. That application layer—there’s so much creativity, and frankly from a business point of view, it doesn’t feel nearly as competitive as the infrastructure or hardware layers. So there are just so many opportunities there.
Defensibility and moats in AI startups
So how do you think about it as an entrepreneur building on the application layer as these models become increasingly commercialized? What do you think are the new moats, or how should you build defensibility?
What I’m seeing is that there are startups building a very thin layer on top of very powerful APIs, which leads to good ideas that are not defensible. For example, last December, Lensa was an app where you could upload a handful of pictures of yourself and have it render a picture of you as an astronaut, a scientist, or something else. It was a fantastic idea and its revenues took off, but then they collapsed just as quickly. I think of an earlier revolution with the iPhone, when I paid $1.99 for an app to turn my phone into a flashlight. It was a great idea, but it wasn’t a defensible business. So I think in the large language model era as well, there are going to be thin application layers, but what I’m more excited about is that, just as with the rise of the iPhone, people figured out you could build Airbnb or Uber or Tinder—businesses that are difficult to build but very valuable. I think with the rise of generative AI as a technology, the opportunity is to build those really difficult but also really valuable products for lots of people. Deep learning started to work really well maybe ten or fifteen years ago, and we haven’t even finished figuring out the use cases for that technology yet. Now generative AI is another major tool with even more opportunities to figure out exciting use cases.
Economics of LLMs and cloud AI
How are you thinking—if you look at the more foundational layer, how should you think about the unit economics of these large language models? They have a fairly short shelf life and require a lot of investment. What are the unit economics on the foundational layer?
It’s been interesting. For example, when I look at the ChatGPT pricing—or more precisely, the API call for the most popular model, GPT-3.5-Turbo—if you assume someone reads 250 words per minute and you figure out the economics, it turns out I estimated it costs about eight cents to generate enough text to keep someone busy for an hour reading. So the economics are fantastically cheap in terms of the cost of an API call to generate a lot of text to keep someone entertained or usefully engaged for quite a long time. One of the things about cloud businesses is that they’re so efficient that to users it can feel pretty cheap to use a cloud service, even though cloud businesses are fantastically profitable. This is good for everyone building applications. And I do think that with Google’s release of PaLM 2 and different models like that—OpenAI, Anthropic, Amazon, and many other great companies—the large language model layer looks competitive. I wish all these companies luck, but that competition also creates a lot of opportunities for entrepreneurs building on the application layer.
The zero-cost knowledge work future
What do you think happens when the cost of knowledge work approaches zero? You were speaking of generating text, and that could similarly be applied to a lot of the knowledge work that we do, as the cost goes toward zero. What happens then?
I think certain parts of knowledge work are becoming very inexpensive. There have been a couple of fascinating studies I’ve read closely—one by OpenAI and one by my collaborator Daniel Rock at the University of Pennsylvania—showing that, unlike earlier waves of automation where lower-wage workers tended to be more exposed, this time it’s actually higher-wage workers who are more exposed to automation via large language models. One of my radiologist friends, Curt Langlotz, said that radiologists who use AI will replace radiologists who don’t use AI. I think it will be like that in a lot of professions: people who figure out how to use AI will have significant advantages over those who don’t.
Productivity, polymaths, and human augmentation
That’s such an exciting future, where you have this polymath in your pocket that knows all of the world’s knowledge, all of your knowledge, all of your company’s knowledge, and can help support you in real time. There have been studies published recently showing 30% to 80% productivity gains. What would we do with all of this extra time that we have?
So far, humanity has not run out of ideas to generate even more meaningful things for us to do. One of the rules I live by, and what I find really exciting, is the idea of making humanity more powerful. While it doesn’t always work out well, on average it works out well the vast majority of the time. I think large language models are making humanity more powerful. I use ChatGPT quite frequently myself, I play with bots and use Microsoft’s Bing Edge product, and I find all of these tools make me more productive as a person. It’s good to see hundreds of millions of people around the world all enjoying these benefits already.
I love this early story from Minsky and Engelbart, where they run into each other and Minsky was telling Engelbart about all of this stuff he would help computers achieve with AI. Then Engelbart replied, “You’re going to do all of this for machines; what are you going to do for humans?” I feel that we’ve reached a point now where we are getting radically augmented in our abilities, and this idea of augmented intelligence is really getting applied with ChatGPT more recently.
Real-world uses for generative AI
How do you use ChatGPT day-to-day?
I have many diverse, specific uses for it. Sometimes I’ll ask it a factual question and, depending on its answer, sometimes fact check it. I use it as a thesaurus all the time and for brainstorming when I’m trying to write something. I don’t use it that much for debugging code, but I see more friends doing that. The use cases are surprisingly diverse and it’s fantastically cheap. I also use it to test out different business ideas. I tend to use the API calls rather than the web user interface. Last weekend, I was playing around with it for many hours and I think I spent about twelve cents. Sometimes it’s just fantastically cheap to explore ideas in multiple different application verticals.
I wonder if you might be one of the best prompt engineers on the planet. It would be fun to have a prompt engineering competition. What makes a good prompt?
Several weeks ago, my team at Deeplearning.ai, working with Isa Fulford at OpenAI, released a free online class on ChatGPT prompt engineering for developers, to try to share best practices on how to write prompts. One thing I’m excited about—and I think its impact is vastly underestimated—is that, even though ChatGPT as a web interface and consumer product is widely discussed, large language models are incredibly powerful developer tools. I think the impact on how the next generation of machine learning or AI applications will be built is still vastly underestimated. That’s why our free online course, offered in collaboration with OpenAI, hopes to share more best practices with people.
Prompt engineering: skill or disappearing art?
Do you think that will be embedded natively into the models instead, that they’ll take our prompts and then rewrite those to better prompts—or do you think there’s a role for us to also become better prompt engineers?
I think there’s a huge role for people to play in becoming better at prompt engineering. I’m not sure prompt engineering itself will be a major job category—just as tuning parameters isn’t a job category for neural networks—but for many machine learning engineers, knowing how to use prompting as a tool will be an important part of their job. Machine learning engineers that can use this tool will be more effective than ones that don’t know how to use it.
Were you surprised that it ended up being natural language that we instruct these models with?
I think the rapidity with which it took off when ChatGPT was released took me by surprise, and it took a lot of people by surprise. But the academic research leading up to ChatGPT had been going on for a while. When GPT-3 came out, before ChatGPT, my colleagues in NLP and I felt, “Wow, this is exciting. This is a potential game changer for how to approach machine learning on text data.” Not in everything, but for text data, it was very exciting. Then, when OpenAI rolled out ChatGPT as a consumer product, it caught the popular imagination. But a lot of the sparks of that technology being exciting—like learning with very low data—were there before the consumer product.
What would you say has contributed the most to its success?
It was interesting when the UI of ChatGPT became so easy to use that people in all walks of life started using it. Unlike early waves of adoption, now we see many CEOs themselves using it. It wasn’t the IT person going to the CEO to tell them to check this out; it was the CEO checking it out themselves and then telling their IT department, “We need to get on this.” In fact, the Wall Street Journal talks about how Coursera’s CEO, Jeff Maggioncalda, has been playing a lot with ChatGPT to think about how it should affect education. I think it’s exciting that there’s so much CEO-driven action, and this is unlike most waves of technological adoption.
The future of education and AI
How do you think it will impact education? Do you think the medium will change radically? I remember the story from the early days of TV where we were essentially just shooting radio shows. I think that’s a bit analogous to the early days of MOOCs, where we were basically shooting classes. Do you think we’ll radically change the development of this content and how it’s created? Will you basically record yourself, synthesize it, feed all your knowledge into a model, and then have an agent, a teacher? Or what do you think will be the instruction model?
I don’t know. I think many organizations like Coursera are playing around with these ideas right now. There are some things that seem to work well, like having a coach so that you can have someone to ask questions, help quiz you, and help you learn along. But whether this will even more fundamentally change the educational experience remains to be seen. I think teacher’s assistants as well as personalized student coaches would be valuable, both to help teachers and to help individual learners. Beyond that, I and many others are excited to try to figure that out now.
Neuroscience, human learning, and AI
One thing that I was quite underwhelmed by when starting to get into the neuroscience of how we learn was how little we understand. Have you learned anything from studying it more from a machine learning perspective where you’ve thought, “I wonder if this is actually how humans learn as well?”
I think I’ve probably learned more about how humans learn by watching my kids—two and four years old—than through my work in machine learning. Ten or fifteen years ago, I was reading a lot of neuroscience papers because I wanted to understand a bit more of how the brain works and translate that to artificial neural networks. But since then, the whole field of AI has really pivoted away from looking at neuroscience, because as you pointed out, Joel, we have almost no idea how the human brain works. So trying to take inspiration has been a really tough thing to do. For the last decade, AI has been driven much more by engineering principles than by biological inspiration.
But I’m also quite intrigued by how well some fairly simple models from neuroscience, like reinforcement learning and artificial neural networks, ended up working. Do you think there are any areas that are under-explored?
Take artificial neural networks: the ones we use today are unlike anything that the brain does. They are a vastly oversimplified model of what the brain is. I actually suspect—though I can’t prove—that there are multiple algorithms that can lead to fantastic levels of intelligence if you feed them the right data. Human intelligence and neurons, our brains, took some local optimization path driven by evolution, and we seem to be pretty intelligent. But when I look at the algorithms we use, I feel like there’s got to be a better algorithm than deep learning or transformers or whatever. All the flavors of algorithms, when fed a lot of data, can do fascinating things. So I think there have to be many different algorithms.
Do you think there are better learning algorithms than the ones we have as humans?
Oh, I’d be shocked if there isn’t. Given the random local search process of evolution, the human brain still has lots of weird things left in our bodies. For example, we still have the remnants of a tail from when our ancestors had longer tails. So, humans have evolved to this pretty good local optimum, but I’d be completely shocked if the human learning algorithm were anywhere near the optimal learning algorithm. Scaling Up: The Limits of Current Models
And where do you think the cap is? Do you think there’s any cap in terms of the models that we use now, or can we just keep training them on larger amounts of data and expect the same correlation when it comes to the intelligence?
I think machine intelligence is turning out to be very different than human intelligence. What we saw in the last decade was that large supervised learning models with large labeled datasets could do fantastic things. This decade is turning out to be the decade of large pre-trained models. We can take a transformer model, train it on massive amounts of text, and then very quickly, through prompting or fine-tuning, adapt it to new tasks. Or take a vision transformer, train it on massive amounts of images, and then with a little bit of tweaking or prompting or fine-tuning, adapt it to different vision tasks. Is this a path to artificial general intelligence? Probably not. I’m pretty sure we still need new algorithmic inventions. But over the last year, we’ve made one year of fantastic, wildly exciting progress—maybe on what will turn out to be a 30- or 50-year journey toward AGI.
So you think it’s 30, 40 years out?
Maybe 50\. I think it’s decades out. But it’s okay, because the machines we have are so fantastically exciting and valuable.
What would change your time horizon on that?
I think we still have some fundamental breakthroughs in algorithms to make. Yann LeCun often points out that a teenager takes maybe 20 hours to learn to drive a car. When will we have an algorithm that can take dozens of hours to learn to drive a car? Or where someone can read a book and then answer lots of questions later? Or read a sequence of books and then write a PhD dissertation after deep knowledge? How do we get AI to do those things? I hope we’ll figure it out, but I just don’t see a clear path to doing all of those things yet.
AI safety, regulation, and existential risk
But listening to that, and also listening to your conversation with Yann, I take it you’re not scared. What do you think are the key arguments against the six-month halt?
I thought the proposal for a six-month pause in training AI models more powerful than GPT-4 was a bad idea for two reasons. One is the premise that it’s dangerous seems misguided. But also, a six-month pause is impractical. If you have a 30-year journey, how does a six-month pause help? And how do you even implement this unless governments start legislating pauses to technology, which seems very anti-competitive. Even leaving aside the impracticality of a six-month pause, I want to get at the premise that AI is an existential risk to humanity. Sometimes people think there could be a hard take-off: today AI doesn’t work, but one day it suddenly works and takes over the world in 48 hours. Technology doesn’t work like that; it advances slowly. I hope we’ll build AI more intelligent than me, than any person, someday, but it’s not going to happen overnight. We’ll see it coming for a long time. Humanity has lots of experience building and steering organizations more powerful than any single human—corporations, nation states. For the most part, democracy manages to steer most corporations and most nation states—not all, sadly—but in a pretty positive and constructive way. With AI becoming more and more capable, I’m confident we’ll do that too.
And what are the actual existential risks to humanity?
I don’t know if maybe someday an asteroid will strike the planet, or if climate change causes massive depopulation, or the next pandemic comes. I think AI will be a key part of the solution to these things. For humanity to survive and thrive for another thousand years, I think our odds are higher with fantastic AI than without. So if you want humanity to thrive and survive, let’s make AI go as fast as possible.
Are there any scenarios that you’re more scared about, more near-term?
Yes, I worry about AI fairness, bias, misinformation, polarization of society, and concentration of wealth and power. These are real issues, real risks of AI. The AI community is rapidly improving on most of these dimensions. Large language models are well known for sometimes making things up, hallucinating, or producing toxic or highly inappropriate speech. But they’re actually getting much safer. It’s not perfect, but it’s much better now than even a couple months ago. So I think AI is making rapid progress on all these things, but fairness, bias, and concentrated power are real problems. Excessive hype about evil sentient AI killer robots—the Terminator scenario—unfortunately distracts us from the real present risks that we should actually be spending time working hard to address.
Regulation, transparency, and open source
But one thing I’m also struck by is that it’s not as if any of those issues are new. We’ve already been dealing with misinformation for a very long time. If you were to give some tips for regulators in terms of where they should put their focus or double down on areas they’ve already been focusing on, what would those be?
I have a controversial point of view on this. I feel like a good set of regulations to get started would be to cast more transparency on AI. I see a risk of, to be candid, European regulators screwing it up. There was a recent draft regulation—with some very problematic things relating to release of open source large language models. But I do think there’s a lot to be said for transparency. Today, there are AI systems that are very black box, or corporations running AI systems that are very black box. Without transparency into, for example, whether an AI system amplified hate speech or contributed to problematic things, it’s impossible to craft thoughtful regulation. So the first move should be transparency, hopefully in a way that doesn’t stifle innovation or knock any region of the world back while others are rapidly moving ahead.
I’m sure you read the leaked memo from Google on open source as well. What role do you think open source will play in this? Is that actually the biggest risk? Because of course we can regulate the companies, but this is going to be much tougher to do with the open source community.
The closed-source large language models are ahead of the open source ones, but not by that much, one would argue, and the open source community collectively just has a lot more people working on this. The rate of progress is fantastic. While there are definitely risks to some of these technologies, so far I think the benefits vastly outweigh the risks. I’m not at all dismissing the risks—bias, fairness, AI safety, let’s take it seriously—but we have to appreciate the good and the massive value it’s creating even while we take a hard look at the risks and do our best to ameliorate them. When Stable Diffusion released image generation models in this very open way, it sparked off tremendous creativity and there are so many applications now possible because they did that.
The next frontier: literacy, democratization, and societal value
What do you think are the key things we should be learning over the next few years?
I feel like the technology for large language models and image generation—moving into video—increasingly, still feels very early and raw. When I chat to friends at OpenAI or with Harrison, CEO of LangChain, there are different things where it’s actually very obvious there will be significant improvement in the technology over the next few years. That’s exciting. And even as the technology races ahead, the opportunity to build more and more applications is also insanely exciting. So all that is a massive value creation. One thing I worry about is job displacement and people getting left behind. I think whatever we can do as a society to bring everyone along can only be helpful. But I’m also nervous about, as you were mentioning, regulators taking unwise actions that actually hinder large parts of the world rather than lift up citizens the way I hope regulators will.
That’s what I find very inspiring as well about democratization. It used to take quite a lot of time to get on with your courses and learning that. But now there’s this democratization where, if you understand natural language, you can start interacting with these models. You made a parallel to literacy a while back. Could you elaborate a bit on that parallel?
I see it as AI coding is the new literacy. Hundreds of years ago, we thought maybe not everyone needs to be literate—why not just go to the holy building and listen to the priest or priestess and the monks read the Holy Book to you? Just sit in the audience and listen. Fortunately, we figured out as a society that we’re collectively much richer if almost everyone can learn to read and write. I think today we’re still in that high priests and priestesses era in AI, where most people feel like, “I don’t need to build AI; let the high priests and priestesses of the tech world, the engineers and the big tech companies, build AI for us.” But with the availability of data in today’s world, everyone—from a corner store proprietor to a high school student collecting biology experiments—has data. If we can democratize not just the use of AI, but access to custom AI and custom creation of AI, then everyone can be in a better position to leverage the unique data they have. That will make society much richer.