Babbage: Fei-Fei Li on how to really think about the future of AI

Hello.

This episode of Babbage is available to listen to for free, but if you want to listen every week, you'll need to be an Economist subscriber. And good news, we're running a Black Friday sale between Thursday the 23rd and Monday the 27th of November. You can sign up to Economist Podcast plus for half price. For just $2 or pounds or euros a month, you'll get access to all of our award winning podcasts. For full details, search online for Economist Podcast plus. The Economist. Chatgpt maybe you've heard of it.

If you haven't, then get ready.

Sometimes it can feel like artificial intelligence just appeared out of the blue a year ago.

This promises to be the viral sensation.

That could completely reset how we do things.

For most people, ChatGPT was a shock. This algorithm could produce sentences almost like a human. It was uncanny and perhaps a bit scary. Our guest today though, well, she saw it all coming. In fact, Fei Fei Li laid the critical groundwork for the technology that underpins large language models and image creation models, and which has accelerated progress in AI in the past decade. Fei Fei Li is a computer scientist at Stanford University and the founding co director of the university's Institute for Human Centered Artificial Intelligence. She made her name in the field as a pioneer in developing ways for computers to see and recognize objects. Computer vision is nowadays so good that it's embedded in everything from border checkpoints to to driverless cars, and it's even involved in helping us make better video calls. Computers have become even better than people at recognizing human faces, which has also led to a host of ethical questions on how this technology should be used in public. In developing the field of computer vision, though, Fei Fei Li also sparked something much bigger the use of enormous data sets as a way to train AI models in the mid-2010s that led to the revolution in deep learning. And deep learning is at the heart of the biggest AI models today. In our conversation and in her new book, the Worlds I See, Fei Fei Li talks about the challenges of her early work in computer vision, how it led to chatbots and the large language models we see today, and how we should all be thinking about the opportunities and the risks of the next generation of AI.

I worry about it a lot, especially as a technologist who actually understand what this technology is, and I think we have to take responsibility to govern it as well as to mitigate these risks.

I'm Alok Jha and this is Babbage from the Economist. Today, Fei Fei Li, a pioneer of AI on its Future. Fei Fei Li, thank you for joining me.

Thank you, Alec, for inviting me.

Fei Fei, I think it's worth talking about the current age of AI that everyone is fascinated by at the moment. And I'd love to hear your thoughts on generative AIs, the type of AIs that power things like ChatGPT and Bard from Google and others. They're all based on deep learning neural networks of the type that you know were really revolutionized by your work. And you must have seen the language models getting better and better as well. What was your reaction to the public release of ChatGPT in 2022? The world seemed to go crazy about it, but I wonder what you thought.

My world went crazy as well. So as a technologist, my colleagues and friends very deeply in the weeds of AI, we probably saw this a little earlier than the world. When GPT2 was released, we especially my colleagues here at Stanford, were recognizing that this way of training transformer models with huge amount of diverse data will give us the kind of powerful algorithm like ChatGPT eventually in ways we've never seen before. Of course, we didn't Predict the precise November 2022 would be the release. We didn't predict that, but we actually anticipated that by creating a center in my institute called center for Foundation Models to look at the different aspect of these algorithms. So from that point of view we did anticipate. But yet the public awakening, the inflection moment of the entire world in response to this technology was still a huge moment. It accelerated our work and concerns about how we talk about this, how we educate our students, to our policymakers, to our business executives, to our journalists and reporter friends. It accelerated our work in engaging with, from Washington D.C. to Brussels in terms of the policy and regulatory framework surrounding this technology. And really completely serendipitously, just a few months later, my book was shipped to the publisher. And I think it really was a moment that reinforced my own belief that no matter how powerful the technology is, like GPT, like technology, it has to be grounded in a human centered approach.

Fei Fei, you're well known for starting your career in the attempts to try and build computers that can recognize objects and scenes, this thing of computer vision. Now, just before we go into the technological aspects of that, what do you.

Think it means for a machine to.

See something and recognize something?

Yeah. I want to invite you to imagine what it means to see, period, 540 million years ago. The simple animals living under the water don't see. They don't have something like ice they just float around. They don't get to feel the world, they don't sense the world. The world is in kind of a darkness that they don't even know it's darkness. And then around 500 million years ago, they start to develop a little pinhole almost in their region where they would call head and start to collect lights. And that moment must be such a fascinating or even poetic philosophical moment of animal life. Right. When you get to receive information from the world. And some zoologists actually hypothesize that the ability to see as one of the oldest sensory capabilities of animals was what set off the evolutionary arms race of animal speciation. Because once you can see, you can see dinner. You can become somebody else's dinner.

You can avoid being dinner as well.

Exactly. You start hide, you start to run, you start to evolve. And the course of intelligence, the nervous system, started. So seeing is a cornerstone of intelligence. It's part of how we understand the world, part of how we survive. And by the time you get to humans, we are such visual animals, we use seeing as a way of learning, as a way of interaction, as a way of making things happen, as a way of socialization, as way of communication. So this is why vision fascinates me. Cracking the fundamental questions of vision is cracking the questions of intelligence. And that's why I specialized in computer vision.

And so intelligent systems, in your telling, have to be able to see because that's such an important part of how they behave or how they can survive.

Essentially, by and large, yeah.

So what does that mean in terms.

Of a machine then?

How do you build a system, a.

Computer system that can, quote, see?

Yeah. Well, one thing we should recognize is that cameras take pictures. And that's a very physical process of allowing photons to land on some kind of sensors. Used to be films, now it's digital to register color, to register illumination or grayscale lighting. And that is only a very small part of seeing. And we have that with our own eyes. We have retina that gets stimulated when we see lights and of different color. But real seeing happens in the computation of making sense of the sensory data. Otherwise they're just numbers or they're just electrical poles coming from the retina. But now I'm talking to you. The different colors and shades and pixels coming from you make me recognize. I'm talking to Alec. I'm talking to a person who's wearing a blue shirt. When you nod, I get very happy. And that is making sense and understanding of what we are seeing. And that is A non trivial feat because there are so many things, there are so many possibilities mathematically, you know, if you want to assemble pixels in different composition, it's infinite. Yet we actually see the world in such richness and beyond understanding. We actually interact with the world. When you make a sandwich this morning, you have to do so many things with your visual intelligence from, you know, opening the fridge, understanding where the ingredients are, bringing them out. You, you have to use your eye to guide your hand. If you close your eye, your hands will not be able to grab that piece of tomato or bread. And then you have to use a knife to spread the butter. And it just goes on and on. Right.

So you've been working on this for some time and I just wonder if.

You could just give me a brief.

History of how computers started to be able to see.

We began as a field in the 60s and 70s, trying to just make sense of very simple lines or simple geometric shapes. Yet as we move into the more ambitious realm of computer vision, we start to look at tackling the problem of recognizing objects. And that was really my PhD study. I remember the first data set I worked with had four different objects. Airplane, leopard face, and the car. You would ask me, how did you pick these four objects?

How did you pick these four objects?

First of all, I didn't. My PhD advisor said, Fei Fei, here's someone else's data set. And they made these four objects. And it's already hard enough to get a few dozen or a couple of hundred pictures. That's back 2000. It's really hard. Digital cameras were just about to take off. There was no smartphone. The Internet was barely born. Throughout my entire PhD, it was working on these very small datasets. And not only my PhD was small dataset, my entire field was working on this small data set.

In 2006, Fei Fei Li began to create a database of images, each one carefully labeled to describe what was in the picture. ImageNet, as the database was called, was part of a project to develop AI algorithms that could be trained to see the world and recognize objects. The purpose of ImageNet was to widen the set of reference images for the computer science community from four to many thousands.

ImageNet gets you to train algorithms to recognize tens of thousands of different kind of objects. It also is not only a data set that is ginormous, 15 million images across 22,000 classes of object. It's also a mind shift for how we train AI systems. ImageNet was the turning point of AI's history. That's recognizing how critical it is to use big data.

Every year, the ImageNet team would run a competition to see which algorithms were best at recognizing objects they'd never seen before. In 2012, when the competition was being held in Italy, Geoff Hinton, an AI researcher from the University of Toronto, entered his algorithm, Alexnet. It was so good that Fei Fei Li rushed from her home in California to watch the results.

Their algorithm was actually a classic class of algorithm. It was called neural network, or in this case, convolutional neural network. It was invented in the late 1980s, but they combined that with Imagenet with two GPUs and made huge progress in terms of reducing the error of the tests and winning the competition. And that was when I recognized that it was the watershed moment for AI's evolution.

Alexnet used a type of AI built on many layers of neural networks.

Every layer is consisted of little nodes that function like neurons, similar to neurons, mimics neurons. I would not say it replicates neuron. You begin with an image. The first layer of neurons take the values of the colors of each pixels of the image, and then all these neurons compute. And then you keep computing. And Alexnet has about seven layers. And after seven layers, it says, oh, after all this, I see cat or I see a microwave. And that's really what this neural network does. But what is really hard is how do you teach this neural network to grab information that's important for cats versus microwaves versus chairs?

The process was called deep learning.

So at the beginning, what you do is you feed this neural network, in this case, millions of images. Every image has a label, right? Some are cats, some are microwaves, some are chairs. And then, you know, you have to learn to differentiate cats from microwave from chairs. And if you're wrong, you know you're wrong because you know the ground truth label. And the algorithm sends back a signal and say you have to correct this parameter that gets you this value, and you do it many, many, many, many millions of times. And eventually you learn a network with enough mathematical values that do a decent job in capturing CATS information or microwave information. And then you use that to go about recognizing object in the world.

Deep learning would go on to become the basis for the AI models that have taken the world by storm in the past year.

Now, you talked about your history in starting with your colleagues, this idea of how to recognize objects. And it zooms all the way forward to, you know, smartphones that can recognize faces. We see supermarkets that don't need checkouts. Robots can be guided around factories. It's just really permeated so many parts of life and industry.

What do you think the most exciting.

Use for computer vision is today?

Yeah Alec, that's a great question. Just think about humans and animals. We use vision for everything. I genuinely think vision will empower and enable people and machines to do a lot of things eventually. I think if we look for what we call embodied AI, you know, robots and doing machines, I think vision is their cornerstone. Without making cars to see, we won't have self driving car. Without making cameras to see, we won't be able to guide, say surgeries without sensing in large scale, we won't be able to map out the biodiversity of our planet. We won't be able to go under the deep ocean for exploration. We won't be able to, you know, train algorithms to tell cells with pathology from healthy cells. So we really have a boundless possibility using vision. I myself am working in robots in robotic learning. I'm very excited by that. I'm also working in healthcare where we use camera sensors as guardian angels to help doctors and caretakers to watch our patients for safety and other clinically relevant conditions.

Can you give me a timescale for the embodied AIs that we are talking about? When robotics and everything will really start to integrate?

Yeah. You know Alec, I've been asked this question a lot and frankly I don't even think my own answers are necessarily consistent. It shows how hard it is to predict the future.

That's moving fast, right?

Yeah, it's moving fast and I absolutely believe it in my lifetime because that's why I'm working on it. I hope I live long.

Well, I'm sure you will. It's also possible that machine vision is going to have lots of perhaps worrying uses and in fact they're already here. Things like unwanted facial recognition in public or you know, using them to guide weapons, for example. I wonder, as one of the people who has made all of this possible, how much concern do you have about these sorts of negative uses? If you like, do you think about them at all?

Absolutely. That's why I wrote the book. That's why for so many years I've been advocating for human centered AI. Our relationship with AI is like our relationship with all tools. It's a double edged sword. Humanity, it's in our DNA to be innovative. We want to be innovative because we want to make life work better. Right from the, the moment we discover fire or use that first stone axe, we are inventing tools and now we're in the intelligent machine age. But our fundamental relationship with tools haven't changed is that we can use it for good and we can use it to do harm. Some harms are intentional and some harms are unintended, but still equally bad. So I worry about it a lot, especially as a technologist who actually understand what this technology is. I see how it can be wrongly used, how can be wrongly communicated, how can be harmful to people, and I think we have to take responsibility to govern it as well as to mitigate these risks.

We'll look at how to solve those challenges in AI with Fei Fei Li next. But first, just a reminder that this is a free episode of Babbage to listen to us every week, you'll need to become a subscriber if you already have a subscription to the Economist, you just need to link your podcast app. There's a link to a video explaining how to do it in the show Notes. If you're not yet a subscriber, you can make the most of our Black Friday deal and sign up to Economist Podcast Plus. For those of you that are subscribers, we've posted a little bonus for you in the Babbage feed. Every couple of weeks, we'll bring you one article from the printed edition of the Economist. Read aloud the story we've picked for you this week looks at how to improve scientific research. We argue that more experimentation is needed in funding mechanisms to supercharge science. Find that article in your podcast app. Today we're speaking to a pioneer of artificial intelligence, Fei Fei Li. Earlier in our conversation, she explained the need for transparency and better communication around new generative AI technologies. As we've explored before on Babbage, generative AI raises a lot of challenges for the society. Earlier this month, the British government hosted a summit on AI safety at Bletchley park, just outside London. This was where the country's mathematicians broke the German Enigma code during the Second World War.

28 countries, including the US and China, have signed up to the Bletchley Declaration, an attempt to kickstart a global approach to manage the very latest technology, amid warnings from Texas tech giants that it presents the greatest existential threat to humanity.

And in America, President Joe Biden has issued an executive order that sets out new standards and regulations for how AI should be built in the future.

Signing an executive order that would require developers working on the most powerful AI models to share their safety test results with the government.

There are a wide range of worries and risks about generative AI and computer vision. How do people like Fei Fei Li, who brought these technologies into the World think about their ethical challenges. Let's rejoin my conversation with her.

So, Alec, look, I think we all have responsibility. It's such a complex and powerful technology. It's impossible that there is one solution. To me, education and communication is a critical part of it. I don't think we have been doing enough public education or enough good public education. I don't think going after hyperbole is a helpful way of public education. This is a very nuanced technology, but we really do need to educate the public much, much better. On also the risks from disinformation to bias, to privacy concerns, to jobs changes. But that's one part of responsibility. We also have responsibility as developers and inventors of this technology. How do we consider the ethical values? How do we consider responsible and transparent development? And then we have responsibility in deploying this technology for business that's creating products and services. What do we bring to our consumers, to our community? And last but not the least, civil society and governments have the responsibility for governing, for guardrailing, for creating the right framework or ecosystem to use this to do good but avoid catastrophic risks that we might be dealing with because of the power of this technology.

What are your concerns about technologies like ChatGPT? There are lots and lots of things that they will do that will benefit, and we can talk about those. But when you see these technologies and the things they're capable of and then the conversation around them, what are the things that sort of worry you about their capabilities?

Yeah. So like you said, I do want to give a very nuanced approach. It has a lot of powerful, positive empowerment and applications and possibilities for people. But you're asking about the concerns. We talk a little bit about bias. Human civilization has been biased and now we have powerful machines that can propagate our historical bias. And that's all good. That would hurt people, especially people from underrepresented, underserved, undervoiced communities. Whether it's automated machine hiring algorithm or a skin disease detector, by AI or bank loan assessment, you can see that the buyers could potentially propagate. And that's just one bucket of catastrophic risk. There's another bucket, for example, disinformation. I'm deeply, deeply concerned. Of course, humans have been propagating disinformation since the beginning of civilization. But technology lowers the entrance bar, right? So you can now scale up so easily. You can now create fake videos, fake audios, fake podcasts, and that would be detrimental to a society that is built upon trust. There is a civil contrast of trust, of information and that would be detrimental. How do we combat that? And then there's jobs. Clearly there will be job disruption. There will be also job empowerment. This technology can increase productivity, but in the meantime, productivity doesn't necessarily translate to a fair distribution of prosperity and wealth in workforce and our society. Our history everywhere, globally has seen this movie over and over again. So how do we deal with this wave of technology shift? We can go on and on.

We asked some of our listeners to send in some questions about things that they were worried about. When it comes to large language models and the capabilities, there seems to be a lot of concern about hallucinations. I mean, these are the things where chatbots and AI models will just invent information to answer the questions. And one of our listeners, Andrew, a microbiologist in California, he asks, can large language models ever be fixed to mitigate the problem of hallucinations? Is it something that's just a technology problem or is there something deeper going on here?

That's a great question, Andrew. First of all, if you look at the progression of, say, OpenAI and other large language models, this problem is being combated. You know, every release of, say the GPT products has been better with hallucination. And a lot of this is what they call the reinforcement learning from human feedback training or post training that is trying to combat hallucination. But is this a pure technical issue to solve? I believe it's more nuanced. For example, it really does depend on data, right? So if you feed, let's say I say the sun is green and I feed the algorithm a lot of data about the sun is green, well, the algorithm is going to learn the sun is green. That is not technically hallucination because it is trained on that. So starting with data integrity, we have potential issues. Of course, because of the way the model is made, it might always hallucinate just out of the patterns it has learned. So we do also have to have a layer of responsible use. How do we communicate with users? How do we actually provide sources of information? These are important questions from product and service point of view.

I wonder what the next sort of iterations of generative AI look like. Are images and video going to be useful data sources for things like language models in the future? Are they going to be the thing that transforms it next time?

Absolutely, Alec. Actually, if you look at the most recent releases including GPT4V, it's already beginning to be multimodal. And a lot of this training have used the images and videos and there has been really interesting work in these Diffusion models in generative AI images. I mean, you know, you type a sentence like put me in a sushi restaurant and you actually can get pictures of sushi restaurants. So we're already seeing that multimodality.

Slightly terrifying, actually, in a way, yes. For your misinformation point, yes.

I'm not going to deny that actually, both language and now images and videos, as we talked about it, are definitely fertile grounds for mis and disinformation. But from technology point of view, multimodality is absolutely the future. And then the ability to translate the physical world into a digital one, and vice versa, is going to bridge the work of robotic learning with today's more software layer AI. Right. So there's a lot to come from technology point of view. But you're right though, every one of them comes with its own risk and potential misuse. And that is the concerning part.

Just one more point on this, which is generative AI from your point of view has so many advantages and there's so much to go. But you've pointed out the things like misinformation, disinformation, bias, these things are real concerns to people right now and should be regulated against and people should be aware of them. But so much of the conversation around generative AI and the future of AI is around these so called existential risks. You know, that's going to destroy the world somehow or it's going to turn us all into paperclips or whatever else. And even though there's no evidential pathway towards those things, people seem to be very concerned about them, including governments. And there have been AI safety summits in the uk. President Biden's issued his executive orders. The heads of these companies go to these government agencies all the time to talk about these things. I wonder what you think about this.

Did you think that people focus too.

Much on these catastrophic and existential things just because they're kind of terrifying to the detriment of worrying about the real risks that people are facing now?

I had this public debate with Professor Geoff Hinton, who's a personal hero of mine as a technologist and he definitely is one of the vocal people talking about the potential existential possibilities. And I actually respect that. In the meantime, a respectful philosophical pondering about the future of this technology doesn't substitute pressing urgent, social, catastrophic issues. We need to take responsibilities now to combat, and we have been saying things like bias, jobs, weaponization, disinformation. And there is another social issue that is very urgent and pressing is the lack of investment in public Sector development of this technology and public sector has always been a more trusted source of honest transparent assessment and evaluation of this technology. And not a single university in the US can train a ChatGPT model today. And with this kind of deprivation of public sector, we will have long term crisis in terms of training the right talents, in terms of depriving us from scientific discovery and depriving us from evaluation and assessment of this technology. So I actually think it's critically important, especially for governments to talk more about the immediate pressing social issues as well as creating ecosystem that would benefit the public. For example, investment in public sector or proper, hopefully not overly done, not under proper regulatory framework, rather than for the governments to engage in long term philosophical existential crisis rhetoric. With all due respect to policymakers, they are not trained to do that. We've got a lot of scholars and true deep thinkers that are probably more trained to do that. But the policymakers are tasked to help the public and make the right policies for their regions, countries and world.

And when you say investment by public sector bodies, do you mean trying to have models or clouds of computing that basically are available to public institutions, governments, universities, et cetera, so that not everything happens within companies?

Yes. So President Biden issued an executive order from the White House just days before the UK AI Summit talking about the kind of government measures for AI. Some of them were regulatory measures, but some of them were public investment. Most notably is a pilot program for the National AI Research Cloud that we call NAIR National AI Research Resource. I was actually personally involved with my Stanford colleagues in leading this effort four years ago, 2019. And we wrote an open letter about the need for Nair in 2020. And we worked on an early bill called the Nair Task Force Bill under Trump administration that made it happen. And then I served as one of the 12 task force members for the White House to plan for nair. And now we are also part of the leading voices from public sector to endorse the Create AI act bipartisan bill being considered in both US Congress and Senate.

Fei, Fei, I can't let you go without talking to you about the future of AI and what we should be thinking about and looking out for in the coming years. And I suppose one sort of usual journalistic way of dealing with this is to ask you about artificial general intelligence. This idea of a machine that can do all the things that a human brain can do, should we ever want to build on those things? You know, generative AIs have tricked us into thinking that there's something intelligent going on underneath the Hood, when in fact it's just statistical meanings of language and so on. You know, they've captured the public imagination. But do you think that these are the route to the mythical artificial general intelligence? Or do you think we're going to need more than that?

So Alec, from a technological advanced point of view, I think we are creating very powerful machines. And as we become more multimodal, as we create embody AI, robots and agents, we're creating machines that will be able to do a lot of what humans can do. I can't wait for a machine to help me to clean toilets and bathrooms. But in the meantime, human intelligence and human lives is actually very, very high dimensional, very layered, right? The kind of feelings we have for each other, the kind of creativity that sparkled our artists, our scientists, our writers, musicians, empathy, compassion. I have no way of even grasping as a technologist how to quote unquote, model that, how to write an algorithm to do that. And I think that will be an important question for each individual. If we create machines that can do some of the tasks humans do today, does that mean it's replacing the essence of being human? I actually don't think so. Just like you know, humans move, but we've created from bicycles to cars to rockets, it didn't take away our agency of moving. We created powerful tools that help us to move or even move us. So I think these are important questions to ponder. I'm not downplaying the important both side, the risks and possibilities of AGI. I think it's really worth talking about. I just want to make sure that we see ourselves with agency, with multidimensions, with depth and nuances. That is how we define who we are.

How about a sort of extreme question? Do you think that AIs could ever become sentient? Are they going to become living in some way or conscious in the way that you have in science fiction all the time?

This is not an extreme question. This is what I say that I respect this kind of discussion, I respect this kind of scholarly work. There's definitely already some level of cognitive understanding of language, let's say of these algorithms. But are they like sentient or aware or conscious in the way of humans? Absolutely not yet. Is there a way to create silicon based computational algorithm that will get to that level of self awareness? I remain open minded, I think there's a long way to go, but, but I think it's a worthy philosophical, scholarly study because we'll always be curious about it. Humans are curious. Humans go after these kind of problems and it's also important to study this because like we said, only when we know something, we know how to evaluate, assess and govern it.

I think that sounds like a good place to end. You seem like a very optimistic person when it comes to AI, whilst also acknowledging the troublesome aspects, which I think is probably the best way to be about it.

So Fay Fear Lee, thank you very.

Much for your time.

Thank you Alok.

And thank you for listening. If you want to keep up to date with the very latest goings on in the world of AI this week, then head over to our sister podcast Money Talks. My colleagues there will be looking at the business side of artificial intelligence and what the recent turmoil at A OpenAI means for the future of the industry. Money Talks comes out on Thursdays. Finally, don't Forget that as COP28 rapidly approaches, we'd like to know if you've got any climate related questions. Get in touch with us by emailing podcastseconomist.com and make sure you put Babbage in the subject line. You might even get your questions answered on a future show this week. Babbage was produced by Jason Hoskin and mixed by Johnny Allen. The Executive Producer is Marguerite Howell. I'm Alok Jha and in London this is the Economist.