Geoffrey Hinton in conversation with Fei-Fei Li — Responsible AI development

Well, good afternoon everyone. I love the buzz that is here in the room today. Welcome to the MaRS Discovery District, this wonderful complex for this very special radical AI founders event, co-hosted by the University of Toronto. My name is Meric Gertler and it is my great privilege to serve as president of the University of Toronto. Before I begin, I want to acknowledge the terrain in which the University of Toronto operates. For thousands of years, it has been the traditional land of the Huron-Wendat, the Seneca and the Mississaugas of the Credit. Today, this gathering place remains home to many indigenous peoples from across Turtle Island and we are very grateful to have the opportunity to work and gather on this land.

Well, I'm really delighted to welcome you all to this discussion between Geoffrey Hinton, University Professor Emeritus at the University of Toronto, known to many as the godfather of deep learning, and Fei-Fei Li, inaugural Sequoia Professor in the Sciences. Computing at Stanford University, where she is co-director of the Human-Centered AI Institute. I want to thank Radical Ventures and the other event partners for joining U of T to create this unique and special opportunity. Thanks in large part to the innovative work of Professor Hinton and his colleagues, the University of Toronto has been at the forefront of the academic AI community for decades.

Deep learning is one of the major advances driving the rise of AI, and many of its key

development

s were pioneered by Professor Hinton and his students at U of T. This tradition of excellence, this long tradition, continues to this day. present. Our faculty, students, and graduates, along with partners at the Vector Institute and universities around the world, are advancing machine learning and driving innovation. Later this fall, our faculty, staff, students, and partners will begin moving into phase one of the beautiful new Schwartz Reisman Innovation Campus across the street. You may have noticed a rather striking building on the corner with the official plan to open for early next year.

This facility will accelerate innovation and discovery by creating Canada's largest university innovation centre, made possible by a generous and visionary gift from Heather Reisman and Gerry Schwartz. The innovation campus will be a focal point for thought leadership in AI and will house both the Schwartz Reisman Institute for Technology and Society, led by Professor Gillian Hadfield, and the Vector Institute. It is already clear that artificial intelligence and machine learning are driving innovation and value creation across the economy. They are also transforming research in fields such as drug discovery, medical diagnostics and the search for advanced materials. Of course, at the same time, there is growing concern about the role AI will play in shaping the future of humanity.

So today's

conversation

clearly addresses an important and timely topic, and I am very pleased that you all have joined us on this momentous occasion. Without further ado, let me introduce you to today's moderator, Jordan Jacobs. Jordan is Managing Partner and Co-Founder of Radical Ventures, a leading venture capital firm supporting AI-based companies here in Toronto and around the world. Previously, he co-founded Layer 6 AI and served as co-CEO prior to its acquisition by TD Bank Group, which he joined as Chief AI Officer. Jordan serves as director of the Canadian Institute for Advanced Research and was among the founders of the Vector Institute, a concept he came up with with Tomi Poutanen, Geoff Hinton, Ed Clark and a few others.

Distinguished guests, please join me in welcoming Jordan Jacobs. (Audience applauding) - Come on, up. Thank you very much Meric. I wanted to start by thanking several people who helped make this possible today, the University of Toronto and Meric, Melanie Woodin, Dean of Arts and Sciences, and several partners who made this possible. This is the first in our annual four-part series of AI Masterclasses for Founders that we teach at Radical. This is the third year we've done it and today is the first this year. We do it in person and online. We have thousands of people watching this online.

So if you decide you need to start coughing, maybe go outside. We do this in partnership with the Vector Institute and we are very grateful for their participation and support with the Alberta Machine Intelligence Institute in Alberta and with Stanford AI, thanks to Fei-Fei. So thank you all for being great partners. We hope this is a really interesting discussion. This is the first time Geoff and Fei-Fei, who I like to consider friends, that I can talk to, but it's the first time they've done this together publicly. So I think it's going to be a really interesting

conversation

.

Let me quickly do some deeper explanations of your background. Geoff is often called the godfather of artificial intelligence. He has won the tour award. He is a professor emeritus at the University of Toronto, co-founder of the Vector Institute, and was also a mentor to many of the people who have become leaders in AI globally, including at large companies and in many of the world's top research labs. world. world in academia. So when we say godfather, it really is, there are many types of Geoff's children and grandchildren who lead the world in AI and that all goes back to Toronto.

Fei-Fei is the founding director of the Stanford Institute for Human-Centered AI and a professor at Stanford. He is an elected member of the US National Academy of Engineering, the National Academy of Medicine, the American Academy of Arts and Sciences. During a sabbatical from Stanford in 2017/18, she took on a Google VP role as chief AI/ML scientist at Google Cloud. There are many, many other things we could say about Fei-Fei, but she also has an incredible number of students who have become leaders in this field worldwide. And what's really important, and for those of you who haven't heard yet, Fei-Fei has a book coming out in a couple of weeks.

It's called, it's coming out on November 7th and it's called "The Worlds I See, Curiosity, Exploration and Discovery at the Dawn of AI." I've read it, it's fantastic. Everyone should go out and buy it. I'll read you the back cover that Geoff wrote because it's much better than what he could say about it. Here is Geoff's description. "Fei-Fei Li was the first computer vision researcher to truly understand the power of big data, and her work opened the floodgates for deep learning. She offers an urgent and clear account of the astonishing potential and danger of AI technology that she helped spark and her call to action and collective responsibility is desperately needed at this crucial moment in history." So I urge you all to pre-order the book and read it as soon as it comes out.

That said, thank you Fei-Fei and Geoff for joining us. - Thank you, Jordan. (audience applauds) - Okay, so I think it's not an exaggeration to say that without these two people, the modern era of AI does not exist, certainly not in the way it is developing. So let's get back to what I think is the big bang moment. AlexNet ImageNet, maybe Geoff, do you want to tell us from your perspective that moment 11 years ago? - Well, in 2012, two of my very smart graduate students won a competition, a public competition, and showed that deep neural networks could perform much better than existing technology.

Now, this would not have been possible without a large data set with which to train them. Until that point, there had not been a large dataset of labeled images, and Fei-Fei was

responsible

for that dataset. And I'd like to start by asking Fei-Fei if there were any problems putting together that data set. (Audience laughs) - Well, thank you Geoff, thank you Jordan and thank you University of Toronto for this, it's a lot of fun to be here. So yes, the data set that Geoff mentions is called ImageNet. And I started building it in 2007 and spent the next three years practically with my graduate students building it.

And you asked me if there were any problems building it, where do I start? (Fei-Fei laughing) Even at the conception of this project I was told that it was really a bad idea. I was a young assistant professor. I remember it was actually my first year as an assistant professor at Princeton and, for example, a very respected mentor of mine in the field, if you know the academic jargon, these are the people who will write my tenure evaluations, actually told me Really, good-hearted, please don't do this after I told you what this plan was in 2007. - So that would have been Jitendra, right? (Audience laughs) - The advice was, "You might have trouble getting tenure if you do this." And then I also tried to invite other collaborators and no one in machine learning or AI wanted to even approach this project and of course there was no funding. - I'm sorry. (audience laughs) - Just describe ImageNet for people who aren't familiar with what it was. - Yeah, so ImageNet was conceived around 2006, 2007, and the reason I conceived ImageNet was actually twofold.

One is that, and Geoff, I think we share a similar background, I was trained as a scientist, for me, doing science is chasing the North Stars. And in the field of AI, especially visual intelligence, to me, object recognition, the ability of computers to recognize that there is a table in the image or a chair is called object recognition, it has to be a problem of North star in our field. And I feel like we really need to make a dent in this problem. So I want to define that North Star problem, that was an aspect of ImageNet.

The second aspect of ImageNet was recognizing that machine learning was really kicking around at the time, that we were building really complex models without the kind of data to drive machine learning. Of course, in our jargon, it's really the problem of generalization, right? And I recognize that we really need to reset and rethink machine learning from a data-driven point of view. So I wanted to go crazy and create a data set that no one has ever seen in terms of quantity, diversity and all that. So ImageNet, after three years, was a curated dataset of images from the Internet totaling 15 million images in 22,000 concepts, object category concepts.

And that was the data set. Just for comparison, at the same time in Toronto we were creating a dataset called CIFAR-10 that had 10 different classes and 60,000 images, and it was a lot of work, generally paid handsomely by CIDAR at five cents an image. - If you turn the data set into a competition, just tell us a little bit about what that meant and then we'll fast forward to 2012. - Right. So we built the dataset in 2009. We just turned it into a poster at an academic conference. And no one paid attention to him. So I was a little desperate at that point.

And I think this is the way to go. And we opened it up, but even with open source, it wasn't really getting better. So my students and I thought, well, let's push the competition a little further. Let's create a contest to invite the global research community to participate in this object recognition problem through ImageNet. So we did an ImageNet competition and the first comment we got from our friends and colleagues was that it's too big. And at that moment you can't put it on a hard drive, much less in memory. So we actually created a smaller dataset called ImageNet challenge dataset, which has only 1 million images in 1000 categories instead of 22,000 categories, and I think it was released in 2010.

You guys noticed that in 2011, did you? TRUE? - Yes. - Yes. - And in my lab we already had deep neural networks that worked quite well for speech recognition. And then Ilya Sutskever said: "What we have should really be able to win the ImageNet competition." And he tried to convince me that we should do that. And I said, well, you know, it's a lot of data. And he tried to convince his friend Alex Krizhevsky, and Alex wasn't really interested. So Ilya pre-processed all the data to get it just the way Alex needed it. - You reduced the size of the images. -Yes.-Yes.-He reduced the images a little. - Yes, I remember. - And they pre-processed it just for Alex, and finally Alex agreed to do it.

Meanwhile, in Yann LeCun's lab in New York, Yann was desperately trying to get his students and postdocs to work on this data set. Because he said: "The first person to apply convolutional networks to this data will win." And none of his students were interested. Everyone was busy doing other things. And so Alex and Ilya kept going and we discovered, by participating in the previous year's competition, that we were doing much better than the other techniques. And then we knew that we were going to win the 2012 competition. And then there was this political problem, that we thought that if we showed that neural networkswin this competition, the Computer Vision people, Jitendra in particular, would say, well, that just shows that it's not a very good data set.

So we had to get them to agree in advance that if we won the competition, we would have proven that neural networks worked. I actually called Jitendra and we talked about the data sets we could use. And my goal was to get Jitendra to agree that if we could make ImageNet, then neural networks would really work. And after some discussion and him telling me to do other data sets, we finally agreed, okay, if we could do ImageNet then we would have proven that neural networks work. Jitendra remembers when he suggested ImageNet and he was the one who told us to do it, but it was actually a bit the other way around.

And we did it and it was amazing. We obtained just over half the error rate of standard techniques. And the standard techniques have been perfected over many years by very good researchers. - I remember the standard technique at that time, the year before it was the support vector machine.with shortage. - Good. - That was, so you sent the results of your competition, I think it was at the end of August or beginning of September. And I remember getting a phone call or an email one afternoon from my students who were running this because we have the test data that we were running on the server side.

The goal is that we have to process all the entries to select the winners and then, I think it was in early October of that year that the International Computer Vision Fields Conference, ICCV 2012, was held in Florence, Italy. We already booked a workshop, annual workshop at the conference. We will be announcing the winner, it is the third year. So a couple of weeks before we will have to process the equipment. Because it was the third year and, frankly, the results of the previous two years did not excite me, and I was a nursing mother at the time.

Then I decided not to go to the third year, so I didn't book any tickets. I'm too far away from myself. And then the results came, that night, a phone call or an email came, I don't really remember. And I remember saying to myself, damn it, Geoff, now I have to get a ticket to Italy. Because I knew it was a very significant moment, especially with a convolutional neural network, which I learned as a graduate student, as a classical algorithm. And of course, at that time there were only middle seats in economy class for flying from San Francisco to Florence with a stopover.

So it was an exhausting trip to go to Florence... Sorry. - But I wanted to be there. (Audience laughs) Yes, but you didn't come. - No (audience laughs) Well, it was an exhausting trip. - But did you know that this would be a historic moment? - Yes, in fact I did. - You did it and you still haven't come. But you sent Alex. - Alex, yes. - Yes. - Who ignored all your advice? - Who ignored my email several times, because I thought, Alex, this is great, please do this visualization, this visualization. He ignored me. But Yann LeCun came and it was because those of you who have attended these academic conference workshops tend to book these smaller rooms.

We booked a very small room, probably just the middle section here. And I remember Yann had to stand at the back of the room because it was so full, and Alex finally showed up because I was so nervous that he wasn't even going to show up. But as you predicted in that workshop, ImageNet was under attack. In that workshop there were people openly attacking, this is a wrong data set. - In the room? - In the room . - During the presentation? - In the room. -But not Jitendra, because Jitendra already accepted that she counted. - Yes, I don't think Jitendra was in the room, I don't remember.

But I remember it was a very strange moment for me because as a machine learning researcher I knew the story was brewing and yet ImageNet was under attack. It was a very strange and exciting moment. And then I had to get in the middle seat to go back to San Francisco because the next morning. - So you mentioned some people I want to come back to later. So, Ilya, founder and chief scientist of OpenAI, and Yann LeCun, who later became head of AI at Facebook (now Meta), and there are other interesting people in the mix. But before we move forward and look at what created that boom moment, let's back up a bit.

They both started this with a very specific goal in mind, which is an individual and I think an iconoclast, and they had to persevere through the moments that you just described, but throughout their careers. Can you just come back, maybe Geoff and start, give us the background of why you wanted to get into AI in the first place? - I did psychology when I was a student. I didn't do very well. And I decided that they would never discover how the mind works unless they discovered how the brain works. That's why I wanted to discover how the brain worked and I wanted to have a real model that worked.

So we can think of understanding the brain as building a bridge. There are experimental data and things you can learn from experimental data, and there are things that will do the calculations you want, things that will recognize objects. And they were very different. And I think of it as you want to build this bridge between the data and the competence, the ability to do the task. And I always saw myself starting at the end of the things that work, but trying to make them more and more like the brain, but still work. Other people tried to stick with things justified by empirical data and try to have theories that might work.

But we're trying to build that bridge and not many people were trying to build a bridge. Terry Sejnowski was trying to build a bridge from the other end and we got along very well. A lot of people trying to do computer vision just wanted something that worked, they didn't care about the brain. And a lot of people who care about the brain wanted to understand how neurons work and so on, but they didn't want to think too much about the nature of the calculations. And I still see it as we have to build this bridge by getting people who know the data and people who know what works to connect us.

So my goal was always to do things that could do vision, but do vision in the way that people do it. - Okay, so we'll come back to that because I want to ask you about the most recent

development

s and how you think they relate to the brain. Fei-Fei, so Geoff just to set a framework for where you started, from the UK to the US and Canada, mid to late '80, you get to Canada in '87, along that route, funding and interest in neural networks. , and the way the approach you are taking is like this, but I would say mainly like this: -He went up and down. - Fei-Fei you started your life in a very different place.

Can you explain a little bit about how you got into AI? - Yes, I started my life in China, and when I was 15, my parents and I came to Parsippany, New Jersey. So I became a new immigrant and where I started were the first English as a second language classes, because I didn't speak the language and I only worked in laundries, restaurants, etc. But he had a passion for physics. I don't know how it got into my head. And I wanted to go to Princeton because all I knew was that Einstein was there, and I got into Princeton, he wasn't there when I got into Princeton. - You're not that old. - Yes.

But there was a statue of him. And the one thing I learned in physics, beyond all the math and all that, is really the audacity to ask the craziest questions, like the smallest particles in the atomic world, or the limit of space-time and the beginning of the universe. . And along the way I discover the brain of Roger Penrose, a third year student, and those books. Yes, you may have opinions, but at least I've read those books. - It was probably better that you didn't. (Audience laughs) - Well, at least I was interested in the brain. And when I graduated I wanted to ask the boldest question as a scientist.

And for me, the most fascinating and bold question of my generation of the year 2000 was intelligence. So I went to Caltech to get a double PhD, practically double, in neuroscience with Christof Koch and in AI with Pietro Perona. So I echo what you said, Geoff, about bridge, because those five years allow me to work in computational neuroscience and look at how the mind works, as well as work on the computational side, and try to build that computer program that can imitate the human brain. So that's my journey, it starts from physics. -Well, then his journeys intersect at ImageNet 2012.-By the way, I met Geoff when he was a graduate student. - Well, I remember, I used to go visit Pietro's laboratory. -Yes.-In fact, he offered me a job at Caltech when I was 17 years old. - You would have been my advisor. - No, I wouldn't, not when I was 17 years old. - Oh, it's fine. - Okay, so we come across ImageNet, I mean, in the field, everyone knows that ImageNet is this big explosion moment, and after that, the big tech companies come in and basically start buying your students and you, and incorporate them into companies.

I think they were the first to realize the potential of this. I'd like to talk about that for a moment, but moving forward a bit, I think it's only since ChatGPT that the rest of the world is catching up to the power of AI. Because you can finally play with it. You can experience it, in the boardroom you can talk about it and then go home, and then the 10 year old just wrote an essay on dinosaurs for fifth grade with ChatGPT. So I think that kind of transcendent experience of everyone being able to play with it has been a big change.

But in the intervening 10-year period, there's a sort of explosive growth of AI within big tech companies, and everyone else doesn't realize what's going on. Can you tell us about your own experience? Because you experienced a kind of ground zero after ImageNet. - It's hard for us to get into the picture that everyone else doesn't realize what was happening, because we realized what was happening. So it took a long time for a lot of the universities that you would have thought would be at the forefront to realize that. So MIT, for example, and Berkeley, I remember even speaking at Berkeley, I think, in 2013, when AI was already having a lot of success in computer vision.

And then a graduate student came up to me and said, "I've been here for like four years and this is the first talk I've heard about neural networks. They're really interesting." - Well, they should have gone to Stanford. - Probably, probably. But the same thing happened with MIT: they were strongly opposed to having neural networks. And the ImageNet moment started to wear on them and now they are big proponents of neural networks. But it's hard to imagine now, but around 2010 or 2011 there were the Computer Vision people, very good Computer Vision people who were very firmly against neural networks.

They were so against it that, for example, one of the main magazines, the IEEE PAM recognition... PAM? - PAM. I had a policy of not refereeing papers on neural networks at one point. Just send them back, you don't referee them, it's a waste of time, it shouldn't be in PAM. And Yann LaCun submitted a paper to a conference where he had a neural network that was better at identifying and segmenting pedestrians than the state of the art. And it was rejected. And one of the reasons he was rejected was because one of the referees said, "This doesn't tell us anything about the vision." Because they had this vision of how computer vision works, which is to study the nature of the vision problem, formulate an algorithm that will solve it, figure out how to implement that algorithm, and then publish a paper.

In fact, it doesn't work. I have to defend my field, not everyone. Not all. - So there are people who are... - But most of them were firmly against neural networks. And then something remarkable happened after the ImageNet competition, that is, they all changed in about a year. All the people who have been the biggest critics of neural networks started creating neural networks, much to our chagrin, and some of them did it better than we did. So this (incomprehensible) thing at Oxford, for example, created a better neural network very quickly. But they behaved as scientists should behave, which is to say that they firmly believed that this was all rubbish.

Thanks to ImageNet, we were eventually able to prove that it wasn't and then they changed. That was very comforting. - And just to continue, what you're trying to show, you're trying to label using the neural networks, these 15 million images precisely, you have them all labeled in the background so you can measure them. The error rate when he did it fell from 26% the year before, I think, to about 16%. - Yes. - I think it's 15.3. - Well. And then it continues... 15.32. (audience laughing) - I knew you would remember. - What randomization? - Geoff doesn't forget. And then in the following years, people are using more powerful neural networks and it continues to decline to the point where it exceeds... 2015.

There is avery intelligent Canadian student who joined my lab, his name is Andrej Karpathy. And one summer he got bored and said, "I want to measure how humans are doing." Then you should go read his blog. So he had all these images of humans having test parties, I think he had to bribe them with pizza. with my students in the laboratory. And they got to an accuracy of about 5%, and that... Was it five or 3.5? - Three. - Three. 3.5 I think. - So humans basically make mistakes about 3% of the time? - Well well. And then I think in 2016, I think a resounding thing happened. - Yes. - Yes, it was resounding, that year's winning algorithm surpassed human performance. - And finally you had to withdraw the competition because it was much better than the humans they had- - We had to withdraw because we ran out of funding. - OK well.

It's a different reason. - A bad reason. -I still run out of funding-Instantly that student began his life at the University of Toronto. -Yes.-Where he went to his laboratory and then he was head of research at Tesla. -Well, first of all, he came to Stanford to be a doctoral student. And last night we were talking, in fact there was a revolutionary speech, in the middle of this. And then he became part of the founding team of OpenAI. -But then he went to Tesla. -And then he went to Tesla. -And then he thought better of it. - He is back.

But I do want to answer your question about those 10 years. - Well, there are a couple of new developments on the way. - Good. - Transformers. - Good. - So the article on Transformer is written, the research done, the article written within Google, another Canadian is a co-author there, Aidan Gomez, who is now the CEO and co-founder of Cohere, who I think was 20 years old. He was an intern at Google Brain when he was a co-author of the paper. So there is a tradition of Canadians participating in these developments. But Geoff, you were at Google. When the article was written, was there an awareness at Google of how important this would be? - I don't think there was, maybe the authors knew it, but it took me several years to realize how important it was.

And at Google people didn't realize how important it was until BERT used transformers, and then BERT got a lot better on a lot of natural language processing benchmarks for a lot of different tasks. And that's when people realized that transformers were special. - So in 2017 the article about transformers was published. I also joined Google and I think you and I met in my first week. - Good. - I think most of 2017 and 2018 was searching for neuroarchitecture. - Good. - I think that was Google's bet. - Yes. - And a lot of GPUs were used. So it was a different bet. - So just to explain that neural architecture search essentially means this: you get a lot of GPUs and try a lot of different architectures to see which one works best and automate it.

It is basically an automated evolution for neural network architectures. - It's like a new hyperparameter. - Yes. - Yes. - And that led to... - Good way. - Pretty big improvements. - Yes. - But nothing like transformers. And transformers were a huge improvement for natural language: the neural architecture search was primarily done on ImageNet. - Yes. - Yes. - Then I will tell you about our experience with the transformers. So we were doing Layer 6 of our company at the time, I think we saw a preview of the document and we were in the middle of a fundraising and a bunch of acquisition offers and we read the document.

And not only me, but also my partner who studied with you and Maksims Volkovs, who left the group's laboratory. And we think this is the next version of neural networks, we should sell the company, start a venture fund and invest in these companies that will use transformers. So we estimated it would take five years to achieve adoption beyond Google. And then from that point on, it would be 10 years until all the software in the world would be replaced or integrated with this technology. We made that decision five years and two weeks before ChatGPT came out. So I'm glad to see that we were good at predicting, but I have to give credit to my co-founders, who I thought understood what the paper was about, but were able to fully explain it. - I should correct you, I don't think Tomi has ever studied with me.

He wanted to come study with me, but a colleague in my department told him that if he came to work with me, it would be the end of his career and that he should go do something else. -So he took the classes, and this is my partner who in the late 90s was doing a master's degree at U of T and wanted to go study with Geoff, he studied neural networks. And his girlfriend, now the father of his wife, who was an engineering professor, said, "Don't do that, neural networks are a dead end." So instead, he took classes but wrote his thesis on cryptocurrency. (Audience laughs) Okay, so...

Are you still going to talk about the 10 years? Because I think there is something important. - Yes, then go ahead. - I think there is something important that the world overlooked in these 10 years between ImageNet, AlexNet and ChatGPT. Most of the world sees this as a 10-year technology, or we see it as a 10-year technology, in big tech things are brewing. I mean, it took sequencing after sequencing of the transformer, but things are brewing. But I do believe that, for me personally and for the world, it is also a transformation between technology and society. In fact, I think I personally went from scientist to humanist in these 10 years.

Because after joining Google for those two years in the middle of the transformer papers, I'm starting to see the social implications of this technology. It was the post-AlphaGo moment and very quickly we got to the AlphaFold moment. That was where the bias was spreading, there were privacy issues. And then we started to see the beginning of misinformation and disinformation. And then we're starting to see conversations about work within a small circle, not within a big public discourse. That was when I felt personally anxious, I think, you know, 2018. Uh-oh, it was also right after Cambridge Analytica. So, that huge implication of technology, not AI per se, but algorithm-driven technology in elections, was when I had to make the personal decision to stay at Google or return to Stanford.

And I knew the only reason I was going to come back to Stanford was to start this human-centered AI institute to really understand the human side of this technology. So I think this is a very important 10 years, even though it's not in the public's eyes, but this technology is starting to infiltrate the rest of our lives. And of course, in 2022, it will all be seen in the light of day how profound this is. - There is also an interesting footnote about what happened during that period, where you, Ilya and Alex eventually joined Google, but before that there was a large Canadian company that had the opportunity to gain access to this technology.

You love us? I've heard this story but I don't think it's ever been shared publicly. Maybe you want to share that story for a second? - Okay, so the technology that we were using for ImageNet, we developed it in 2009 to do speech recognition, to do acoustic modeling, a little bit of speech recognition. So you can take the sound wave and make something called a spectrogram, which simply tells you at each moment how much energy there is at each frequency. You're probably used to seeing in spectrograms. And what you would like to do is look at a spectrogram and guess which part of which phonamion is expressed in the middle frame of the spectrogram.

And two students, George Dahl and another student I shared with Gerald Penn called Abdo, had a longer name, we all called him Abdo, who was an expert in speaking, George was an expert in learning. During the summer of 2009, they created a model that was better than what 30 years of speech research, and very, very large teams working on speech research, had been able to produce. And the model was a little better, not as big as ImageNet's gap, but it was better. And that model was later ported to IBM and Microsoft; George went to Microsoft and Abdo went to IBM, and those big speaking groups started using neural networks at that time.

And I had a third student who had been working on something else, named Navdeep, Navdeep Jaitly. And he wanted to take this speech technology to a big company, but he wanted to stay in Canada because of complicated visa reasons. So we contacted Blackberry, RIM and told them we have this new way of doing speech recognition and it works better than existing technology and we would like a student to come to you over the summer and show you. how to use it so you can have the best voice recognition on your cell phone. And they said that after some discussions, a rather tall Blackberry said, "We're not interested." So our attempt to give it to the Canadian industry failed.

And then Navdeep took it to Google, and Google was the first to put it into a product. So in 2012, around the same time we won the ImageNet competition, George and Abdo's acoustic model for speech recognition, the acoustic model was available, there was a lot of work to make it into a good product and get it to have low latency, etc. . which came out on Android. And there was a time when Android suddenly became as good as Siri at speech recognition and that was a neural network. And I think for high-level people in big companies, that was another ingredient.

They saw that it had this spectacular result for vision, but they also saw that it was already available in a product for speech recognition that worked very well there as well. So I think that combination makes the speech, makes the vision, clearly it will do it all. - We will not say more about Blackberry. - That was a shame. It was a shame the Canadian industry didn't do it; I think we would still have had Blackberries if that had happened. (Audience laughing) - Very well, we'll leave it there. (Audience laughs) I thought it was a story, I had heard this story before, but I thought it was important for the rest of the world to know some of what was going on behind the scenes, why this technology didn't stick.

Canada even though it was offered for free. Okay, so let's move on. Now we have post transformers, Google is starting to use them and develop them in different ways. OpenAI, where his former student Ilya had left Google, was one of the founders of OpenAI with Elon Musk and Sam Altman, Greg Brockman and a few others. Ilya is the chief scientist and Andrej is his student as co-founder. So, you're working together in a very small team to basically take turns, well, initially the idea was that we would build AGI, artificial general intelligence, eventually the transformers paper comes out, you start adopting transformers at some point and you start making tremendous progress internally. , they don't actually share publicly what they're capable of doing in understanding language and a host of other things.

They had efforts in robotics that emerged. Pieter Abbeel ended up creating Covariant, a company in which we later invested and other things. But this is how the linguistic part advances, and advances, and advances. People outside of OpenAI don't really know to what extent this is going on. And then ChatGPT came out on November 30th of last year. So 10 months ago. - Well, GPT-2 caught the attention of some of us. In fact, I think when GPT-2 came out, my colleague Percy Liang, an LP professor at Stanford, I remember him coming to me and saying, "Fei-Fei, I have a completely different understanding of how important this technology is." . es." So, to Percy's credit, he immediately asked HAI to set up a center to study this.

And I don't know if this is controversial in Toronto, Stanford is the university that coined the term basic models, and some people call it LLM (big language model). But going beyond the language, we called it the foundation model. We created the research center for the base model before, I think before 3.5 came out. Definitely before ChatGPT. basic model for those who are not familiar - That's really a great question. Base model, some people feel that it has to have a transformer. I don't know if you use... No, it just has to be a lot of data. - Very large, pre-trained with a large amount of data.

And I think one of the most important things about a basic model is the generalization of multiple tasks. You're not training it, for example, machine translation. machine translation is a very important task, but the basic type of model like GPT is capable of doing machine translation, it is capable of doing conversations, summarization and blah, blah, blah. That is a basic model and we are seeing it now in multimodality. We are seeing a vision, in robotics, in video, etc. So we created that. But you're right, the public saw this in... 10 months ago. - What did you say? - 30th of October. - November, I think. - November. - Another very important thing about basic models, which has been in cognitive science for a long time, the general opinion was that these networksNeural neurons, if you give them enough training data, they can do complicated things, but they need a lot. of training data.

They need to see thousands of cats. And people are much more efficient statistically. That is, they can learn to do these things with much less data. And people don't say that much anymore because what they were really doing was comparing what an MIT undergraduate can learn to do with a limited amount of data to what a neural network starting with random weights can learn to do with a limited quantity. of data. - Yes, that is an unfair comparison. - And if you want to make a fair comparison, you take a basic model which is a neural network that has been trained with a lot of things and then you give it a completely new task and ask it how much data it needs. learn this completely new task?

And that's called short-time learning because it doesn't require much. And then you find out that these things are statistically efficient. That is, they compare quite favorably to people in the amount of data they need to learn how to perform a new task. So the old nativist idea that we come with a lot of innate knowledge, and that makes us far superior to these things, you just learn everything from the data. People have pretty much abandoned that now because you take a basic model that had no innate knowledge but a lot of experience and then you give it a new task and it learns quite efficiently.

You don't need large amounts of data. - You know, my PhD is on one-shot learning, but it's very interesting, even in the Beijing framework you could do pre-training, but it's only in the kind of neural network pre-training that can really give you this multitasking. . - Good. - Well, this basically happens on ChatGPT, the world experiences it, which was only 10 months ago, although for some of us it seems... - It seems longer. - Much longer. - It seems like an eternity. - Because suddenly you have this, you had this big bang that happened a long time ago and I think for a long time no one really saw the results, suddenly, I mean, my comparison would be that there are planets that form, and stars that are visible, and everyone can experience the results of what happened 10 years before, and then transformed, etc.

So the world suddenly gets very excited about what I think for a lot of people is magic. Something they can touch and experience and that will give them feedback in whatever form they request. Whether they include text prompts and ask for an image, video, or texts to be created, and ask for more texts to answer things you might never expect and get those unexpected answers. So it feels a little like magic. My personal opinion is that we have always moved the finish line on AI. AI is always what we couldn't do, it's always the magic. And as soon as we get there and say that's not AI at all, or there are people around who say that's not AI at all.

We move the goal line. In this case, what was your reaction when it came out? I know part of your reaction was that you left Google and decided to do different things, but when you first saw it, what did you think? - Well, as Fei-Fei said, GPT-2 made a great impression on all of us. And then there was a steady progression, I had also seen things within Google before GPT-4 and GPT-3.5 that were as good as PaLM. So that in itself wasn't much of an effort. It was more that PaLM impressed me within Google because PaLM could explain why a joke was funny, and I always used it as, we'll know you really get it when you can explain why a joke is funny. fun.

And PaLM could do that. Not for all jokes but for many jokes. - And then- - By the way, these things are now pretty good at explaining why jokes are funny, but they're terrible at telling jokes, and there's a reason they generate text one word at a time. So if you ask them to tell a joke, what they do is try to tell a joke. So they will try to tell things that seem like a joke. So they say, a priest and a badger walked into a bar and that sounds a bit like the beginning of a joke and they go on telling things that sound like the beginning of a joke.

But then they get to the point where they need the punchline. And, of course, they haven't thought about the future, they haven't thought about what the punchline will be. They're just trying to make it sound like they're leading up to a joke, and then they give you a pathetically weak punchline, because they have to come up with some punchline. So while they can explain jokes because they can see the whole joke before saying something, they can't tell jokes, but we'll fix that. - Well, then I was going to ask you if the comedian is a job of the future or not.

Do you think soon? - Probably not. - Alright. - Anyway... - So, what was your reaction? And again, you've seen things behind the scenes along the way. - A couple of reactions. My first reaction is that of all people: I thought I knew the power of data, and I was still old because of the power of data. That was a technical reaction. I thought, damn, I should have created a bigger ImageNet. No, but maybe not, but that was really... - You still could. - Financing is the problem. Yes, that was the first thing. Secondly, when I saw the moment of public awakening to AI with ChatGBT, not just the moment of GPT-2 technology, I generally thought, thank goodness we have invested in human-centered AI for the last four years .

Thank God we have built a bridge with political leaders, with the public sector and with civil society. We haven't done enough, but thank goodness that conversation had started. We were participating, we were leading a part of it. For example, we, as a Stanford Institute, led a critical national AI research cloud bill that is still moving through Congress right now. - Not really now. - Senate, Senate, it's by chamber, so at least it's moving the Senate because we predicted the social moment for this technology. We don't know when it would happen, but we knew it would happen and honestly, it was just a sense of urgency.

I feel like this is the moment when we really have to live up to not only our passion as technologists, but our responsibility as humanists. - And so you two, I think the common reaction of both of you has been: we have to think about both the opportunities of this and its negative consequences. - So for me, there was something that I realized that I didn't realize until very late, and what got me much more interested in social impact was, as Fei-Fei said, the power of data. These big chatbots have seen thousands of times more data than any person could ever see.

And the reason they can do that is because you can make thousands of copies of the same model, and each copy can look at a different subset of data, and they can get a gradient of how to change its parameters, and then they can share all of those gradients. So each copy can benefit from what all the other copies extracted from the data, and we can't do that. Suppose you have 10,000 people and they go out and read 10,000 different books, and after everyone has read one book, everyone knows what's in all the books. We could become very intelligent that way, and that's what these things do and makes them far superior to us. - And there is education.

There are some studies we are trying to do, but not in the way. - Yes. But education is useless, that is, it is hardly worth paying for it. (Audience laughs) - Except the University of Toronto and Stanford. (audience laughs) - I've tried to explain to my friends that Geoff has a very sarcastic sense of humor and that if you spend enough time with him, you'll understand. But I'll let you decide if that was sarcastic. - So the way we exchange knowledge, generally speaking, is a kind of simplification, but I produce a sentence and you figure out what you have to change in your brain, so you could have said that, that is if you trust me.

We can also do that with these models. If you want one neural network architecture to know what another architecture knows, which is a completely different architecture, you can't just give it the weights. Then you get one to imitate the other's result, that's called distillation and that's how we learn from each other. But it is very inefficient, it is limited by the bandwidth of a sentence, which is a few hundred bits. Whereas if you have these models, these digital agents that have a trillion parameters, each of them analyzing different bits of data and then sharing the gradients, they're sharing a trillion numbers.

So you are comparing the ability to share knowledge that is in billions of numbers with something that is in hundreds of bits. They are just much, much better at sharing than we are. - I guess Geoff... I agree with you on the technological level, but it seemed like that was the moment that made you feel very negative. - At that moment I thought, we are history, yes. - Yes, I am less negative than you. I'll explain it later, but I think that's where... Well, let's take a second, let's talk about it. Explain why you are optimistic and let's understand why you are more pessimistic. - I am a pessimist because pessimists are usually right. (Audience laughs) -I also thought he was pessimistic.

We have this conversation. So I don't know if I should call myself an optimist. I think... Look, when you came to a country when you were 15, now you speak one language and you start from $0, there's something very pragmatic about my way of thinking. I think technology, our human relationship with technology, is much more complicated than an academy would normally predict, because we come to academia in the ivory tower, we want to make a discovery, we want to build a piece of technology, but we tend to be purist. But when a technology like AI hits the ground and reaches the societal level, it inevitably becomes messily entangled with what humans do.

And this is where you might call it optimism, it's my sense of humanity. I believe in humanity. I believe in not only the resilience of humanity, but also the collective will; the arc of the story is sometimes uncertain. But if we do the right thing, we have a chance, we have a great chance to create a better future. So what I'm really feeling right now is not delusional optimism, but a sense of urgency of responsibility. And one thing, Geoff, I think I really hope you feel positive is that you look at the students of this generation, in my class I teach 600 undergraduate classes every spring on the introduction of deep learning and computer vision.

This generation, compared to five years ago, is very different. They come to our class not only wanting to learn about deep learning transformers, AI generation, but they want to talk about ethics, policies, understanding privacy and bias. And I think that's where I really see humanity rising to the occasion. And I think it's fragile. I mean, look at what's happening in the world, in Washington, it's very fragile, but I think if we recognize this moment, there is hope. - Then I see the same thing. - Oh God. - I no longer teach university students, but I see it in younger teachers. - Yes. - So, at the University of Toronto, for example, two of the brightest young professors went into anthropology to work on alignment.

I hope Roger Grosse comes back again. And AI, for example, now works full time on alignment. - Yes. - So there really is a big change now, and I think it is unlikely that I will have ideas that will help solve this problem, but I can encourage these younger people around 40 years old. - Thank you. - To work on these ideas and they're really working on them now, they're taking it seriously. - Yes, as long as we get the brightest minds, like many of you, looking at this problem in the audience and online, that's where my hope comes from. - So, Geoff, you left Google largely so you could go and talk about this freely in any way you wanted.

And basically... Actually, that's not true, that's the media story and it sounds good. I left Google because I was old and tired and wanted to retire and watch Netflix. (Audience laughs) And I had the opportunity in that moment to say some things that I had been thinking about liability, and not having to worry about how Google would respond. So it's more like that. - If we have time we will return to the Netflix recommendation. - I was going to say. - But in the meantime, but you came out and started talking quite significantly... - Yes. - In the media.

I think you've both probably talked to more politicians in the last eight months than in your previous lives, from presidents and prime ministers, to congress, parliament, etc. Geoff, could you explain what your concern was, what you were trying to do? achieve by expressing it and do you think it has been effective? - Yes, people talk about the risk of AI, but there are many different risks. So there is a risk that it will eliminate jobs and not create as many jobs. And then we will have a whole underclass of unemployed. And we should care a lot about that because the productivity boost that AI will bring will not be shared with the people who lose thejobs.

The rich will get richer and the poor will get poorer. And even if you have a basic income, that is not going to solve the problem of human dignity for many people who want to have a job to feel like they are doing something important, including academics. And that is a problem. Then there is the problem of fake news, which is quite a different problem. Then there's the issue of battle robots, which is also quite a different issue. All the big defense departments want to make battle robots, no one will stop them and it will be horrible.

And maybe eventually, after we've had a few wars with battle robots, we'll get something like the Geneva Conventions, like we did with chemical weapons. It wasn't until after its use that people were able to do anything about it. Then there is the existential risk. And the existential risk is what worries me. And the existential risk is that humanity will be wiped out because we have developed a better form of intelligence that decides to take control, and if it becomes much smarter than us. So there are many hypotheses here. It is a time of enormous uncertainty. You shouldn't take anything I say too seriously.

So if we do something much smarter than us because these digital intelligences can share much better, so they can learn much more, we will inevitably get those smart things to create subgoals. So if you want them to do something to achieve it, they'll figure it out, well, you have to do something else first. For example, if you want to go to Europe, you have to get to the airport. That is a secondary objective. So they will set subgoals and there is a very obvious subgoal: if you want to do something, get more power. If you get more control, it will be easier to get things done.

And so anything that has the ability to create subgoals will create the subgoal of gaining more control. And if things much smarter than us want to take control, they will, we won't be able to stop them. So somehow we have to figure out how to keep them from wanting to take control. And there is some hope. These things didn't evolve, they're not nasty competitive things. They are however we make them, they are immortal. So with a digital intelligence you just store the weight somewhere and you can always run it again on other hardware. So we have truly discovered secret immortality.

The only problem is that it is not for us, we are mortal. But these other things are immortal. And that might make them a lot nicer because they don't worry about dying and they don't have to... Like the Greek gods. - Well, they look a lot like the Greek gods, and I have to say something that Elon Musk told me. This is Elon Musk's belief that yes, we are the kind of bootloader for digital intelligence. We are this relatively dumb form of intelligence that is smart enough to create computers and artificial intelligence, and that will be a much smarter form of intelligence.

And Elon Musk believes that he will keep us around because the world will be more interesting with people in it than without it. Which seems like a very fine thread to hang your future on. But it is related to what is said Fei-Fei, it is very similar to the model of the Greek god, that the gods have people around them to have fun with. - Okay, can I comment on that? - Yes. (Audience laughs) - Nothing you said was controversial. - Yes, no, not at all. That's why I want to discard your four concerns: economy, work, misinformation and weapons, and then extinction.

Greek Gods - - (indistinct) discrimination and prejudice. - Okay, then I want to divide them into two buckets. The extinction of the Greek god is the bucket of extinction, everything else I would call catastrophic. - Yes, simply catastrophic. - Catastrophic danger. And I want to comment on this, I think one thing I really feel is that my responsibility as someone in the AI ​​system, the ecosystem, is to make sure that we're not talking hyperbolically, especially with public policy makers. Geoff, with all due respect, extinction risk is a really interesting thought process that academia and think tanks should be working on. - That's what I thought for many years, I thought it was very far in the future and having philosophers and academics working on it was great.

I think it's much more urgent. - It could be, but this process is not limited only to machines. We humans are in this messy process. So I think there are many nuances. For example, we talk about nuclear energy. I know nuclear is much more limited, but if you think about nuclear, it's not just the theory of fusion or fission or whatever. It's really about getting uranium or plutonium, the systems engineering, the talents and all that. I'm sure you saw the movie "Oppenheimer." So here, if we go down that path, I think we have a better chance of fighting than fighting because we are a human society.

We are going to put guardrails, we are going to work together. I don't want to paint the picture that tomorrow we're going to have all these robots, especially in robotic form, in physical form creating the machine lords. I really think we need to be careful here, but I don't disagree with you that it's something we need to think about. So this is the extinction cube. The catastrophic risk, I think, is much more real. I think we need the smartest people and the more, the better to work. So just to comment on each of them, the use of weapons, this is really real.

I completely agree with you. We need an international partnership, we need potential treaties, we need to understand the parameters. And this is about humanity, as much as I am optimistic about humanity, I am also pessimistic about our ability to self-destruct and destroy each other. So we have to get people working on this, and our friend Stuart Russell, and even a lot of the AI ​​experts, are talking about this. And the second group you talk about is misinformation. This is again, I mean 2024, everyone is looking at the US elections and how AI will develop. And I think we have to address the issue of social media, we have to address the issue of misinformation.

Technically I'm seeing more work now. Technically, digital authentication is actually a very active area of ​​research. I think we need to invest in this. I know Adobe is, I know academia is, I think it's necessary, I hope there are new companies in this space that are looking at digital authentication. But we also need policies. And then jobs. I could not agree more. In fact, the most important work that I think is really at the center of our AI debate is human dignity. Human dignity goes beyond how much money you earn and how many hours you work.

In fact, I believe that if we do this right, we will move from the labor economy to the dignity economy in the sense that humans, with the help of machines and in collaboration, will earn money through passion, personalization and experience, instead of just that. jobs that are really tiring and tiring. And this is also why human HAI at Stanford has a fundamental principle of human augmentation. We see this in healthcare, one of the biggest early days of ChatGPT. I have a doctor friend at Stanford Hospital who reached out to me and said, "Fei-Fei, I want to thank you for ChatGPT." I said I didn't do anything.

But he said we are using GPT's medical summary tool because this is a huge burden on our doctors and is taking time away from patients. But thanks to this I have more time. And this is a perfect example, and we'll see more of it. We might even see this in blue-collar workers. So we have a chance to get this right. I would add another concern in the catastrophic concern: it is actually about power imbalance. One of the power imbalances that I'm seeing right now, and that's exacerbating it as an enormous velocity is shutting out the public sector.

I don't know anything about Canada, not a single university in the US today can train a ChatGPT in terms of computing power. And I think combining all the US universities, GPT-A100 or H100, probably no one has it, but A100 can't train a ChatGPT. But this is where we still have unique data to cure cancer, fight climate change, and conduct economic and legal studies. We need to invest in the public sector. If we don't do it now, we will fail an entire generation and leave that power imbalance in such a dangerous way. So I agree with you.

I think we have a lot of catastrophic risks and we need to address this. That is why we need to work with policy makers and civil society. So I don't know if I'm saying this in an optimistic or pessimistic tone, some were pessimistic about myself now, but I think there's a lot of work that goes into doing this. - Well, optimistically, since you've both talked a lot about this over the last six or eight months, there's been a big shift, both Geoff and you said, key researchers focusing on these issues and then public and policy changes. in such a way that governments are really taking it seriously.

I mean, you're advising the White House and the U.S. government, you've talked to them as well, and you've sat down with the prime minister or maybe several prime ministers, and they're listening carefully, in a way that they don't. I would have necessarily done 10 months ago, 12 months ago. Are you optimistic about the direction this is going? - I am optimistic that people have understood that there is this whole series of problems, both catastrophic risks and existential risks. And I completely agree with Fei-Fei that catastrophic risks are more urgent. In particular, 2024 is very urgent. I'm pretty optimistic that people are listening now, yeah. - Yes I agree.

I think they are listening. But first of all I want to say: who are you listening to? Once again, I see an asymmetry between the public sector and the private sector, and even the private sector, who listens to? It shouldn't just be big tech companies and celebrity startups, there is a lot of agriculture sector and education sector. And secondly, after all this noise, what is good policy? We talk about regulation versus non-regulation, and I don't really know where Canada is, there's always the United States innovating and Europe regulating. Where is Canada? - Probably in the middle. - Okay, good, good for you.

So I really think we need incentive policy, develop the public sector and unleash the power of data. We have a lot of data locked in our government, whether it's wildfire data, wildlife data, traffic data, climate data, and that's incentivization. And then there is good regulation, for example, we talk a lot about how you have to be very careful when regulating, where do you regulate upstream and downstream? For me, one of the most pressing regulatory points is when the rubber hits the road, it is when the technology is now in the form of a product or service.

You are going to meet people, whether through medicine, food, financial services or transportation. And then there is the current framework, which is not far from perfect, so we need to leverage the existing framework and update it rather than wasting time and possibly making the wrong decision to create a completely new regulatory framework when we have the current framework. some. - Well, we're almost out of time for the discussion part, but we're going to have a long question and answer session. However, before we begin, I will ask two last questions. One is that our view is that this technology will affect virtually everything, and some of the positive impacts are extraordinary.

It will help cure diseases like cancer, diabetes and others. It will help mitigate climate change. There is an enormous amount of things, inventing new materials. I see someone here who is focused on that, who can help in the energy, aerospace and pharmaceutical sectors. And that's a big effort at the University of Toronto. But there is this whole world of new things that couldn't be done before and that can now be done. Basically, it is about advancing science in a way that was previously part of fiction or imagination. Are you optimistic about that part? - I think we are both very optimistic about it.

I think we both believe it will have a huge impact in almost every field. - So I think for those in this room who are really studying, it's an incredibly exciting time to come in because there's an opportunity to be involved in limiting the negative, the negative consequences, but also to be involved in creating all of those opportunities to solve some of the problems that have been with us since we existed as a species. I think, at least from our perspective, this really is one of the most extraordinary moments in human history. I hope that those of you who are embarking on your career will really go out and pursue the more ambitious things.

You can also work on optimizing advertising and other things, or making more Netflix shows, which is great. But also... Geoff would like that. -Yes. So would my mother, who I think has exhaustedto the editorial staff, if you don't use fade, but use shift, I got a complaint from someone who said, I tried it and it didn't work. And they use shifting instead of fading. And the thing is that if we understand fading means changing color and staying changed. But if you say change, it will change color but it might change again. Therefore, it doesn't give the same answer if you change instead of fading away.

It is very sensitive to the wording. But that convinced me that he really understood it. And there are other things that are done. So there's a good question that people have come up with recently: a lot of chatbots don't respond well and some people don't respond well, but GPT-4 does, I mean, you see, I'm answering the question: GPT-4 understands, which It has some relation to what you asked, right? - Yes, yes, precisely. -So the question is like this, Sally has three brothers, each of her brothers has two sisters, how many sisters does Sally have? And most chatbots get it wrong. - What about humans? - Well, I just gave a fireside chat in Las Vegas and the interviewer asked me for an example of things chatbots got wrong.

So I gave him this example and he said, “Six,” and that was a little embarrassing. -We won't ask him his name, I'm just joking. - No, then people are wrong. - Yes. - But I don't see how you can do it well without being able to do a certain amount of reasoning. You have to build a model. -Yes.-And Andrew Ying has these examples where when playing Othello, even if you just give him strings as input, he builds a model of the board internally. So I think they really get it. - And to go a step further, does understanding cross the line of intelligence? - Oh no. - You said yes. - Yes, I mean that I accept the Turing intelligence test.

People only started rejecting the Turing test when we passed it. -So that's the moving goal line he was talking about. Well, do you want to answer? I want to respond quickly. First of all, I also applaud you for asking such a good question. I'm going to respond in addition to Geoff's answer because I think what Geoff is trying to push is really how we evaluate the fundamental intelligence level of these large models. But there are a couple more dimensions. One is, again, the Stanford HAO foundation model research center is creating these evaluation metrics. You're probably reading Percy Helm's articles and all that.

I think also this technology is getting so deep that some of the benchmarks are more complicated than you think ImageNet benchmarks are, for example, in collaboration with the government now, for example, NIST, the Institute US National Standards, What is the T? ? - Technology. - And technology, tests or something like that. We need to start benchmarking against socially relevant issues, not just core capabilities. One more thing, I want to open the aperture a little bit, is that beyond LLMs there is so much technology towards the future of AI that we haven't built good benchmarks for yet. I mean, again, my lab is doing some robotic learning, Google just published the robotic learning paper yesterday.

Therefore, there is a lot more research being done in this space. - Thank you. - Okay, I know we have a lot of questions online. Maybe I'll take a few more in the room and then maybe someone from Radical can read a question or two online. Okay in the room, let's go for one that's not too far from the last one here. Just here. Well. (unintelligible chatter) Here comes the microphone. - Hello, I'm Vishaam, I'm a graduate student at the University of Guelph doing my thesis in AI and agriculture. So based on something you mentioned, universities do not have enough funds to train grassroots models.

Same question: I want to work in artificial intelligence and agriculture. I am passionate about it, but I don't have enough resources to do it. I could think of a very good architecture, but I can't train it. So maybe I can go to the industry and present the idea to them, then I have no control over the idea. I don't know how they are going to apply it. Do you have any advice on how to handle the situation? - Make a start-up. - If you can get... - Start a startup, that's what we're here for. Oh, sorry, I'll let you answer. - If you can get a basic open source model, you can fine-tune one of those models with far fewer resources than it took to build the model.

Therefore, universities can still adjust those models. - That's a very pragmatic answer for now. But this is where we've really been talking to higher education leaders, as well as policymakers investing in the public sector. We have to have a national research cloud, I don't know if Canada has a national research cloud, but we are putting pressure on the US, we need to bring researchers like you to be able to access the national research cloud. But the advantage of not being a company is that you have more opportunities to get your hands on unique data sets, data sets, especially for the public good, and play that card.

You could work with government agencies, work communities and whatever because the public sector still has the trust and take advantage of that. But for now, yes, fine-tune the open source models. - Thank you. Thank you so much. - Well, let's answer a couple of questions. We have thousands of people watching parties online at Stanford and elsewhere, so let's see if we can get questions from some people online. Leah is going to ask this question on behalf of someone online. By the way, she has worked incredibly hard to make this happen along with Aaron Brindle, so thank you both. - Thank you Jordan. (Audience applauds) Very good, thank you.

We have hundreds of AI researchers online and they are people who are building AI companies. So, the first most voted question was from Ben Saunders or Saunders. He is currently CEO of an artificial intelligence startup. And his colleague was actually a student of Geoffrey Hinton in 2008. And he's asked about

responsible

construction, and a lot of these questions have to do with responsible construction, and they're thinking about what actions can help them as teams to be good managers for the good versus bad, and what does it really mean to be a butler? - Big question. So there is a responsible AI framework, there are many frameworks, and I think someone estimated that a few years ago there were like 300 frameworks from the nation state, the state to the corporations.

I think it's really important for all companies to create a responsible framework. There are a lot of things you can borrow, even Radical is doing one and creating the value framework that you believe in and recognize that the AI ​​product is a system. So from the beginning, defining the problem, the data set, the integrity of the data, how to build models, the deployment and creating a multi-stakeholder or multi-stakeholder ecosystem, whatever team helps you build this responsible framework and also create associations. Partnerships with the public sector like academia, like us, partnerships with civil society that cares about different dimensions, from privacy to bias.

So really try to leverage it, you both have a point of view as a company, but you're also part of the ecosystem and you partner with people who have this knowledge. That's my current suggestion. - I'll add to- Do you want? - No, that was a much better answer than I could have given. - I'll just add a little bit. I think to Fei-Fei's point, working with people who are interested in this, I think there are people in the investment community who are thinking and leading this, in our case Radical, we have written in each term sheet an obligation of the company to adopt responsible AI.

Initially, when we did that, some of the lawyers who read it and said, what is this? And I tried to cross it out, but we put it back. But we've also been working on a responsible AI investment framework that we're going to launch quite widely. And we've done it in partnership with several different organizations around the world. We've met with 7,000 AI companies in the last four years and I think we've invested in about 40. So we've seen a lot and tried to build a framework that others can use in the future. And we will open source the code so we can develop and improve it.

But I think there's a lot that individual companies can do simply by reaching out to like-minded others. Do you want to ask another question? - Yes, great, there are so many questions, so unfortunately we will only answer a couple of them. But from that, many of these questions have to do with the relationship with the industry, considering the important role that industry and the private sector now play in the development of models. And some people even wonder: should researchers and different engineering positions also take management courses today? - Sure. - I have to tell you a story from when I was at Google, I managed a small group and we received reports every six months from the people who worked for us.

And one of the reports I got was, "Geoff is very nice to work with, but he could benefit from taking a management course, but then he wouldn't be Geoff." (Audience laughs) That's how I feel about management courses. (audience laughing) (audience applauding) - I don't have a better story than that. (Audience laughs) - We have about a minute and a half left, so maybe we'll do one more in the room if we can. We'll see. Do you want to drink? Yes, no, at your side. I'm sorry. Alright, hopefully ask quick and then we'll get a quick answer. - Thank you.

And it's a pleasure to be here. Good to see you Fei-Fei. My name is Elizabeth. I work at Cohere. So my question is, from a private sector perspective, do we work with everyone to bring NLP big language models to society at large, to specific public sectors and research institutions, universities that have a lot of talent and a lot of data. What's the best way to find the kind of mutually beneficial relationship where we can contribute and they can contribute? Thank you. - Give them some money. (audience laughing) - Thank you. (Audience applauds) - Or H100, we'll take H100.

But look, it's very important. I advocate for public sector investment, but I probably also advocate more for partnership. We need the government, the private sector and the public sector to work together. So over the last four years at Stanford HAI, one of the main things we've done is create an industry ecosystem. And there are a lot of details that we can talk about offline, but if I talk to university or higher education leaders, I think we need to accept them. We need to accept that responsibly. Some people will have different ways of calling it, but I think this ecosystem is very important.

Both sides are important. Create that partnership, be the responsible partner for each other. And resources are a big thing. We would appreciate it. - Thank you. - Okay, with that we're running out of time. I want to thank you both. I always feel very privileged to be able to call you both friends, Fei-Fei, your partner, and Geoff, your investor, and have these conversations privately with you. That's why it's great to bring them both together and let other people hear what they have to say. So thank you both so much for doing this. Hopefully it was as informative for you as it was for me. (Audience applauds) And we'll turn it over to Melanie Woodin, Dean of Arts and Sciences at U of T. - Thank you very much, Jordan.

So, Geoff, Fei-Fei and Jordan, on behalf of everyone in the room tonight here at MaRS, and the thousands who joined us online, we are deeply grateful for such an insightful conversation tonight. I can say, and I think many of us know, that being part of a university community offers an endless set of opportunities for interesting conversations and lectures. And as Dean of a College of Arts and Sciences, I have the pleasure of attending many of them. But I can say without reservation that tonight's conversation was truly incomparable. And of course, this conversation couldn't be more timely.

Geoff, when you shared your concerns with the world about the threats of superintelligence, we all listened to you and we all did what we could to try to understand this complex issue. Whether it was reading opinion pieces, watching his video, or reading a lot about journalism, we really tried to understand what he was telling us. So hearing directly from you and Fei-Fei, who spent so many years leading the way in human-centered AI, is really powerful. With that, thank you both and thank you to everyone here for attending this afternoon, and many thanks to Radical Ventures and the other partners who make this evening possible.

And with that, the talk concluded and we invite those of you who are here with us in person to join us in the lobby for light refreshments. Thank you for joining us. (audience applauding) (audience chatting indifferently)