This episode is brought to you by Pique Tea. That's p-i-q-u-e. I have had so much tea in my life. I've been to China. I've lived in China and Japan. I've done tea tours. I drink a lot of tea. And 10 years plus of physical experimentation and tracking has shown me many things, chief among them that gut health is critical to just about everything and you'll see where tea is going to tie into this. It affects immune function, weight management, mental performance, emotional health, you name it. I've been drinking fermented pu-erh tea specifically pretty much every day for years now. Pu-erh tea delivers more polyphenols and probiotics than you can shake a stick at. It's like providing the optimal fertilizer to your microbiome. The problem with good pu-erh is that it's hard to source. It's hard to find real pu-erh that hasn't been exposed to pesticides and other nasties, which is super common. That's why Pique's fermented pu-erh tea crystals have become my daily go-to. It's so simple. They have so many benefits that I'm going to get into and I first learned about them through my friends Dr. Peter Attia and Kevin Rose. Pique crystals are cold extracted using only wild harvested leaves from 250-year-old tea trees. I often kickstart my mornings with their pu-erh green tea, their pu-erh black tea, and I alternate between the two. The rich earthy flavor of the black specifically is amazing. It's very, very — it's like a delicious barnyard. Very peaty if you like whiskey and stuff like that.
They triple toxin screen all of their products for heavy metals, pesticides, and toxic mold contaminants commonly found in tea. There's also zero prep or brewing required as the crystals dissolve in seconds. So you can just drop it in your hot tea or I also make iced tea and that saves a ton of time and hassle. So Pique is offering 15% off their pu-erh teas for the very first time exclusive to you, my listeners. This is a sweet offer. Simply visit piquetea.com/tim. That's p-i-q-u-e-t-e-a.com/tim. This promotion is only available to listeners of this podcast. That's peaktea.com/tim. The discount is automatically applied when you use that URL. You also have a 30-day satisfaction guarantee, so your purchase is risk-free. One more time, check it out. Peaktea, that's P-I-Q-U-E-T-E-A.com/tim.
This episode is brought to you by ButcherBox. ButcherBox makes it easy for you to get high-quality, humanely raised meat that you can trust. They deliver delicious, 100% grass-fed, grass-finished beef, free-range organic chicken, heritage breed pork, and wild-caught seafood directly to your door. For me, in the past few weeks, I've cooked a ton of their salmon, as well as two delicious barbecue rib racks in the oven. Super simple. They were the most delicious pork ribs I've ever prepared, and my freezer is full of ButcherBox. When you become a member, you're joining a community focused on doing what's better for all. That means caring about the lives of animals, the livelihoods of farmers, treating our planet with respect, and enjoying better meals together. ButcherBox partners with folks, small farmers included, who believe in going above and beyond when it comes to caring for animals, the environment, and sustainability. And none of their meat is ever given antibiotics or added hormones. So how does it work? It's pretty simple. You choose your box and your delivery frequency. They offer five boxes, four curated box options, as well as the popular custom box. So you get exactly what you and or your family love. Box options and delivery frequencies can be customized to fit your needs. You can cancel at any time with no penalty. ButcherBox ships your order frozen for freshness and packed in an eco-friendly, 100% recyclable box. It's easy. It's fast. It's convenient. I really, really enjoy it. And best of all, looking at the average cost, it works out to be less than $6 per meal. Skip the lines for your Thanksgiving turkey. This holiday, ButcherBox is proud to give new members a free turkey. Just go to butcherbox.com/tim to sign up. That's butcherbox.com/tim to receive a free 10 to 14 lb turkey in your first box.
Hello boys and girls, ladies and germs. This is Tim Ferriss. Welcome to another episode of The Tim Ferriss Show. My guest today is Eric Schmidt. Eric Schmidt, that's S-C-H-M-I-D-T on Twitter @ericschmidt, is a technologist, entrepreneur, and philanthropist. He joined Google in 2001 helping the company grow from a Silicon Valley startup to a global technology leader. He served as chief executive officer and chairman from 2001 to 2011 and as executive chairman and technical advisor thereafter. Under his leadership Google dramatically scaled its infrastructure and diversified its product offerings while maintaining a culture of innovation. In 2017 he co-founded Schmidt Futures, a philanthropic initiative that bets early on exceptional people making the world better. He serves as chair of the Broad Institute and formerly served as chair of the National Security Commission on Artificial Intelligence. He's the host of Reimagine with Eric Schmidt, a podcast exploring how society can build a brighter future after the COVID-19 pandemic. With co-authors Henry A. Kissinger and Daniel Huttenlocher, Eric has a new book out titled The Age of AI and Our Human Future. You can find him again on Twitter @ericschmidt and at ericschmidt.com. Eric, welcome back to the show.
Thank you. I really look forward to this conversation.
I have been looking forward to this and I want to confess first and foremost that I have tremendous insecurity around my lack of clarity on AI. So I am really looking forward to digging into many facets, but before we get to that, I want to pick out Henry Kissinger. How did you come to collaborate with Henry Kissinger?
About 12 years ago, I met him at a conference called Bilderberg. And my father worked for the Nixon administration when I was very young. And my father had Henry as a hero. He said he was the most brilliant, hardest working, and he was enigmatic because he has both a very favorable reputation, but also a very controversial reputation. So we chatted and he said, "The only problem I have with Google is I think that you're going to destroy the world." And I thought, "Well, that's a challenge from Henry Kissinger." So we invited him to Google, where he gave a speech. And he confronted the employees directly on the manipulation that Google was doing in his view of the public discourse. And his criticisms were so apt, and the people enjoyed it so much that we struck up a friendship. And although we're politically very different with very different backgrounds, I've come to learn that working with someone this brilliant at any age is phenomenal. But when your co-author is 98, it's a special treat.
I looked up the age. I'm glad you mentioned it. 98. To what do you attribute him remaining cognitively sharp into his late 90s? Is that just good hardware out of the box? Is there more to it?
He works harder than a 40-year-old. I can tell you that he gets up in the morning and he works all day. He has dinner with his wife and his family, and then he works at night. And he keeps up that pace 7 days a week at 98. I am convinced that the secret to longevity is being a workaholic.
And the reason I say that is that Henry Kissinger at the age of 90 knew nothing about the digital world, although he had a lot of opinions about it. But he has mastered the digital world and artificial intelligence with the alacrity and the speed of people who are just getting into it now. That's unique to him. That's a gift. And that's why his analysis of our world is so incredibly interesting to me.
All right, we're going to spend a little more time on Henry because I must scratch this itch. So you are very good at systematizing thinking from first principles, whether it's systematizing innovation or hiring, thinking of things at scale. There's a method to not the madness, but the outcomes. Was Henry's ability to learn so quickly based on some approach that he has to first principles or a framework of any type that you've seen exhibited in him that he applies to new domains?
There've been some studies about the age at which you are your most productive professionally. And as you know in math and science, brilliance tends to show up young, in their 20s. They tend to get early awards. Historians, however, seem to get better with age. Maybe it's the accumulation of perspective and the accumulation of reasoning and the depth of wisdom that is represented by increasing age. And so Dr. Kissinger has both the benefit of being a brilliant historian and also having changed history and lived in the moment. And today he spends a great deal of his time with people talking to him about current affairs and judging them with his historical principles in mind.
It is from that basis of insight that when he looked at AI, he said, "This is a very much bigger thing than people think it is." And I said to him, "Why?" I honestly didn't know. And he said, "Because we're discussing artificial intelligence as though it's a technology. This is like the beginning of the Renaissance." And I said, "Tell me about the Renaissance. What else do you say to a historian who's famous?" And we started to talk about the Renaissance, and he said that the Renaissance is really about the age of reason. It's about individuals being able to think through their systems. It's about society allowing experts to criticize other people. Before the Renaissance, decision-making was essentially hierarchical and from a king or a religious leader. That change allowed us to develop intellectual thought. He is arguing that we're entering a new epoch similar to the Renaissance, this age of artificial intelligence, because humanity has never had a competitive intelligence similar to itself but not human.
I'd love to hear you elaborate just a bit, and then we're going to dig into a whole slew of questions that I have in front of me. How you thought about the composition of the co-authors on this book, and certainly would love to learn more about Daniel. I know a little bit about Daniel Huttenlocher, but I'd love to hear more. And we could start with Dr. Kissinger. I realize calling him Henry's probably going to give me bad karma, so I'll start with Dr. Kissinger. Is it his broad familiarity with history as well as his knowledge of geopolitics and statecraft that you were hoping to augment everything else in the book? Maybe you could just speak to how you think about what each party brings to this project?
In our collaboration, the book suggestion was actually from Dr. Kissinger. And what he basically said is that we have an opportunity to architect the questions that need to get answered in the next 10, 20, 30 years. I would parenthetically offer that we didn't understand when we invented social media 15 years ago or so, the extraordinary and compound benefits and costs of social media to our society and particular political discourse. So, armed with that knowledge that the tech industry had invented a tool that outstripped the governance of it, at least initially and maybe for a long time, Dr. Kissinger said, "This is an opportunity to ask the right questions."
We recruited Dan Huttenlocher, who is the Dean of MIT computer science, partly because he's a good friend and partly because he's such a good scientist, he will make sure that our claims are accurate. And Dan worked very hard to get the path of AI correct. Many of the books about artificial intelligence are speculative, but I'm not aware of any book that has both the geopolitical, social, and historical context, but also has the technology correct.
Just to look at the opposite of speculative from a first-hand perspective, could you tell a story or two about what you've seen or experienced firsthand with AI?
Well, you use AI today whenever you use Google Search, Google Ads, spelling correction, Google Translate. The story on AI that's not really well understood is that in the 1960s and 70s when I started, AI was going to happen within a decade. And my friends who were AI obsessed got their PhDs in this area. And then everything stopped. It stopped working. And there was a period of about 20 years which is known as the AI winter where the systems didn't work. And then a series of mathematicians in the 80s and 90s invented what is today known as deep learning. I'll spare you the technical details. But the important thing about this deep learning is it allows the manipulation of patterns at scale that allowed these algorithms to work.
And the big breakthrough was in 2011 with a process called ImageNet where there was a contest to see if computers could see better than humans. And today computers can see better than humans. Their vision is literally better. And I didn't realize at the time how important sight was for everything. A car should be driven by a computer. The doctor should use an AI system to examine you and then give him or her recommendations on your care. I'd much rather have the computer look at my skin rash or my retina in my eye because we now know from many, many tests that humans make observational mistakes, even the best, but computers when properly trained don't. So it was from this insight that you could do vision at scale that you began to be able to do prediction at scale. And all of a sudden now we're beginning to see systems that can predict the next thing.
And computers have gotten very good at predicting what will happen next. The most recent — well, there sort of three events in the last three years that really are the index points. The first is that — and Dr. Kissinger wrote an article after AlphaGo called the end of history basically, and inspired by the fact that Go, it was a game that humans had played for 2,500 years. It was thought to be incomputable. Not only did a computer solve the game but it beat the top humans both in Korea and China. I know because I was physically there. Computer against human and it was great fun.
But in that process the computer invented some new moves and strategies that had not been known to humans for 2,500 years. Now that's a big deal. The next thing that happened was that at MIT a set of synthetic biologists and computer scientists did a very complicated trick involving going through a hundred million different compounds and figuring out which compounds would create a reaction for antibiotic use. And they came up with using this technique a new drug that could not be foreseen. It's called Halicin and it appears to be the next broad scale antibiotic. We haven't had one in roughly 40 years.
The third thing that happened was that a group called OpenAI built what are now called universal models where they read everything they could find on the web into something called GPT-3. This is called basically a transformer and these generative models can generate things. So, all of a sudden, we have a computer that can speak what it knows. And these models are interesting because you train them and you don't know what they know. And furthermore, they can't tell you. You have to ask them. So, it's like a teenager. And many people think that these universal models are going to profoundly change our understanding of language and thought because they only get better with scale.
So, we've now got four or five companies, couple big ones, couple startups that are building what are called trillion parameter models. These trillion parameter models cost a hundred million dollars or so to make. That's how exciting this new area is. So, you've got strategy in the form of Go, you've got medicine and drug discovery, really scientific discovery in the form of AlphaFold, and now you've got language models and learning models. And we believe collectively that in the next 10 years, this is going to come together and transform everything.
I'm just taking a breath to let that settle in. GPT-3 really got my attention in part and I want you to disabuse me of any misunderstandings, but I saw interviews generated between people who are no longer living. So, I saw Marcus Aurelius interviewing — let's just call Mark Twain — Mark Twain interviewing Abraham Lincoln and I thought to myself, once we have enough audio and video online, this will really simplify my guest recruitment for the podcast because I won't actually need to reach out to the living human. It'll be great. I won't have to keep banging on the door of Oprah or anyone else. I'll just be able to generate an interview. It'll save me recording time, too.
You say that in jest, but we're busy building that. And so, let me just be clear how the Tim Ferriss Show would work. So, you have an enormous amount of video and audio data. So, there's enough to learn to build a Tim Ferriss question and answering thing where you would have not only your tone and the fun that you represent, but it would also represent the sum of your insight. So, it would say, "I was speaking with so-and-so, and he or she said this, what do you think?" Now, in this particular scenario, this is long after your own passing and all your guests passing, too. There's every reason to think that we'll be able to mimic your intelligence and charm and wit and that of your family, your parents, your grandparents, historical figures. That will be fun. The question is, will it really change society?
Or how will it change society? I recall being in a small meeting, private meeting, but the content I'm going to mention is public, and the discussion centered on the legal cases that will be forthcoming related to a lot of this technology, and there was at the time something related to — Now, I should be clear. I want to take a backstep in a minute and define artificial intelligence and what that is so that it doesn't get conflated with other things. But in terms of developing technology, deep fakes, copyright, defamation, there's been a lot in the news, at least in the last few years, related to deep fakes of Taylor Swift and lawsuits, and we are all going to have to contend with entirely new breeds of legal cases, I would imagine. So, that's another facet of this that is going to be incredibly fascinating, terrifying, and complex, possibly.
I would love for you, if you don't mind, to just define for me, quite frankly, artificial intelligence, because as I'm sure you have seen and can imagine, there are a lot of terms that kind of get copied and pasted into startup decks whenever they're hot. That would be deep learning, machine learning, AI. What is AI and what is it not?
The simplest explanation for AI is it's a system that gets better through learning. It is busy learning something. And that's probably the easiest and current definition of it. When I say AI to my non-technical friends, they typically think of a movie that they saw. And the movie always has a robot, and the robot goes awry, and that robot is slain by a female scientist who triumphs. And I propose a variant of that movie where the computers all conspire against humans and start killing humans. The humans notice this, the computers run away from the humans, and a female scientist figures out how to unplug the computers one by one by one. That's not what AI is. That's a movie fiction. And it'll be a very long time, hopefully never, that we'll have to deal with that.
What AI really is is a system of knowledge that is implemented inside the cloud, so it's around you all the time, and it's very good at looking at large data sets and predicting things. It's very good at looking and finding patterns that humans can't see. I mentioned earlier that the computers were very good at vision. That was the first breakthrough. But what they're now doing is looking for patterns of correlation that humans can't see. So, let's imagine that on Tuesday you do something and on Friday someone else does something and it turns out that you're magically linked in the universe such that Tuesday causes Friday, but it's not a pattern that's apparent. Computers can discover that pattern, that needle in a haystack particularly well.
When you operate this stuff at scale, you end up with systems that look human-like because they can aggregate data and they can think what you think. They can generate a solution. There's a technology called GANs, generative adversarial networks, where the computer can generate candidates and another network says not good enough, not good enough, not good enough until they find it. GANs have been used for example to build pictures of actresses and actors that are fake, but look so real you're sure you know them. So one of the concerns, one of the real concerns about AI, is that it's going to be very difficult to tell the difference between information and misinformation.
And I'll give you an example. Let's assume that all of your readers and listeners are obviously human and that the standard rules of humans apply to all of us. There's something called anchoring bias, there's something called recency bias. You know, people aren't completely rational computers nor should they. We don't want them to be. We love humans. But let's imagine the computer gets enough of such people and it figures out that if I say the following outrageous thing first, you'll always believe me. And the computer discovers this and so all of a sudden everything it does, boom, boom, boom.
I thought a lot about the way politicians speak. If you watch carefully, they take a set of phrases and they repeat them over and over again. They're simple phrases. That's anchoring. They're trying to get the audience, their voters, to start with this fact and then judge past it. Well, computers will be incredibly good at exploiting that. That means that the world, our social world around us within the next decade, will become impossibly confusing because there's so many actors that will want to misinform us. Businesses, politicians, our opponents, for fun, God knows. We don't know how to manage that.
When you mentioned things coming together in some fashion in the next 10 years, could you speak to if and how generalized or general artificial intelligence fits into that? Because I know that's a big question for a lot of people, including those who don't really understand the technology, when generalized artificial intelligence will arrive, so to speak.
I think it's important to establish some of these terms. AGI stands for artificial general intelligence. And AGI refers to computers that are human-like in their strategic and capability. Today's algorithms are extraordinarily good at generating content, misinformation, guidance, helping you with science, and so forth. But they're not self-determinative. They don't actually have the notion of who am I and what do I do next? If you ask GPT-3 who it is, GPT-3 will say, "I'm a computer." But it won't give you a philosophical basis for its existence, its purpose, and how it's determining where it goes. That is the distinction between AI and AGI, in my view.
Now, within the community, the AGI optimists think that within 10 years we will have such computers. The pessimists think that it will be far longer if ever. So, let's say 15 years from now, it may be possible to have computers that have a sense of self-determination, that they know roughly what they're trying to achieve against a broader objective. Many people think that this will become an enormous competition between humans. Because remember that those computers will be faster learners than humans. They'll have access to more data than humans. Other people believe that they'll be fundamentally flawed. Because they won't have the nuance, they won't have the background, they won't have the cultural background of the countries that we all grew up in and the cultures we grow in. We don't really know.
My own view is that this will occur. But that the amount of computation required to do this will be so great that there'll only be a few of them. And furthermore, they'll become so important that they'll be like nuclear weapons. And let me tell you why. Let's imagine you have a truly evil person who encounters this AGI. What's the first question they'll ask them? Tell me how to kill a million people. Obviously a terrible thing to ask and answer. And the AGI, because it doesn't have morals, could actually answer. It could actually say, well, do this, this, this, and this. And furthermore, it could articulate something that's not generally known because it knows everything.
So I think a fair reading of these AGIs is that in the most extreme form, they're going to be so powerful that they're going to have to be protected. We don't want them broadly used. They're going to have to be used in specialized scenarios. Now, such an intelligence will be enormously valuable for drug discovery, material science, climate change, making the world more efficient, making people more educated. And imagine with an AGI, you could say, "Tell me how to teach a million children English better." And he could figure that out. Because it could in its own thinking to the greatest thinking, it could look at all the patterns, figure out how to do it, and you could actually ask it to write the program.
The beginning of this is today there are companies that are offering tools which will help you write code. They don't know what code you're trying to write, but once you start, it can fill in a bunch of it for you. That's the beginning of this phenomena. In the extreme case, there are scientists and basically science fiction around the future that could use something called the singularity, which is the point where the computer evolution is so much faster than humans. Now, that's speculation. When you talk to people in my world, most people think the singularity will occur 20 years from whenever you ask them that question.
So, they would say 2041.
Where do you fall, if I may ask your personal opinion, assuming, and this is a big assumption, that there are people on the techno-optimist and pessimist side who are very technically sophisticated? What is your current best guess or intuitive feel, however you want to answer it, on the arrival of something we would consider AGI?
The top scientists, the greatest inventors in this area, collectively believe that we need one or two more breakthroughs to get to volition, to get to consciousness in the way that humans do it. When you wake up in the morning, you have so many choices of how to spend your time. How do you choose? How do you handle unplanned situations and all of that? Most people believe that the computers that I'm talking about will become enormously valuable partners.
My physicist friend, I said to him, "What would you like the computer to do?" And he says, "It's really simple. There are so much physics being written about now that I don't have time to read it all. And I want the physics assistant to go and read everything, figure out what disagrees with each other, suggest things for me to think about while I'm sleeping." The biologists have similar questions. People who are philosophers want new insights. Read everything. Help me think through this. I believe that in the next decade that will be the primary achievement in AI. And that is extraordinary. Because it means, for example, that we can answer questions in physics, math, medicine, so forth, that have been unanswerable. We've not been able to understand the behavior of subatomic particles. We can do it now. The enormous breakthroughs, but that's not the same thing as real human intelligence. I think it's going to take another breakthrough or two.
I know this is maybe trying to look into the crystal ball with too much intensity, but could you elaborate on if those breakthroughs are known, what they are, what the problems are that need to be solved? Are they known problems?
There are people who are working really hard on the question of goals. And the computer systems have knowledge that they cannot explain, which we can think of as intuition. So, the system gets trained and it knows things, but if you ask it why it knows it, it can't tell you. That's intuition. So, the question here is how much more computational power and scenario planning do you have to do to look at all the different choices that you have every day? There are people who believe that it's perhaps a hundred times more computing power. And computing power goes up, it doubles every 2 to 3 years right now. And so, that gives you a sense that within 15 years it's possible to imagine that amount of computing power required to do this.
And the reason we use computing power is one, we don't have a better metric, but also because the computer doesn't think the way humans do. The computer looks at scenarios and eliminates. So, it says, "Well, what if this? And what if this? And what if this?" And the computer itself is eliminating this choice, this choice, this choice, this choice, and finally settles on a good choice. So, many people believe that the compounding of knowledge will take what we say a couple of orders of more magnitude. Which is why these computers are likely at least initially to be few, extremely expensive, and very, very large.
Does the 15-year hypothetical time horizon factor in — and this is me getting into very slippery territory since I'm non-technical — but does that factor in the potential applications of quantum computing?
For at least 20 years, people have been focusing on this notion of a different kind of computer called a quantum computer. And quantum computers are hard to explain, but basically think of them as they do all the computations at the same time. So, instead of going ka-chunk, ka-chunk, ka-chunk, all the ka-chunks go at the same time. And the term quantum supremacy means that everything occurs at once, rather than taking the computer time to go through each of the mathematical calculations.
The reason that this is of such great interest is that one of the core aspects of security in the computer age is the difference between multiplication and division. And with quantum computers, you could break all of the codes and all of the secrets and all of the keys and all of that that we all use every day. So, a lot of people think that national security groups are busy working hard to do that. The consensus in the industry is that we are 8 to 10 years away from that.
And the way they get there is that the current quantum computers have an error rate because it's not perfect when they operate. They're essentially mimicking a natural process digitally. And Google showed last year that they could do this on a computer that would have taken a million years that took roughly 10 seconds because of it. So, we know it's possible. The question now is how do you actually build a real one that's useful? If you had a real one, you could instead of simulating a physical process — the typical one is called annealing where two metals basically merge, and this is crucial for high-strength steel and titanium and things like this. And is of great business interest. But the same thing would apply to AI if it worked because the algorithm, instead of looking at all these different scenarios, would examine them all at the same time and pick the best one. Technically, the one with the lowest energy state.
The current quantum learning is not proceeding very quickly. It's a very hard problem. My assumption is that quantum learning will occur in our lifetimes, but not very soon.
Thank you. And for people listening, you want to do a real deep dive on quantum learning and quantum computing. I have a conversation with Steve Jurvetson that goes very deeply into this subject matter, and it really does start to be a mind bender when you get into the details. It is quite something.
Just a quick thanks to one of our sponsors, and we'll be right back to the show. This episode is brought to you by ShipStation. The holiday season is fast approaching, and we know that people will be buying more stuff online than ever before. All of these trends to e-commerce have been accelerated due to COVID and much more. If you're an e-commerce seller, are you ready to meet the demands of a record-breaking online shopping season? Be ready with ShipStation. ShipStation.com is the fastest, easiest, and most affordable way to manage and ship your orders. In just a few clicks, you're managing orders, printing out discounted shipping labels, and getting your products out fast. Happier holidays for you and your customers. ShipStation takes the hassle out of holiday shipping. No matter where you're selling, on Amazon, Etsy, your website via Shopify, or other platforms, ShipStation brings all of your orders into one simple interface. And ShipStation works with all of the major carriers, USPS, FedEx, UPS, even international. You can compare and choose the best shipping solution every time, and you can access the same postage discounts that are usually reserved for large Fortune 500 companies. It's no wonder that ShipStation is the number one choice of online sellers. And right now, my listeners, that's you guys, can try ShipStation free for 60 days when you use offer code TIM. Just go to the homepage, ShipStation.com, click on the microphone at the top of the homepage, and type in TIM, T-I-M. That's it. Go to ShipStation.com, then enter offer code TIM. ShipStation.com. Make ship happen.
I want to ask you, you mentioned what philosophers might ask for in terms of augmentation or help. You mentioned the biologist, the physicist. I am going to ask you, I'm just planting a seed, what Eric Schmidt might use AI technology for in the future. But before I get there, I want to just look at a snapshot of current day. What are some of the coolest or most impressive things that you've seen AI figure out on its own?
I mentioned this new drug and drug discovery will accelerate. Combination of the mRNA achievements plus the ability now to essentially replace the way the drug lab works. My sort of stereotype is the chemist wakes up in the morning and says, "Let's try the following seven compounds." They try the seven compounds, none of them work, and at 5:00 they go home to have dinner and think, watch television, and the next morning they think of another seven. Well, the computer can do, as I mentioned, 100 million in a day. That's a huge accelerant in what they're doing.
I'm very interested in the development of humans together with AI systems. And the example I would offer is you have a 2-year-old and the 2-year-old gets a plush toy. It happens to have AI inside of it. And by agreement as this child ages every year, they get a better toy. And of course, the toy gets smarter. We don't know at all what happens when a child's best friend is not a human or a dog. We don't know what that does to the child's bonding to other children, to their parents. I mean, you know frustrated parents and the kid is busy and they give them a computer and do whatever you want. But imagine if that computer is learning, talking, thinking, educating at the same time. It's a godsend, right? But what is it teaching? What is it teaching? What are its norms? What are its values? Will such a child end up being very stilted with real humans and really comfortable with digital? We honestly just don't know.
We know that people get attached to inanimate objects. There are many religions where inanimate objects have what we would think of as a bit of a soul. We mentioned this in the book. But we don't know. I'll give you another example with elderly. A lot of studies indicate that the elderly are very lonely. It's sort of sad. And imagine if their best friend is a digital friend. What does that do? Does that extend their lives? Does it make them crankier? Does it make their loneliness more perverse? I don't think we know yet.
Yeah, or is that digital friend an emulation of a relative?
That's right. So now we get our grandmother as an example and we recreate her husband. And what does it do when she can chat with her husband who's now passed away? We honestly don't know.
When I go through and I look at these technologies, there really four qualities in AI that are different. The first is it's imprecise. It's imprecise in that it can't tell you exactly what it's doing and it makes mistakes. Don't use it for life-critical decisions. You want a human who can make mistakes consulting a computer. You don't want the computer flying the airplane. And maybe never, but certainly not for a long time.
It's dynamic in the sense that it's changing. It can learn. It can assess the situation around it and change its behavior. This is the AI now. It has emergent behavior. Emergent means things that come out that we don't expect. Strange things. Now, we've seen this a bit with social media. I don't think anyone expected the government interference in the elections in 2016. That was an emergent behavior. It was unexpected by humanity using these tools. And those were not AI-powered. The final point is it's capable of learning.
So, you've got a system using the child scenario. You've got a child that's got a teddy bear that looks like a teddy bear, but it's imprecise, dynamic, emergent, and capable of learning. What's the teddy bear learning? So, the teddy bear's watching TV, too. And the teddy bear says, "Look, I think this show sucks." And the kid says, "I agree with you." Again, we don't understand the implication, especially on formative behaviors. There's a whole other set of arguments about national security and how governments will work and how dominance games will play and who will be winners and losers from this technology that we don't really understand. And we talk about that.
Well, let's actually touch on that for a second. And this is something on the minds of a lot of folks. Certainly, I know a lot of investors who are trying to understand AI in the capacity of acting as an investor. I'm not saying we go there, but you have spoken to Congress about, say, China's announced ambitions with AI. 10 to 15 years isn't that far away. It's really close, and certain things move very slowly. Certain facets of, say, government move quite slowly, regulation. What types of corners are most important to look around? And I'd love to hear you speak a little bit more about the geopolitical components.
I was fortunate to be the chairman of an AI commission for Congress. We just released our report, very proud of it. And we studied a lot about where the world is in AI, not so much on these AGI questions, but the more tactical things that I've been discussing. We concluded that the United States is slightly ahead of China in these areas. That China has a national program to focus on this. They're pouring literally billions and billions of dollar equivalents into this. They generate four times more engineers than we do just because of population size, and they're extremely focused on dominance of AI by 2030, which is soon.
In our report, we speak a great deal about what the government needs to do to help. It starts with more research, access to more data, making sure our values are represented so we don't end up with systems that have a prejudice and so forth and violate both our laws as well as our morals. We talk about partnerships and all of this kind of stuff. The reason this is so important is that pretty much every national security issue will be tainted or controlled by AI in the future. If you think about cyber attacks, the most obvious war scenario in the future is the following. North Korea decides to attack the United States. It begins its attack. China decides this is a bad idea and blocks the attack. America's defenses are awake, alert. America announces a counterattack which stops this. The entire war that I just described took about 10 milliseconds.
Give you another example. In the military, there's a presumption of human control, which is very important. I was part of a team that wrote some of the AI ethics that are now used by the US military, which I just obviously strongly support. And they say we want the principle of human control, human authorization. So, let me give you an example. You're on a ship and a new kind of missile, let's say a very fast hypersonic missile, is coming in, which people are developing. Countries are. And it's coming in so fast that the ship can't see it, the humans can't see it, but the AI has figured it out. So, the AI says to the commander, "In 15 seconds, a missile is going to be showing up and it's going to destroy you and your entire crew. I recommend that you press this button to launch the counterattack."
And the human says, "14, 13, 12, 10, what do I do?" And at 3 seconds, that human's going to press that button. So, the compression of time, the amount of data, and the potential for error creates a whole set of problems around military doctrine. Now, everything that we live in today, deterrence, all of the other things that all of us are familiar with, great power competition, are concepts that were invented over the last 100 years ago. And Dr. Kissinger actually invented much of the containment strategy in the '50s and '60s before he was in government as part of a team that sort of invented all this work.
So, I asked him, "What does this look like now? How would you address this now?" And his answer is, "Let's get an equivalent team of people and try to figure these things out." Containment, for example, doesn't work. Containment is about keeping another country from getting something because these algorithms and this software is pervasive. It leaks and it leaks through ideas and through discovery, not just from criminal leaking. It's not like Los Alamos where you could keep a secret. So, the way to maintain competitive dominance, that is national security, is to invest in these areas, both in terms of data and algorithms, and to be excellent, but also to begin some kind of dialogue about what the limitations of automatic war are.
So, another example, and I'll make one up now. So, the scenario I described of the 10 millisecond war — so let's say that China in this case...
develops an AGI ahead of everyone else. And this AGI is thought to be so powerful that it could defeat any of our defenses. Then, the logic on the American side would be to do a preemptive attack to prevent that possibility. And that's destabilizing of great power competition. So, you want to avoid an arms race. An arms race in AI could look a lot like the arms race that we went through in the '50s and '60s. And people have forgotten how much of our military-industrial complex, how much money and so forth, was devoted to build more than 30,000 nuclear weapons, all of which could destroy the world many, many times over. Many, many rounds of negotiation got those numbers down to 3,000, 4,000 such weapons, which are still plenty to destroy everything. It's an example of overreach. And so, I worry that because we don't have an agreement on even what the rules are, what the landscape of limitations are, we don't have diplomats who can have the conversations. And no single national security group, no single country is going to self-limit and say, "Oh, we're not going to do that." This is not Costa Rica, which doesn't have a military. So, the natural course of logic will be the development of these incredibly fast and potentially destabilizing weapons, for which we don't have a language.
I think I'll just act as a stand-in for listeners. That sounds very terrifying. It does sound not just terrifying, but challenging to address. And I may come back to that and questions of policy and geopolitics. I want to ask you the personal question that I alluded to earlier before I let that go, which is what could you achieve or what would you want to achieve or ask of AI in the future that would allow you to do things you cannot do today?
Almost every hard problem in our society is based on either the computers can't figure it out, we don't have enough data, we honestly don't know how to solve it. I'd like to get some breakthroughs. So, I'd like for example to get better climate models. In my philanthropy, I funded a group at Caltech which figured out a way to predict climate better using an AI system that would model clouds. I didn't know this, but it turns out clouds are really complicated. And so, they used an AI system to approximate how clouds would behave, which allowed them to solve the prediction problem. It's called Clima, C-L-I-M-A. Over and over again, I would like the AI system to educate me and entertain me and keep me curious about the dynamism of the world. Speaking personally, when I look at the news feeds, we're obsessed about politics and President Trump and so forth, to the point where you get the impression that there's nothing else going on. But one of the great things about humanity and our world is it's incredibly dynamic. We never hear what those people are doing. I'd like to see if we could figure out a way to advance against these really hard problems. Let's get some solutions around mental health. Let's get some solutions around drug abuse. Let's get some solutions around the rise of inequality. And let's have our computers help us with those solutions.
And let's just talk about this — is something you know a lot more about than I do, but how programming may be able or incapable of addressing morals and ethics. And I'll give an example, probably a bad example and not technically accurate, but it seems like we're already at a point with say the development of autonomous vehicles where questions that might have been presented in a freshman philosophy class, like the trolley problem, actually need to be programmed on some level in the sense that cars need to make decisions if there is an accident. They need to say choose between going on the sidewalk at high speed and hitting three elderly or swerving right and hitting one child, let's just say. Not that we would ever want to be in that type of situation, but these things happen. How do you foresee computer scientists attempting to, if this is even going to be an objective, instill ethics and morals into AI? And does it just require a level of self-awareness that perhaps we don't have? I'm curious how you might think about that.
There are many, many computer scientists working on this. Every major computer science program I'm aware of has a computer science and ethics research project. There are a lot of problems. One of the first problems is the computer science and AI are based on learning data, literally learning, and therefore the information that they're learning has biases. So, whatever prejudice and bias and religious problems that society has around it, the computer will absorb. So, the research is how to identify those biases, because remember in many cases these systems cannot tell you what they know and how to mediate them.
It's pretty clear to me that there's going to end up being the system that knows everything, and then there's going to be a supervisory system that limits it. So, two different systems. One will be the knowledge system, and another one which will keep it within some guardrails. So, the question that was asked is not a permissible question, don't send it to the system.
And you're going to have all sorts of problems. My favorite self-driving car problem is here in New York. Eventually, all the cars are self-driving because it's such a crowded grid, and traffic moves perfectly. The engineers at Google and everywhere else have figured out exactly how to optimize under a set of assumptions, the aggregate and average delivery time of a human from place to place. It's a perfect computer science solution. So, now we've got a woman who's going into labor who needs to get there faster. Is there a button in the car that says, "I'm pregnant, I'm about to give birth, and I need to get there faster." Okay, so she presses the button, and somehow it works. How do we make sure that a gentleman who's lying cannot press the same button?
In the trolley problem, all of these are of the form, how do you choose one life for the other? The correct answer is, don't kill anything. And so, we're going to have to find ways to avoid being in the situation where you have to choose between three old people and one child. And these are the great debates of philosophers, and how do we value life, and so forth. But, I think we should start with the premise that the system should be designed such that it optimizes overall happiness and overall wealth, in the societal sense of wealth.
As an example, we tolerate double parking. We tolerate people speeding. We don't track every car and give them an automatic ticket every time they go above the 30 mile an hour speed limit by 1 mile. But technically that would be easy. If you want to eliminate the vast majority of crime in our society, put cameras in every public space with face recognition. It's by the way illegal to do so. But if you really care about crime to the exclusion of personal freedom and the freedom of not being surveilled, then we could do it. So the technology gives you these choices and those are not choices that computer scientists should make.
Those are choices for society. I use the surveillance cameras as an example to say that in Britain they're very widely accepted. When you're in Britain on a street, you're on camera. In the United States, it's partial. Sometimes you are, sometimes you're not. But in Germany with a long history of the Stasi, they are very violently opposed to such surveillance. These are three different democracies that have made the choices in different ways. So I don't think that computer scientists and the tech industry should make these decisions, but I do think what we should tell people is our tools give you this range of choices. If you look at leadership in China, China is a couple generations ahead in surveillance. They can actually spot you in your gait now, the way you walk, not just your face recognition. This is not something we want to lead in.
I'm just letting that all sink in for a second.
This is one of those conversations for me. It's great. There's so much to chew on and I'm going to take a 90 degree turn back to the physicist and the biologist and others who will be able to consult AI in the way I think it was presented to help assimilate new information, new developments, get up to speed. They can't possibly even with physical constraints, keep up or digest the amount they would like to digest.
Now, many people, Dr. Kissinger certainly over decades and many, many others, spend a lot of their time ingesting information. And that could be in the form of reading, it could be in the form of conversations like this. I suppose there are two questions. One is, how do you think thought workers or humans will spend their time once that ingestion is dramatically reduced or removed? And along with that, if we get to the point where that is possible, and there's a question of how it would be conveyed to the human, how far are we from there simply being a direct brain-computer interface where that information is somehow seamlessly integrated into our consciousness without being spoken or imparted to us externally?
