Kai-Fu Lee: The Future of Humanity in a World of AI | Endgame #279 transcript

GITA WIRJAWAN: Hi, friends. Today we're honored to be graced by Dr. Kai-Fu Lee, who's one of the world's few leading experts on AI. And he's also the founder, CEO, and chairman of 01.AI. Kai-Fu, thank you so much. KAI-FU LEE: Hi, thank you, Gita, for inviting me. - It's good to see you again in such a short span of time. I want to probe a little bit about how you grew up. You've grown to be a high-quality intellectual, and you've professed in AI for decades already.

But what influenced you at an early age to study and to pursue science? - Well, I was born in Taiwan, so I spent my six years of elementary school there. I think what was helpful was a very Chinese-driven competitive environment, which built a good foundation. But I really didn't know I could love a subject. And then, as I went to the US for middle school, I found that my math was really outstanding. In retrospect, probably because too much knowledge was drilled into my head, and I memorized so many things.

But it gave me confidence: "Hey, I'm really good at math. This is interesting. This is fun." Better than a lot of the other kids. The teacher wants me to help out. And then I started taking university courses in high school. So that gave me a lot of interest in math. And also, in high school, I was very lucky. This was in the 1970s. My high school actually had a computer. It was a retired IBM 360. I got to learn to program in high school.

And that's when I said, "Hey, this is even better than math." So I went to Columbia as an undergraduate, initially as a math major, then switched to a computer science major. And really, I just thought computer science was a lot of fun. I can make computers do things. But what's even better than that was my sophomore year and junior year, I took two classes on artificial intelligence: one on natural language processing, one on computer vision.

Both were so fascinating that I thought this had to be it. It's not just computer science. It's this very new field called artificial intelligence, which encompasses both natural language processing and computer vision. So that was when I really embarked on that journey, which has not ended even as of now, still continuing on the last mile as AI is about to become superintelligent. - Wow. Just to go back a little bit, Taiwan is known for producing so many high-quality scientists.

What do you think would have made Taiwan good at allowing that scientific discovery to permeate into society? - I'm not sure about any one reason, but I think there's a series of reasons. In 1949, many people went from the mainland to Taiwan. These are the elites of the administration that was running mainland China, which then went to run Taiwan: top governmental officials, top doctors, top businessmen. So it may have had a significant effect on the gene pool, because it took the successful people who sided on the losing side, but nevertheless successful and smart people.

That may have been one small reason. If you look at my own background, I'm not sure about Jensen Huang, Lisa Su, Jerry Yang. They may or may not have similar backgrounds. But certainly, my father was a senior official in the government. That may have been one of the reasons. But another reason is that the Taiwan education system combined a lot of good things from the Chinese education system, but also, I think, imported some aspects of the Japanese approach.

And also, for people younger than me, maybe more for Jerry Yang than myself, or Steve Chen, who was even younger, maybe for them it was the very successful Taiwanese government policy about making the bet on Hsinchu Science Park, which then led to Taiwan Semiconductor. That was a very, very prescient decision made by the minister of finance, the same position you had, Li Kwoh-ting, and his boss, Sun Yun-suan. They were very pro-tech, pro-business.

They studied a lot about what's the most exciting area and decided semiconductors were it. Then they recruited Morris Chang at an unheard-of $1 million a year package back then. I mean, this is probably at least $50 million, if not $100 million, today. And brought him into the government. So a government paying this high a salary. He led Hsinchu Science Park before spinning off with Taiwan Semiconductor. So they made huge exceptions to find the right leader to basically begin the dawn of the semiconductor industry in Taiwan.

And then, of course, many other companies followed. Having Morris Chang as a role model, I think, pushed a lot of people into engineering. And also, Taiwan's education system is just like China's. It was basically a single universal exam, and your grade determines what school and department you go into. People really fought hard to get into the best departments, which were electrical engineering at National Taiwan University and Tsing Hua University.

And then, of course, the job they get is to go do great things at Taiwan Semiconductor. So this creates a virtuous cycle, a role model that says, "Hey, high tech is interesting. It's exciting. It's lucrative. It pays well. Stock options are worth a lot of money in Taiwan Semiconductor." And this was 30 years ago. So I think it created a Silicon Valley-like system of attraction and a system of advancement, and also a lot of respect.

"What do you do?" "I'm an engineer." People are proud to be that. I think that was what got Taiwan off to a good start. - It's amazing how Taiwan was able to show a great degree of open-mindedness. As I chatted with you a few weeks ago, there's four ways to cultivate brain or talent, right? The first is brain train, which is really hinging on a lot of talent indigeneity. That's probably manifested in Japan. The second is brain gain, and how you show open-mindedness.

You open the gate, the way Taiwan, China, Singapore, America, Australia do. And then the third is brain circulation, which is pretty much what China has done by sending more than nine million people in the last 50 years all over the world. And then the last one is brain linkage, which has manifested in India, which has sent more than seven million people. And there's all kinds of ways. I'm just always looking for ideas and ways as to how Southeast Asia could do better scientifically, academically, as it relates to how the evolution of the world is becoming much more exponential.

I want to dig back to your educational journey. You decided to go to Carnegie Mellon. What did you see to pursue artificial intelligence? Were you that prescient about the future, as we're seeing today, in how AI has permeated into so many ways we live? - I think it was more fascination and optimism than prescience. I just saw artificial intelligence. I thought it would be the final step for humanity to understand itself.

Basically, I assumed we would study how the brain works, write it in computer algorithms, and then one day it gets as smart as humans. In doing so, we learn how intelligence works. So that was what drove me, and that's what I wrote in my application letter to Carnegie Mellon. What I learned today is, interestingly, humanity's final step to understand ourselves: that part is correct. But the process of how to get there was completely wrong.

It turns out we, the AI community, built a machine that is different from the human brain structure. It may be inspired by it, but it works differently. It doesn't want to be taught by humans. It wants to use mathematics to map data to data. That's called machine learning. It's called neural networks, deep learning, large language models, all in the same family. So that's what beat every other approach. That was the path I did in my PhD thesis.

And I didn't imagine it would become this powerful. I think we underestimated how much data we had and how much compute we will have, which is what led to the large language model revolution, and the fact that we're basically at the footsteps of AGI, artificial general intelligence. How is this humanity's last step into understanding ourselves? The process, as I see it now, is that the AI community has commoditized intelligence.

AI can think more intelligently than most of us, most of the time. That high intelligence has become commoditized and cheap. It is no longer suitable for any human to say, "I'm differentiated by my intelligence. My goal is to improve it, and that's what sets me apart." Well, it's good to be more intelligent, but it is no longer possible to beat humans, or soon will be completely impossible. We've now reached nearly the last mile of effort to understand ourselves, and the essence of being human is not our intelligence.

It must be something else. So that becomes our final step to understand ourselves, because that which we can duplicate is not the reason that we exist. Whether one believes in evolution or God created the world, we are not put here just because we're intelligent. It's great to be intelligent. We should try to be intelligent, but it is not the thing that sets us apart. So let's look for what sets us apart. That's something I put quite a bit of thought into in my new book. - I'll dwell a little bit deeper into that.

But just the last part of your background, how you've spent time at Apple, and then at Microsoft and at Google. Did you ever think intelligence was going to get commoditized the way we're seeing today, back in the '80s and '90s? - I did not, because we always used the most advanced computers at the time. When I was at Carnegie Mellon, my adviser let me use all the workstations in our group, which was probably a million dollars, a huge amount of money in those days.

When I was at Apple, we had some time on the Cray. Actually, Apple was largely designed on the Cray supercomputer, and that was the world's most expensive computer. So what happened was, interestingly, we were looking at today, that compute was always going to be super expensive to do the best AI. So we never really did the simple thing, which is just exponentially extrapolate it. If compute was increasing at this speed, in a matter of 40, 50, 60 years, intelligence will be free.

That's the number-one thing that we overlooked. It is easily extrapolatable that it would get this cheap. Also, the other uncertain thing was, it's often believed to be very, very hard to get the last 10 or 20%. So to get to a 70% good-enough solution is not hard. That means to exhibit a little bit of intelligence is not hard within a single domain, right? We've been talking to speech recognizers for 40 years. That was my PhD thesis.

That exhibited some intelligence, but every bit of improvement is really, really hard. Researchers seem to have shown tiny incremental improvements year after year since the '80s, so that AGI would be very far away. Never did we imagine at that time, in the '80s and '90s, that there would be a new intelligent system, namely the large language models, that would suddenly become general out of just the huge amounts of data.

It would become general, and the intelligence would jump year after year. We did not see that jump coming. That's another interesting paradox. We certainly knew that AI would get smarter with more data. We did not know, with the whole world's data fed into AI, it would get this smart. We certainly knew compute is reduced every year. So it's an exponential cost reduction. But we did not know that at some point, it would reach very close to zero.

So I think I was too optimistic in just believing AI is humanity's final step into understanding ourselves, back in 1980. But I was too pessimistic that adding more data, adding more compute, these two things alone would get us here. But it did. - I just want to give you a bit of a backdrop on the comparison of how economic development between China and Southeast Asia has taken place in the last 30 years. If you look at the GDP per capita of China, it's grown by ten times in the last ten years, whereas that in Southeast Asia, only 2.7 times, right?

How do you see AI being able to boost Southeast Asia in accelerating its velocity of developing, economically speaking? - I think it's going to create the biggest opportunity for every country. But frankly, currently the odds favor the US and China, because they're by far the two leaders. And also, they have different business models and different methodologies. So very possibly, both can win, just like iPhone and Android.

Both won, so to speak: one in making the most money, and the other in getting the largest volume. So my thought for Southeast Asia would be that it would look for a different game to play. China has played a different game, right? China didn't go build premier labs with hundreds of billions of dollars, wanting to build AGI that beat everyone else and become winner-take-all. China took a very different approach. And China didn't sell these expensive APIs.

China took a different approach, an open-source approach for market share, letting the huge consumer software companies like ByteDance and Alibaba build great consumer apps. And it's choosing to ask the government to be the early adopters of AI. So these are almost the antithesis of the American approach, and they both are working. So I would think that Southeast Asia can find a different approach. I have some ideas, but I think, not having lived in Southeast Asia, I don't want to be presumptuous and say it's the right answer.

But I'll give a short version of that. I think there are so many ways to use AI. The US and China are just two ways to use AI. I think every profession will get disrupted, every job will potentially get challenged. But with the problems that causes for traditional businesses and old ways of doing things come new opportunities. What are the segments that Southeast Asia, or actually different countries, can pursue and do something interesting in those areas in which it has an advantage?

The traditional way to look at it would be asking each country, "What is your primary area of strength? Can you apply AI to it?" That's certainly one approach. But I would prefer to start from scratch and just say, "Hey, look, if we have a bunch of well-educated people, and there are plenty of well-educated people, and a large population size, what can they do to ride the wave of AI to create value that the US and China currently aren't pursuing?" So one example is the so-called one-person company, OPC.

That basically means you take a proven formula of saying, "These are the traditional types of activities that will get replaced by AI because AI can do it much more cheaply." Can each individual come up with, or have someone help them come up with, the business that the one person will start to go after that sector? Usually, these one-person companies are focused on the following: it has to be fully digital. There has to be payment.

The work can be almost fully automatable, with very little human in the loop. And the work will not have a high risk of a huge downside. So if you let AI do it, it might make a mistake, but it would not be catastrophic. Examples of this include using AI to automatically create content. This isn't just documents. We can all do that. But this is creating a short-form drama. I've visited companies that can produce a full episodic drama, using fully AI script creation and content creation.

Two people, one month, they can create essentially a serial drama. But that's one example. Or, based on a couple of photos, create a TikTok video to attract users. To use content creation to create pretty much anything: entertainment, movies, games will all be AI-created. So look for areas in that. Second is things related to user acquisition. Help anybody acquire customers, and they'll pay by the customer. So it satisfies the opportunities.

Third is in advertising. So find ways to reach the right audience using AI. Then there are a few other very high-paying things, for checking for compliance reasons. That's not as large a segment, but it's very high-paying. Or creating esoteric software, because AI can write code just based on prompts. If you know how to prompt to get an application, you can build a special app for a customer that does something completely unique.

Before, no one would produce the software, because it might be only used by one person two hours a day. It's not worth it to pay somebody $100,000 to build it. But now, maybe they'll pay you $300, and you'll spend $100 to build it. So these are all opportunities created because AI can automate and take over parts of work. The people who jump in and find these opportunities first will probably not make a unicorn. They probably will not make a world-famous company.

But they'll at least be able to make enough money to support the family. There's a very good chance they would. A few might be so successful they'll create a brand, a phenomenon, an IPO. That's the unlikely outcome, but it's possible. So it keeps the dream alive. Let anybody, everybody be an entrepreneur. There's no prerequisite other than being smart enough and hardworking enough, and having education, and also getting help in whatever area this might pay off.

So that's one possibility, the digital direction that countries can look into. - There is a stark observation about how the per-unit cost of AI from China is disproportionately cheaper compared to that in the US. The second would be the observation that the Chinese companies are still doggedly open-source, right? Whereas the US companies, they're shifting from open to closed source. And of course, on the back of that is the seeking for profit.

Southeast Asia has 11 economies, of which only two economies earn above that of a developed economy, about $13,200 per capita per year. One of which is Singapore at $94,000. So intuitively, the instinct is to go for anything that's cheaper. Now, is there conviction that the Chinese models are likely to remain open-source and, of course, technologically cheaper for quite a while, that it's going to resonate even more and more with the Global South, not just with developing economies in Southeast Asia? - Right, that's not going to stop.

I can say that with almost certainty. Firstly, the Chinese models are cheaper because the Chinese entrepreneurs and large companies are poor. They couldn't raise the kind of money that OpenAI and Anthropic did, so they were forced to make do with however much money they raised, which is less than 10% of the American giants. So they tried to train with less than 10% of the compute power. They have to squeeze everything, otherwise they couldn't even do experiments.

So the process of using incredibly deep engineering tweaking isn't something that's regarded as elegant or fun. Silicon Valley basically shuns that and says, "That's beneath us. My CEO can raise $10 billion and buy us more, faster computers. Why should I waste time doing this low-level stuff? That's beneath me." But the Chinese engineers, nothing's beneath them. Asians tend to be collectivist societies. So an engineer in the company would view it as their pride to work together to create a company's success.

This engineering tweaking causes the models to be anywhere between 50% and 95% cheaper, with slightly worse and sometimes equivalent performance. So no doubt, it should be the choice for Southeast Asian companies to use. It also has a side effect, or rather a by-product, that is also very beneficial. Because it's an open-source model, a company or a country can just take a Chinese model and install it locally. Then it's yours, because you're not sending anything to servers in China or anywhere else.

It is your own server, so you have achieved sovereign AI. Also, when you have the model in your hands, you can tweak it, for example, to speak better Urdu or Malay or whatever language you want. When it's an OpenAI or Anthropic model, it lives in a remote server, and you can't see its structure or its numbers. So the Chinese open-source models are inexpensive and give you sovereign access. So that's an advantage. Some Americans are advocating for building American open-source models.

I think more competition is always good for the consumer. So we'll see if that happens or not. Now, lastly, on your question of whether the Chinese companies will collectively decide to go closed-source, I think that's virtually impossible. The only way that can happen is if in China it became one winner-take-all. Then, when you become a winner, you have a temptation to now close things off. But given there are 11 companies still training frontier models in China, it's really not likely a single one will win.

It's likely there will be consolidation. There might be three or five, but not likely to be one in the near future. The last thing I'll say is that even if the worst thing happened, people bet on Chinese open-source models, and then the market consolidated to one winner, they decided to be greedy and turned it closed. You're still free to choose. You can still choose any open or closed model. The level of dependency we have on any one model is modest to low.

That is, the switching cost of moving to another model is not so high. And I use multiple models interchangeably. So even if the worst thing happens, you just move to another open-source model if one exists, or another closed-source model if you're not happy with the supplier. So not to worry. While models are very expensive to train, they are, on average, not that expensive to use, especially open-source. And then the switching cost is nearly zero.

It's not like once you got stuck on Facebook or TikTok, it's a little bit difficult to move off because your friends are on it. That's sticky. The large language models do not have that ability to create stickiness. So you can leave it whenever you want to leave it. Feel free to use what's suitable now, and then you can pretty much switch at will. - Explain the repercussions of open versus closed upon sovereignty of data and stuff like that. - Right.

So an open-source model is basically a large file. It's like I created a giant Excel file and I said, "It's free. Download it here." So that's the open-source model. If you're a company in Southeast Asia and you said, "Oh, great. I'd love to use that giant model," you download it. Then it comes with instructions on how to connect it to applications. You do that, and it's installed on your server. If you're a company, it's on your company server.

And if you don't want it hosted in your company, you can have Amazon or another hyperscaler host it, and you can use it on their server. But in any case, no data is sent to the model company. So that's the advantage. Having the model local has that advantage. As I mentioned, once you have the data local, if you want to change the model, you can. It's like I gave you the giant Excel spreadsheet. You can change some of the formulas or numbers, or load in another spreadsheet and use macros to compute something.

It's all yours. It's free. So that's the advantage. With a closed-source model, you don't get to see the model, download the model, you don't get to tweak the model. You just basically send API calls. That is, you have your application call up Claude or OpenAI, which might have its data center in Singapore or the US. Then they would respond. To a user, there's not much difference. But to a company, that means sending your data to Singapore or the US, wherever the server is located.

And many people are not comfortable with that. Some people want to tweak the models, and with the American closed models, you can't tweak them. So that's the trade-off. Now, having said all these good things about the Chinese models, I will say there is one significant problem, very significant problem: these companies aren't making money, right? So they're doing goodwill, free work for the world to use. They do have some revenue, depending on how these models are accessed.

If you download it, you pay nothing, unless you're using it for profit. Then you have to pay them something. But it's kind of on a voluntary self-reporting basis. Secondly, if they do host it on Amazon, I think they get next to nothing. But there are other hosting arrangements in which they get a cut. And sometimes they host their own model. Then they bill the people who use their model, in which case they do make some money.

So if you look at DeepSeek or Qwen or Kimi, they have revenue, and it's going up. But this is creating more loss, even though it's also creating more revenue. So it's higher revenue but higher loss. Not only is there serving cost, development cost, but the cost to train a modern large language model is at least $100 million. And their profits are not going to cover that. Unlike a company like Anthropic, which now has reached annual recurrent revenue at around $600 billion.

Sorry, no. $60 billion. They're going to use it to justify a $2 to $3 trillion valuation. So that would be an unthinkable kind of number. The Chinese numbers are not even 1% of that. So here is Anthropic: in the US they have a closed model, they charge a lot of money, they build a high margin, high profit. And yet companies who use them and got used to them are used to paying for software annually. So they pay for tools, pay for APIs.

I'm not talking about tech companies. I'm talking about normal companies like Procter & Gamble, General Electric. Whether they do it for designing their PowerPoint or writing code or reviewing legal documents, they sign a licensing agreement, or they buy AI software, or they let their people talk to the chatbots produced by Anthropic and OpenAI. And all the subscription fees are coming to $60 billion per year. So here's a company only a few years in existence, making $60 billion a year, projecting a 30 to 50 times price-earnings ratio, which is not ridiculous at this over tenfold growth year over year.

You could question, might it be a little bit overpriced? Perhaps. But at, let's say, a $2 trillion valuation... - It's reasonable? - Yeah, it might. At least based on the pure numbers today, I think it looks reasonable. So the Chinese companies face the challenge that they don't have this kind of valuation or revenue or profit. So how do they justify the R&D resources going forward? Of course, the answer is they have to develop a business model.

But Anthropic and OpenAI seem to have already done that. So that is some catch-up, not so much in model capability, but in figuring out how to make money from the China side. - But given the divergence between the Chinese and the US models, I would argue that perhaps being able to command pricing the way Anthropic has been may not be sustainable. I mean, the competition in China is just brutal. Much more brutal. At some point, intuitively, there's going to be an intersection between players from one side of the world vis-à-vis those from the other side of the world.

That, I think, is only going to entail a much higher degree of competitiveness from a pricing standpoint. - Absolutely. All the skeptics who don't think Anthropic is worth $2 to $3 trillion are using that argument, namely that if there were no Chinese open-source models, people seem willing to pay you your high price because they get their money's worth. So you're going to continue your growth. But now there is a 95% good-enough solution for one-tenth the cost.

And a lot of people are going to either switch completely, or only use your model for their most complex tasks, and let others use the Chinese model. Also, Anthropic and OpenAI will keep improving until, at some point, the user can't tell the difference. If you remember the early days of PCs, right? Every upgrade was, "Wow, what a relief. Everything's so fast." Same with every iPhone release. And then they're all the same.

At some point, I remember I said, "I don't need to update my Mac anymore. It's basically good enough. The slightly better speed and memory just isn't worth the money to upgrade." Until they changed the microprocessor. Okay, that one I had to do an upgrade. So I think at some point, it's possible OpenAI and Anthropic will reach the point of diminishing returns, where the improved model isn't so much better, and they can't increase the price accordingly, or they have to decrease the price, perhaps.

Also, if you compare, let's say, Kimi versus Claude, now the gap might be 100 to 95. And some people might still choose Claude because they want 100. But if the gap continues to narrow, and also they're both above the good-enough line, and maybe the difference is really negligible, then they would lose market share. So yes, that is, I think, the largest problem in the argument of extrapolating from Anthropic's past numbers to future numbers.

But what we don't know is, if you ask me what is the single clear advantage of the American companies, OpenAI and Anthropic, over the Chinese companies, I would say, well, one is they're able to make a lot of money. But I've already talked about that. The other is their talent density. So if you ask all of the US, all of China, "Produce your top 100,000 AI researchers," the two lists will be quite comparable, the US a tiny bit better.

But if you ask, "Produce your top 100 AI researchers," then the US would have a dominant advantage, or even a thousand, because there's so much talent density. When Anthropic has much more talent density than the typical Chinese company, can they invent something that's a breakthrough? And can that breakthrough not be reverse-engineerable? If both answers are yes, then it can not only be worth $2 or $3 trillion, maybe a lot more than that.

But so far, to this point in the last three and a half years, we have not seen anything invented by OpenAI or Anthropic that Chinese companies have not been able to reproduce, at least at the result level. But that's not to say in the future it will be the same. - I would provide the perspective of the Global South, which is only 84% of humanity. That's about seven billion people. They're still, in general, at the lower end of the value chain.

The demand for the top-notch echelon of intellectualization may be there, but not pervasively, right? And if they're presented with a comparison between one and the other, and one costs 10 to 20% of the other, intuitively, the choice is really clear. I want to present two observations about Southeast Asia. The first is with respect to the intellectual propensity, and this is manifested in the educational attainment. And we still want to move up the value chain.

I'm in a camp that believes that AI is really hallucination that's affected by the hypnosis, and the quality of hypnosis matters. But if educationally we're not there yet, and this is empirically manifested in how nine out of 11 countries' PISA scores are still below the global average, that's the proficiency of communication and STEM for 15-year-olds. Out of 10,000 universities in Southeast Asia, only two are positioned in the top 20 in the world, both of which are in Singapore.

The third-best university in Southeast Asia is ranked 56, and that's in Malaysia. But no school or university in Thailand, the Philippines, or Indonesia places in the top 100. So I would argue that the quality of hypnosis may not be as high-quality as that in countries that are intellectually up there. The second observation I want to register to you is what we talked about a few weeks ago: the structural limitations from an energy standpoint, where nine out of 11 countries are still electrified at 5,000 kilowatt-hours per capita or less.

Singapore and Brunei, 10,000 kilowatt-hours per capita. How do you think the narrative of Southeast Asia becoming a stakeholder of AI, as opposed to a spectator of the AI narrative, could be ushered with the help of technological innovation coming from anywhere, from China, from the US? - I can see several ways. One is to have an enlightened leader in one of the countries that sees and understands the trend and finds the resources to make it happen.

I've spoken to some of the wealthier countries and their leadership in Southeast Asia. I have to say I have not yet seen that kind of vision. Singapore is the only one that I think has a reasonable vision. You have to be enlightened, you have to see. I think there's such a preponderance of data that I would want to spend some time explaining to each one. These are all very bright people. The writing is on the wall about the commoditization of intelligence.

And that creates not only the two classes that you talk about, right? The ones who are making history, the ones who are following, and then there are also the ones who are trampled. We have to be aware of that when we talk about the Global South. I think Southeast Asia is in average shape, because there are many much poorer countries. When they have no one making money with AI, and most of the jobs are being replaced by AI, produced locally in the US, Europe, or wherever, with robots and factories, I think it becomes an issue of whether people can even make enough money to survive.

But coming back to Southeast Asia, I think one is enlightened leadership, or one that will listen and believe an expert or a group of experts. That's possible. The second possibility is for there to be some existence proof. "Hey, if you do this, you can get the results. It's been done." So some country, or maybe not a whole country. Maybe the country leadership doesn't have that understanding of AI, but maybe a governor or a mayor does.

Then do an experiment to show, like the example I gave earlier of a one-person company. I know there's great entrepreneurs in Indonesia, for example, a great entrepreneurial community. And many of them are looking to do this. But also, I think people who don't have AI expertise can now do it. So the biggest difference is whether someone can demonstrate. You don't have to have the vision and the brilliance, but when you see the data, you believe.

I think that's a second way. Otherwise, the danger is not that you can't go from a follower country to a leader country, but that you might fall from a follower country into a trampled-on country. - China produces about four to 4.5 million STEM products per year. India, about two million. The US, 800,000. Southeast Asia, only 750,000, of which Indonesia produces about 250,000 STEM products per year. Is that a structural challenge that needs to be remedied? - Well, absolutely.

I think it's really the quality of education that's important. For the past 50 years, STEM has been incredibly important. It led to the engineering jobs, engineering innovations. One could reasonably debate what is the right skill set going forward, STEM being one of them. I would argue the education system needs to shift too, at a meta level. However, let's say for now STEM is the best we've got. Just to improve the STEM scores, AI can be used as a tool to improve the education.

For example, my company, 01.AI, is currently working with a joint venture in the country of Kazakhstan. This is a massive government investment to build AI tutors for all the students. This includes AI tutors that could basically make learning fun and make learning personalized. So someone who learns very fast, a super STEM student, becomes super, super fast. Someone who's having trouble is patiently getting chances to do the problem over and over until they get it right.

And then the personal tutor is as smart as any chatbot, but it doesn't let you cheat. If you ask what's the answer to a question, rather than telling you the answer, it will patiently walk you through until you solve it yourself. So this doesn't replace the teacher by any means. And that's one of the reasons this proposal got accepted. The teachers will become the natural companion of the AI tutor, because the ideal teaching should always be one-on-one.

This way you learn the most, but no country can afford one teacher per student. Now you can have one AI tutor per student. So a lot of the drills and homework, and also learning the materials, answering the questions, can be offloaded to the AI personal assistant. Then the teacher can really talk about bigger things, like how to become a better person, how to work with other kids, how to communicate clearly and better, the soft skills.

And also getting to know some of the kids, so that you're also offering a human side of the one-on-one coaching and mentorship. Maybe, for countries that can afford it, there will be even more teachers than now, because we all never had enough. So we're building this system in Kazakhstan. This is another example where someone enlightened, who understands, bets on the future, can do this now, just because the writing is on the wall.

But it's not here yet. Others who want to wait a year will see the PISA scores in a year. And I think they will be significantly better. I'm really confident, because in China, in the last 15 years, China has already fully installed AI. Not even the latest AI, not even chatbots, older AI has already improved the scores by, I think, quite a few points. So that's the basis from which we add all the new large language model technologies.

We have no doubt this will lead to tremendous improvements. So hopefully there will be one country, at least, in Southeast Asia that will give this a try. - If you walk from Myanmar to the Philippines, a distance of about 5,000 kilometers, you skip over Singapore, around 80% of the heads of households don't have university education. So your only hope is for that student to be exposed to a great teacher who can infuse imagination, ambition, and serendipity.

Because the kid spends about six to eight hours a day with the teacher. Unfortunately, most of the teachers are not as great quality as desired, because they don't make much. I'll give you an illustration. In Indonesia, there's about 3.5 million teachers and professors, most of which make less than $200 a month. So if you're a brilliant student coming out of Carnegie Mellon, you want to be a teacher in some remote village in Kalimantan or Papua, you get an offer from Google for $5,000 a month.

As much as you want to be a teacher back home, you're going to go to Mountain View, because you can put food on the table. So I guess my question is, how do you use AI to remedy this? And how do you allow for a country like Indonesia to transition to a better spot without dislocating the preexisting teachers so as to create sociopolitical instability, while we take the view incrementally, in terms of recalibrating the criteria and recruitment of better teachers going forward? - I believe the AI teaching assistant product can show improved PISA scores within one year.

That is something that will happen, and this is leaving all the teachers in place. One of the other features we have in the product is what we call a multicast, a multicast real teacher plus AI. So we actually project virtually a real teacher into the rural classrooms, so that the rural teacher and the students are together, watching the elite superstar teacher from a major city teaching the course in their own language.

At the same time, the kids are learning, and they're not just watching. They actually have a little device. They can interact with the remote teacher. So you've just brought in a superhuman teacher, enhanced by AI as a cartoon character, making learning fun, interacting with the kids, especially for the lower grades. That has two huge benefits. One is that it helps all the kids move forward, without just being in a position to use the AI tutor, but having a great teacher, a great human teacher, interacting with each kid.

Even though it might be one to 500 at any given time, it's all happening in real time. That's the first advantage. The second advantage, which is even more enduring, is that a rural teacher is in the room watching how the great teacher is remotely teaching the class, so that the rural teacher will also improve. I think that will bring the second-order improvement, in addition to the AI tutor just bringing up the PISA scores.

So I think this can and will work. But to be honest, what I do worry about is for this to have an impact on the GDP, it's going to take a long time, right? These kids don't graduate and get a job for maybe ten years. So it'll create a certain degree of assurance and confidence that eventually everything will be okay. But there may still be a J-curve. - What about the tertiary-level education? How do you remedy that, or how do you improve upon it with the help of AI? - Universities, right?

Yeah, that is a tougher one. I was invited to be an adviser to a brand-new university. I'm not going to name what country so that people don't second-guess what country it's in. I was an adviser to the chairman of the university, and that person is a visionary. He said, "Dr. Lee, come in. We're building something from scratch, and we can do great things with your help." I said, "Do you want to build an AI-native university?" He said, "What does that mean?" I said, "Well, that means AI is running at the core of courseware, AI tutors.

Everything includes AI by default, and AI is not added as a periphery, but it's central. Of course, it's not going to fit everywhere, but we do the best we can." He says, "Yeah, that sounds great." Then he says, "I have a faculty meeting. Even though we're a new university, I already brought in all the department heads and chair professors. I want you to meet them and share your vision." So I went into this room, and it was all famous retired teachers, between 60 and 80, I would say.

All incredibly prestigious, famous in top universities throughout the world. They flew in because of the attraction of working from ground zero. They are all well-intentioned educators, brilliant in their fields. Then I started talking about AI. Then I realized most of them had not used any AI. They certainly don't use it in their research. They don't use it in their teaching. Some of them haven't even used one as a consumer, because of maybe their dedication to their work.

Then I start talking about how I would use AI, and there was a lot of, let's say, a combination of either disbelief or reluctance, or even just not paying attention. So I just felt like this was not going to work out. So I told the chairman, "I'm not going to be able to accept this adviser position." So I think the core lesson from the story is that we found the young teachers in rural areas in Kazakhstan, in the elementary and middle school, while they have not had great education, they're not paid a lot, they're living in very poor places, but they're all happy and eager to learn.

They view the arrival of AI as a great thing, and I have full confidence, at least the teachers I've met, they'll embrace it. But I think these professors, the more esteemed and older they are, the more they will not give up the status quo. So I think it's going to be so much more difficult. You can probably start a new university truly from scratch, or maybe go to one of the more recently founded universities. Like in the UAE, I helped the founding of MBZUAI, which is His Royal Highness's namesake AI university in Abu Dhabi.

That was built truly from the ground up with a focus on AI. Kazakhstan is now building some AI universities. So I have high hopes for these, but how do you modify existing ones? It's so entrenched, and the people don't want to change, by and large. I'm sure some do. And also, gluing AI as an afterthought isn't going to create the change needed. But I think there are certain departments that might be suitable to talk to.

For example, I was talking to a department chair of computer science, and he said, "Kai-Fu, we're going to have to change the way we do things, because our enrollment, our application pool is down 30% from last year, because people are hearing that a computer science engineer can't easily get a job anymore. So they're shifting majors to more mechanical, electrical engineering, which still isn't replaced fully by AI. So what should we do with our curriculum?" So I guess if I were a university president, I would say, take a look at which of the department chairs are in crisis mode, and which department chairs have the vision to change, and maybe just do some experiments based on them, and see how far we can get.

Trying to do a more broad change throughout the university, you're fighting against an infrastructure. I would have to really learn more about universities to know how to change that. As I recently wrote the book AI Native, that's about how to change a company. It's the same problem, but at least the company people are driven to make the company great, to improve profit, revenue. So as long as my methodology leads to that, there will be openness.

And also, in companies, people are a little more willing to let go of the way they used to do things, because if there's a new technology, embracing it leads to better numbers. So I think companies will be easier than universities to change. - Right. Kai-Fu, we've talked about the intersection of AI with academia. I want to just paint the picture generally about Southeast Asia. We've seen declination in the ratio of manufacturing to GDP, the middle class to the overall population, agriculture to GDP.

This would have been on the back of perhaps rotation into services. But I'm not witnessing fantastic up movement on the value chain within the service sector. And we've seen automation, robotization, and now we're imminently seeing physicalization of digital know-how into physical activities, right? I mean, intuitively, the risk of further, if not disproportionately more, dislocation of jobs is there, right? How do we cope with this type of...

Because there is a political ramification here, to the extent that you don't have economic stability, social stability, you're not going to have political stability, right? Like in Indonesia, we've got 150 million workers, of which 75 million workers are informal. The other 75 are formal. The risk of dislocation with respect to both the formal and informal is not to be underestimated once robotization kicks in in a pervasive manner.

We're seeing in the Philippines about 1.5 to 2 million workers that would have been busy with the call centers, the BPO and all that, which is, I think, systemically at risk with AI coming on board, right? How do we prepare for this so that the transition could be, albeit with incrementalism, but judicious? - The invasion of the AI workers will begin with cognitive work, as we've already seen. If it can do software engineers' high-paying work, it can surely do call centers, BPO, accounting, legal, so on and so forth.

So for people holding cognitive work, the writing is on the wall, and they need to fly to safety. And then after that, I think, basically, the blue-collar workers will have a similar problem. The gap between the two might be five to ten years, because AI's ability to overcome the clumsiness of their body and fingers is going to take some time. And also, there's a cost, right? Software is just compute, which is a modest cost that's dramatically reducing.

But the hardware isn't going to have the same curve of cost reduction that software had. So I think five to ten years is probably about the right range of estimate. But not everything is replaced at the same time. So let me first talk about cognitive work. The types of work that will remain are those that have a very strong people component and a very strong openness component. By people component, I mean most of the work is interacting with other people, gaining trust, negotiating sales, solving problems, and things like that.

And then openness means you're working on a problem that isn't about a fixed, repeated problem where you're more or less solving for the same thing each time. So a closed problem would be filling out a reimbursement report. A pretty closed problem would be customer service, but an open problem would be designing a marketing campaign or working out a company merger. So that's the range for cognitive work. And I would be the first to acknowledge that the people-oriented work is where the volume is, because there aren't that many open jobs, but there are a lot of people-oriented jobs.

And I would also argue that, as AI starts to take over the routine execution of tasks at work and in services, people are going to want real people for the services. They're going to long for that human touch, whether it's an existing job like a concierge, or tour guide, or a customer service rep, or partnership specialist, that kind of thing. But also, there will be new opportunities, like a new-style teacher to help kids gain confidence and work with other kids.

People helping to deal with anxiety, therapists, nurses, and also maybe newly created jobs. Perhaps jobs that are unpaid today, like volunteer at an elderly home or volunteer at a foster home. And maybe even at home, homeschooling might be something the government can pay. Where does the government get the money? Well, tax the rich, especially those who got rich on AI, and use that money, convert it into a universal basic income, and encourage people to take on this open and human-oriented work.

And to be more specific, there will be many more human interaction-based service work that will be going up. And I would say domestic help is another area that I know is a significant part of Southeast Asia. Now, going back to your question about blue-collar labor and the physical work, I think there's a little bit more time, so five to ten years. But also, in the same way, look for openness and look for human touch.

So what's a human touch in physical work? Well, a plumber, someone who does disaster relief, someone who fixes things in a hospital. These things all have a human component, and AI can't do that, because even if AI could fake it or emulate it, people won't accept it. They want that human-to-human touch. So the human orientation. And what about openness? An assembly line, that's clearly not open. That's a very closed, repetitive process.

And picking out boxes in Amazon, that used to be kind of not closed, but now it's being closed because robots are getting good enough, because packages are generally similarly rectangularly shaped, and AI could do it. Folding clothes used to be in the open area because every dress is different, but now AI is getting pretty close to figuring out how to fold clothes. So one at a time, AI will encroach from closed domains to open domains.

But there are some that are truly in the very, very open area that AI will take a long time to be able to do. So disaster relief would be the extreme open, because every disaster is different, every location is different. So that remains for humans to do. But another interesting one I brought up earlier: plumber. That sounds like a closed job. There are only so many permutations a pipe could go wrong. However, because the pipes are hidden, it's almost like a detective job.

And also the human interaction. So you have to decide when to take a chance to open the wall. And what if you're wrong? How do you explain to the homeowner that you broke their walls and their pipe's still not fixed? And then when they get mad, how do you respond? So it requires both an ability to navigate an open, unknown environment, which is, every house is different, as well as an ability to interact with people and to calm them down when they're not happy.

So these are the kinds of jobs that will remain. I cannot be sure there will be an equal number of jobs that remain at the end of the day. There's some likelihood it could happen, because AI is also creating new jobs. One of the new categories of jobs is labeling data, because AI needs to learn on massive data. So label the data that AI hasn't learned to label yet. I should be specific, because every time AI is trained, it gets better at labeling the data for that category.

Then you've got to look for the next frontier. So today, a lot of people are being paid to collect data for embodied robotics. That is, you wear gloves, and then you're given an instruction, and you execute the instruction, like pick the glass up and put it to your mouth. And then that is captured by cameras and labeled so that it can be used for training AI models. So AI's need for labeling will continue for a while longer, and who knows what other jobs AI will create. - So to the extent that AI survives and thrives upon pre-existing codification, any profession that's less codified would be less vulnerable to AI.

Is that the right way to think about this? - Yeah. So I think one could say that computer software is pretty codified. It's well understood what a section of code does. You can execute it, simulate it, test it, read it, examine it, debug it. One would also argue understanding law would be in that category. - Understanding tax, accounting, law. - Yeah. So it's more than you think. So I would say it's not just closeness, but it's also the following properties: any domain where past data can be predictors of future data, any domain where verification of correctness is doable, because when that's possible, AI can self-generate data and teach itself, because it can just test itself, see if it got it by verifying, and so on and so forth.

And then also, as you said, the structure and codification that makes it more replaceable. - You alluded earlier that compute capacity is continuously rising. There is also a camp that believes that the cost of computing is coming down, but the cost of transporting data may not be coming down as quickly, right? Do you see imminence in how these two, at the rate they're declining in terms of cost, in terms of increasing in terms of efficiency, do you see these two reconciling with the scarcity of energy in most developing economies at some point? - Yeah, I think the major dissonance is with the cost of energy, right?

Because that's increasingly a larger percentage of inference workload. That is, when you ask ChatGPT a question, it calls an inference engine, which is the GPT model, to answer your question. It's used as opposed to training, which is the process of AI learning its model by itself. So the inference cost is coming down, but the usage is going up. So these things are causing a greater need for data centers, and a lot of people are investing in data centers, and they're deploying a lot of GPUs.

But the cost of electricity is increasingly a larger percentage of the inference costs. And as you mentioned earlier, I think there will be many countries in Southeast Asia, but also the US, the current electrical grid is at capacity. And once you add too much, it will start to break and not be able to serve continuous, most reliable electricity. And that would be a disaster for whatever country in which that happens.

I think the right answer to that problem has to be massive investment in green data centers. Because that actually is a challenging problem, but a typical green data center uses solar and wind near the data center, so that it becomes largely self-reliant. So you're not adding too much more load on the electrical grid. And the other thing is to further make a stronger and larger electrical grid, which has some capacity limit everywhere.

So the US, as an example, I think, is not far from the limit of its maximum on the electrical grid. And then there's some deployment of solar and wind. China, as an example, has much more headroom on its grid, as well as the world's lowest costs in using solar and wind. So in that sense, China is ahead of pretty much every country. And I think countries have to start planning, when they don't have enough data centers: one is how to get people to invest in them.

Two is how to make sure the data centers are used, because some countries are not getting the same adoption of AI as many of the other countries. And then thirdly, how will the electricity problem be solved? And I really think it's going to become increasingly obvious as the green energy cost keeps coming down, because green energy is effectively the cost of making solar panels and a wind turbine. And as with any manufacturing product, the more you make, the cheaper it gets.

So if you buy Chinese solar or wind equipment, it will probably get cheaper and cheaper. At some point, it becomes the obvious choice over the other options. - China has already built about 4,000 gigawatts. It's planning to build an additional 4,000 gigawatts. The US has built about 1,400 gigawatts. They're planning to build an additional 2,600 gigawatts. Relating back to how these big AI companies are doling out, or planning to dole out, I don't know, $3 to $5 trillion just for data centers and GPUs.

It seems like a closed loop, right? It doesn't permeate into the other sectors of the economy. Do you smell a potential bubble at that rate, and at the rate that there's not enough energy intensity? - Yeah, it's possible there is a bubble. I'm not an expert at the data center capacity and investment. But given the heat, the excitedness that people are jumping into this with a large percentage of their net worth, it seems like this has basically the prerequisites of a bubble.

I remember back around 25 years ago, there was a huge buildout in the cables. So there was a belief that there would be cables everywhere, and especially driven by the US, the cables would be the primary carrier. There was not much Wi-Fi back then, so homes really got their internet connection through the cable network. And then the cable needs to be connected continent to continent, so the sea cables were being built.

So there was a huge buildout. A lot of the cable companies were worth a lot of money. And then eventually, what happened was most of them went out of business. But actually, the carcass that they left, which are the cables that got built, became the backbone of our internet. So we might see the same thing again, because this is compute we will use. But to the extent a bubble's created, the companies will be punished, but the assets are there, so the assets will be sold to people who can operate them at a much lower cost.

Then the assets will deliver their value. So I would say the worst case is, if there is a bubble, the compute resources, data centers being built, are not going to be unused, and they'll be contributing, as long as they're built correctly, to the future growth of AI. - Kai-Fu, I've read your books. You had a personal episode more than ten years ago. You were diagnosed with a tumor. - Yeah. - And then I've noticed that your recent book smells differently from the earlier books.

There's more humanity in your tone. Feel free to talk about how that experience changed the way you look at technological innovation, particularly in the context of AI. And the second follow-up question would be whether you can predict that, at some point soon, AI will be sentient. - Okay. Yes, I was always a workaholic, especially during the days where I was starting new things like Microsoft Research, Google China, Sinovation Ventures.

These were all brand-new endeavors. Even though I wasn't necessarily the founder, I was the founding managing director or founding VP of all of these organizations, whether they were multinationals or my own. And in these periods, I worked incredibly hard. And also, I always held it as my motto. If people say, "Well, what is your motto? Why do you live?" My answer was always, "I want to make the greatest amount of difference to the world." I measure things by something a college professor once said to me: "Imagine there are two worlds, one with you in it, one without you.

Your life's purpose is to make the former much, much greater than the latter. That's an easy way to measure your personal contribution." Well, it all sounded great, and that was what I used to drive myself. And when I had difficult decisions, I would ask myself, "Where can I make the greater difference?" And invariably, it always shifted to the work side. It always shifted to my career. And my family and even friends rarely figured prominently in that equation of making the biggest difference, until I got sick.

And then I saw, really, that my family has stayed with me, took care of me, and that their love for me was unconditional, and that I was giving them second priority in my choices, and I was wrong. Also, I had the chance to meet with a very, very famous Buddhist monk. His name is Master Hsing Yun, in Taiwan. And there he asked me what was my motto. And I told him, and he said, "Well, I think it's very dangerous." And I said, "Why?

Why is it dangerous to maximize your difference?" He says, "Well, if you measure everything quantitatively, then you will not be sure if you're really trying to make the world a better place, or if you're being driven by greed, greed for money and fame." Because every effort for greed and fame: "I want to go for that promotion. I want to start a new company. I want to make a great investment because I want the world to be better." But is it really that, or is it because you want to be more famous and richer, right?

"I want to give a lecture to the largest number of people. I want to write a book read by the largest number of people." Those are reasonable ways to measure impact. But when you decide to go for the largest number, biggest impact, is that to satisfy your ego or to serve the world? And what Master Hsing Yun said to me was, "You should ask something a lot simpler. Everything you do, is it good for the world? As long as it is good for the world, do it.

But always ask at the same time: what's in it for me?" And the less there is in it for you, the more assured you can know you're doing it for the good of the world, not for yourself. And that's not to say you can't do things for yourself, but always be deliberate and thoughtful about these two things. Don't let "make the world a better place," "make a difference in the world," become an excuse to get stuff for yourself.

So that combination of these two things really caused me to rethink and bring a human side to this. And that's why, when I wrote about AI, in fact, every book I wrote about AI, I talked about things like, "When AI becomes smarter than us, what's left?" Well, what's left is our right brain, our emotion, our conviction, our love, our taste, our empathy, and our courage. So the human side of things, which is more or less what the right brain generally characterizes, becomes more emphasized on a day-to-day basis, even though my job is building AI products and selling them.

But I now feel that the right brain, especially love above all, is so much more important to us, and it's the reason that we are here. And when we remove intelligence as the single thing that differentiates us, the replacement, or what should have always been there, is really love. That's something AI can never take away. One could argue AI could become so automated that it could have taste, or learn empathy, or have conviction, or learn to do things that are counterfactual.

But one thing it won't have is... I really don't think AI can love. And so the question of, can AI become sentient? Can AI fake sentience? I think probably yes. In five years, it could fool almost all the people. But is it really love? And some would argue that's close enough, but I would argue that's not real love. And also, I think even if AI fakes it pretty well, people won't reciprocate, at least within the next couple of generations.

People are still looking to love people, not to love robots. So as much as I'm super optimistic about AI's advancement and value, I remain highly confident, not 100%, but highly confident, 95% confident, that our definition of our sentience will not likely be replicated by AI in the next 100 years. - One part of our earlier discussion was with regard to the fallacy of AI, in that it doesn't have accountability. And number two, we've seen sort of the shifting of LLMs spitting out stuff that's more politically correct as opposed to the truth.

How do we train humanity that, at the end of the day, the accountability onus is on users? And then how do you train humanity so that they can actually train the model, the platform, to seek truth as opposed to political correctness? - Yes, two different questions. So on the first one, I think people are a perfect counterpart to AI, because AI can execute at the speed of lightning with a high degree of accuracy, but if anything goes wrong, it can't take responsibility.

And anything important in life, someone needs to take accountability, and that's where the human comes in. Of course, we don't want humans to be a scapegoat, right? To be a rubber stamp and a scapegoat. Whatever AI does, I say, "Okay." Something goes wrong, "Okay, I'll take the fall." You want people who are actually driving, running AI with a purpose. And also, they are rewarded when things go well, and punished when things are not.

And this, in my book, is called a DRI, directly responsible individual, a term coined by Steve Jobs. So when you have the right kind of person who has a strong sense of responsibility, who's trustworthy, who's very deeply knowledgeable about the task, he or she is actually running the AIs and understands the AIs, then I think we've got a really good thing going. We have basically all the right characteristics and knowledge of AI combined together, and that person should be empowered to run that particular goal or task for the company or for the startup or the OPC.

So that's the first part. And the second, sorry, was? - Basically being cognizant of the fact that platforms are becoming political-correctness-seeking as opposed to truth-seeking. How do we caution humanity about this? - Yeah, that is the danger of social networks, of what they have become, right? We've seen Twitter, as it becomes X, it actually changed its appearance. It theoretically could target people, so that it keeps people in different camps.

But also, it took on a view of its own that was different from before. So if that's a danger in social networks, it's an even greater danger in AI, because social networks, we scan and we swipe, and we look at a lot of things, but AI is directly giving us an answer. If that answer is biased or incorrect, it will lead to very serious consequences. So the first thing I think we can do as individuals is to update our prompt.

So I have a prompt that removes, as much as possible, sycophancy, that is, kissing up to me. I think one could update and modify the prompt to also maintain a balance between the extremes or present multiple points of view, sort of the ideals of journalism. So that kind of a prompt is probably 200 words, and it will cost you a bit more tokens, but it's worth it. You don't need to copy-paste it into each one of your queries.

In both OpenAI and Anthropic, there are settings where you can set it for each query. So for me, I cared more for removing sycophancy and also for reducing hallucinations, and stating when it knows the fact, or thinks it's right, or is guessing. I think having these two helped me greatly. I think it's not very hard to engineer another prompt that would include maintaining balance and respecting different viewpoints, and even representing different viewpoints, maybe coming from the doctrines of journalism.

So someone could create that prompt and share it and use it. Now, the prompt may not be enough, because sometimes the model has been baked in with this political view, right? Elon Musk claims his models are more programmed to tell the truth. I don't know if that's true or not, but it is certainly possible you can program it to be pro-leftist, pro-right, or centered. So the prompt I said can course-correct, but if the insight's still wrong, your course correction can lead to unpredictable results.

So that's why I think having the open-source model helps, because a computer scientist can use a set of training data to tweak a model with whatever biases it has to be as neutral as you want. Of course, you could also do it the wrong way and make it even more biased. But that's the power of being able to tune models when you have parameters and the model on your server. When you use cloud or OpenAI, you do not have that possibility. - Kai-Fu, I'm sensitive of time.

I'm going to ask two more questions. The first one is with respect to how you aptly admitted that you were incorrect in predicting how quickly AI was going to attain or manifest creativity, right? It came a lot sooner than you predicted. I guess my curiosity is with respect to, how soon do you think we're going to be able to attain AGI? - Yeah, I made the prediction that what humans have to protect ourselves were two things.

One was the human-to-human connection, which I still maintain. The other was creativity, which I stated in 2018. Today, I think only the likes of Picasso, Mozart, cannot be created by AI when that style didn't exist. But that's going to be a very, very tiny percentage. Most of what we call creative, writing a new novel and things like that, poetry, certainly AI has the ability to do, and better than most people. Sorry, you're asking about creativity and... - AGI. - AGI is a very ill-defined term.

If you ask any large language model, it would tell you there are two sets of people. One set strongly believes it can do everything a human does, becomes a superset. The other is more practical. And if you believe in the former, I would have to say there's no date whereby I can put AGI, because by definition it includes sentience. And also, there are many things about us, our sentience, self-awareness, we don't even know how they work in our brain.

How are we going to measure it? We don't even have a clear scientific definition. So I think that definition of all-encompassing AGI is impossible to measure and extremely hard to do, whatever it is. So I wouldn't want to put a date by that at all. I don't even know what it is. And in its simplest definition, I would say it's probably never that we would get there, because we're not a full superset of an ant, right? And so I think a superior species does not need to be a complete superset of an inferior species.

So if you want to take that view, I just don't think superset is a useful way to measure AGI. But the AGI definition I like, that I usually quote, is, for cognitive tasks, when AI can do 90% of the human tasks better than 90% of the people, I think that's close enough to call AGI. I think we're not there yet. I think we're already at the point where AI can answer 90% of the human questions better than 90% of the people.

We're already there. But answering questions and doing things are quite different. I think we're probably still maybe two years away from that. But we'll get low-hanging fruit pretty soon. And then, of course, one could say AGI should also include physical robotics and embodied AI. So that adds another five to ten years, in my opinion. - Last question. Can you re-illustrate Boss AI again, the way you did it a few weeks ago when we were chilling out? - Yeah.

So one of the things I talk about in my book AI Native, as well as ship in my company, called Boss AI, has the following concept: that AI can be a tremendous management tool. A CEO sitting at the top of the company with tens of thousands of people or thousands of people is too distant from the truth. The truth is sometimes buried somewhere. And also, the hierarchies of management, the larger the company, the deeper the hierarchy, the less the boss knows the truth, because the information is passed up the chain in a very lossy kind of way.

People selectively share things with the boss. They may be the bearer of good news, but not the bad news. Or they see a problem, they may not want the boss to know. They want to kind of cover it up or solve it. And actually, what happens is, sometimes by the time the boss hears, it's really blown up. And then they come back and say, "Hey, why didn't you tell me earlier?" And you say, "Well, I'm trying to fix it." And they said, "Well, you should have told me earlier, because now it's too late to change.

Had you told me earlier, I could have fixed it." So these kinds of things continue to really make the communications very difficult. And also communications down. A strong CEO wants nothing but for his or her strategy and direction and orders to be followed and executed. But when every middle manager takes some inputs, he or she will first consider, "Well, what does it mean for my department? How should I communicate to my department?

How do I get benefit for me and my department?" And once you enter that kind of level of selfishness, there will be sugarcoating, modifications, and then the execution may not line up. So Boss AI is intended to give the CEO the God's-eye view to all the data in the company, not only CRM, ERP, but also communications data. That means corporate email, corporate chat, and most importantly, corporate meetings. So I started using Boss AI myself as the first boss, and it's been tremendously helpful.

So my Boss AI really virtually attends every meeting, because in my company, we ask every meeting to be recorded. And now they can be highly accurately speech-recognized, speaker-recognized, and with the script of what was recognized, AI can virtually attend every meeting on my behalf and answer my questions specifically or generally. For example, I could say, "In the thousand meetings my company had in the last week, especially those I didn't attend, what are the top three debates that were had that are important for me to know about?" So it tells me new things.

Or I could say, "In this very important strategic directive in the last month, have there been any signs, whether in meeting notes or emails or chats, that should make me worried that we might not meet the milestone?" So it can give me that. I can ask, "What are the three biggest legal exposures as we apply for IPO?" I could ask, "Who are the people who are best at AI programming, or actually managing AI to write programs in the company?" Then I was surprised some of the non-engineers showed up.

And I said, "How many of them are very generous in sharing with other people?" Because I want to make them the DRI, or give them more responsibilities, and trust them to mentor and coach more people, because they're so good at it and willing to share. And then I could ask, "How many of them might be retention risks? They might leave the company." AI knows because they may have shifted from very proactive, signing up for things six months ago, and now just saying nothing and staying quiet in all the meetings, which is an indication they're not as engaged.

Therefore, retention risk goes up. So all of these things can be answered. And before every meeting, I would be told who's coming to see me, here's their PowerPoint, here's what they are likely to ask for, and here are the questions I should ask them in order to make my decision whether to give them what they ask for. During the meeting, the person's speaking, and with a little bit of a delay, they're fact-checked against the realities.

If they claim sales are a certain number, headcount is a certain number, AI will check if they're off, right? And then I would know whether to reduce my level of believing in him. Then after the session, there are action items and also a pledge, basically a promise ledger, of who promised to do what. And then AI will continue to nudge and remind people who promised to do things, whereas before, they might have hoped that I would forget.

And I forget, but Boss AI doesn't forget. So as a result, my company is managed much better. I sleep better at night. I know what's going on. I can manage processes, and I do so transparently. I tell my team, "I am using this data, and here's how I've used it, and here's how it's helped the company. You're all shareholders, so you're better for it." Because I know some people might be concerned whether I would use it to look for things that are maybe unprofessional, or invade their privacy, or something like that.

But I show a lot of positive examples, and everyone's incentivized by the company stock going up, the company doing well. So the people accept it. I will finally make a note that people from Europe and sometimes the US are not so ready to embrace this style of recording everything. But in Asia, I found pretty broad acceptance. - That's a separate podcast. I'm thrilled to talk about that, but I'm sensitive of your time, Kai-Fu, this has been fascinating.

When is your book coming out, AI Native? - It's out in mid-September in Kindle and then early November in the hard copy. - Okay. I'm so looking forward to getting a copy and reading it. I've enjoyed all your books, especially 2041. - Oh, thank you. - I look forward to seeing you. Thank you so much, Kai-Fu. - Thanks so much. - Friends, that was Kai-Fu Lee from 01.AI. Thank you.