Kai-Fu Lee: Why America's AI Strategy Backfired | The Extended Cut

0:00Allie Garfinkle: I'm Allie Garfinkle and we are here with Dr. Kai-Fu Lee. Dr. Lee is one of the foremost voices in AI. Former president of Google China has held senior roles at Apple and Microsoft. Founder and CEO of O1.ai. Dr. Lee, Kai-Fu, thank you so much for being here. Thank you. We have decided that we're going to kind of have an interesting format here. Here at Fortune, I focus on investors, Silicon Valley. My colleague, Emily Forlini, our senior AI reporter, is an AI expert. And there's always AI news, big AI news today. Emily, we were slacking about this this morning.

0:36Emily Forlini: Yes, this is an enormous story, global story, so we wanted to ask you about it. NVIDIA is investigating its own CEO, Jensen Huang, for having an affair that he has owned up to. He did do this with one of his customers. So she is a data center executive in Asia, and they have had an affair, and she's a suspected spy for China. So what is at stake here for NVIDIA, for the AI industry, and for U.S.-China relations? Well, I just read the news, so I really have nothing to say and no comment. Yep. This is definitely breaking right now. I thought maybe you could just give a little bit of context on the backdrop of some of the tension between the U.S.

1:17Emily Forlini: and China in terms of chips, which is really at the center of this. So it is public that the U.S. has restricted export of the most advanced NVIDIA chips to China. Why is that and how has that landed in China? Why is that such a controversial issue? Well, I think the whole LLM revolution began with the belief that whoever has more GPUs can train on more data, will build possibly AGI that will squash every other company from any country. So with that belief, then making sure that American companies have the most advanced chips and most quantity of chips seemed like a plausible strategy. But it ended up being completely ineffective because it actually did restrict the best chips from China.

2:06China did have to scramble. Chinese companies, I should say, did have to scramble. But in the scrambling, people put their engineering ingenuity to tweak everything to be 10 times faster than the American models. And I think that has led to a set of open source models made in China. That's maybe six to 12 months behind the American models, but it was made cheaply and it's now free for the world to use. So now we're in an iPhone and Android situation where the American companies have closed models, just like the iPhone. They make a lot of money. Chinese companies don't make nearly as much money, but they're going to get the footprint and the market share.

2:49Emily Forlini: And also the open models in China, if you look at websites like Open Router that have rankings of the most used models, the Chinese open models are consistently at the top. So why are people using open models so much worldwide? And when would they use an open model versus a closed model? So, first, I think the top American companies price their tokens extremely high with very high margins. It's somewhat understandable because they spend so much training, they need to recuperate the cost. But when the Chinese companies have figured out how to build nearly as good models for one-tenth or even less the price for both training and inference, then that creates a natural alternative.

3:34So the types of people who like the open source models are the following. People who can't afford or don't want to pay the exorbitant prices. Secondly, they might have a less demanding task for which the open and closed models are equivalent. And third is a really big advantage is that the open weight models you can copy to your server and run it on premise with no data leaking to a large language model companies. And we have seen throughout the short history of LLMs how vertical application companies come out with a very cool product, but that's built on using closed models by sending lots of tokens to the closed models.

4:20And then the closed models gets all the data and train on top and eventually eliminating the business proposition of a startup. So with news like that really understood by enterprises, and also enterprises view their corporate data as sensitive and their competitive edge. So they just don't want to send their data out to a large language model company. They might be willing to send it to a hyperscaler, but then there are obviously people who are okay sending it to the LLM companies, closed models. Some will only send it to hyperscalers, and some really want to keep it on site. So the third group will move towards the open models.

5:09Allie Garfinkle: I am really curious. I feel like there's a lot of noise, there are a lot of headlines, but there's not a lot of context about the AI race between China and the U.S. What's fact? What's fiction? Where do you sort of feel like there are misunderstandings, either in the U.S. or in China? I think there are a lot of misunderstandings. A lot of them. Yeah. So first, there is a belief that the Chinese companies can't make money. But if you go back to the history of mobile internet, as I described in the AI superpowers, the Chinese companies are born to make money. That's the primary purpose. It's not so much to win AGI or win the race.

5:53So in the cases of Pinduoduo, Meituan, ByteDance, these were all companies that were questioned about how they would make money, but they figure it out. So you may reasonably doubt China's ability to do breakthrough, which is valid. The US is way ahead on that, but never doubt Chinese ability to find a way to make money. So that's the first misunderstood area. The second one is there's some belief that China is an entity, that the Chinese government tells the model companies to do things like do open source. That's the way the national strategy wins. This is complete non-existent, made up nonsense.

6:39The Chinese companies are venture funded, just like the American companies. They have Chinese investors, American investors. They just want to serve their investors. And what really happened with the Chinese open source phenomenon was not by design, not by a company, not by the government, but rather when DeepSeek moment arrived that the best model in China was open source. That really raised the question for all the other LLM companies. Can you be DeepSeek? If not, get out of the business. If you can, you must open source. Because there's no situation where an inferior closed source model will ever have any opportunity.

7:17So everybody follows suit and open source. And then the history began and the rest followed. Most are still staying in the race to be the best open model. I think that's the second misunderstood thing. The third is that export control is what limits the Chinese spending on GPUs. It is not. The Chinese companies are naturally parsimonious, not just the startups. It's a piece of culture. Well, it's a piece of culture, yes. And it's also about prudence and survival instincts. That place is higher than the AGI dream. So a startup company is measuring its budget very carefully, raising money when there's a chance, using it prudently.

8:04And now they have money to last three or four years, and we'll see what happens. And then the larger companies, Alibaba and Tencent, and they basically spend a fair amount on training their models, but nowhere like the American companies.

8:20Allie Garfinkle: Do you think the American companies are spending too much? What does too much mean here? Yeah, so here's the American value proposition, open-air, and anthropic. To them, the most valued asset is talent and talent density. Very understandable. To attract the super smart PhDs, you've got to give them GPUs. and you don't want to force them to tweak infrastructure, firmware. That's not what they want to do. So to make these brilliant researchers happy, the cost of training is not just the cost of training. It's the cost to let 1 ,000 people run a million experiments, which costs perhaps five or more times as training.

9:02But that's the American path. And the dream is that this team of genius will make an AGI that would become the most profitable company and the most valuable and remembered company in the world. That is the Silicon Valley model for open AI and Anthropic. And I respect that. And I think it's logical. The Chinese model is more like a study group. None of the companies have the talent density of Anthropic and open AI, but all of them have pretty good talent. So because they can't really compete alone. They do compete against each other, but they're basically open sourcing their, or at least open weight, their model.

9:48And that allows each other to examine the model. They're actually publishing papers when the American companies are much less doing so. So the papers, the talent moving around in the companies, plus the open weight models that you can examine causes a rising tide to lift all boats. So there are actually quite a number, I would say four or five very decent Chinese open-weight models. And that's how it came about, is DeepSeek's success, forcing the hand of every company to have no choice but to also open weight or drop out.

10:25Allie Garfinkle: So this goes back, there are a couple of things here. I mean, the first is, it sounds like the DeepSeek moment was as big a deal in China as it was. here. Well, bigger in China. Bigger in China. It was a pride, a national pride for a lot of people.

10:37Emily Forlini: Was it all over the TV? Like, what did that look like? It was everywhere? People were talking about it? It was cleverly done just before Chinese New Year. So a classic. And then in Chinese New Year, everybody goes home. And then the kids tell their parents and grandparents and just spread like wildfire. And it was previous before Deep Seek, it was believed China was way behind, maybe two years, two and a half years behind the U.S., and even longer before that. But the moment DeepSea came out, it was like one of the top three or four models in the world. And I think that was fun to use. It was good in Chinese.

11:15And then it's a national pride thing. So people say, okay, we thought we were never going to catch up. Now we have a shot. So it was quite a special moment.

11:24Allie Garfinkle: A lot of the VCs who I cover at the DeepSeek moment were truly a gog, I believe is the term, over DeepSeek because of the low costs of training it. And as we sort of think about the news, Anthropics S1 dropped a few weeks ago. The company is set to go public. I am really curious. Anthropics is a spending story for so many reasons, right? When you look at those financials, not only the revenue, but the massive net losses, those massive costs of compute, what do you see? Well, I've heard Dario speak that if you remove the cost of training the last model, the balance sheet actually is quite healthy.

12:06And that is a valid way to look at it, right? You still have to train the last model.

12:10Emily Forlini: Yeah. Well, the last model is - They keep training new models and new models, and they kind of replace themselves. So a lot of people are talking about, is that sustainable? Yeah. So his explanation is training the last model is going to bring revenue in the next two years. Then spending more money to train that model will make money in the two years after that. So they see a two-year horizon. Sometimes the models are coming out every couple of months. Yeah. And that's the other thing that anthropics should rightfully be proud about. It's because they have figured out some sort of RSI, recursive self-improvement.

12:48So AI is improving AI. So when he actually did the interview where I quoted, I think it was truly one and a half years per release. Now it's going faster and they have better RSI than a lot of other companies.

13:05Emily Forlini: So you think they have RSI, so RSI is here? They have some, well, they're still human in the loop. But when the model is, let's say, 80 % trained by AI, 20 % by the people, it'll be trained three or four times faster. So the improvement will actually accelerate. So these are the valid arguments for the anthropic valuation. Another valid argument is that there are a lot of American enterprises who are willing to pay the exorbitant prices. And that creates a very amazing growth of the revenue stream. If it weren't for the big spending and losses, it would be a beautiful spreadsheet, P &L. So these are the valid cases.

13:50Emily Forlini: That's very interesting that you think the first thing you said, that their valuation depends on them developing RSI, recursive self-improvement. Partly. Partly, yeah. That's interesting. I have noticed they're putting out more papers to show that they are, and maybe that's perfect timing. So I think these are the valid reasons. So I can see reasonable people making an argument, they're buying. But on the flip side, there are also issues because one, their model is better than the open source models, but not by a huge margin by six to 12 months. So people could say, I'll settle. I'll take a slightly worse model and save a bunch of token costs.

14:31and also their models are kept secret and then people would say, hey, I want on-premise. I hear that increasingly in the enterprise users. So the question is, will the enterprise users keep using more tokens or will some of the tokens move to the open-weight models? So will their growth really go at, they clearly won't grow at the same rate they've grown the last two years.

15:01Emily Forlini: And they said that in the S1. They said, that's one of the risk factors. They said, we will not be growing at over 100 % going forward. So they were at 46%, I think, between Q2 and Q3. Well, that's still a great number. It's an amazing number. Yeah. And one of the big things though was compute. So they said they're going to spend almost a trillion dollars in compute, which is a crazy number. But then we also had the midterm elections where people in the US are saying that compute is going to become slower because they see the Democrats in the U.S. as having less favorable attitudes towards data centers specifically.

15:34Emily Forlini: And there might be more regulation, more permitting, communities getting more support to push back on those products or those projects. So how was that received in China? I wasn't in China during this period. I think people - No chatter? I actually don't think people watch that news. Okay. That's good. It's really, I mean, every country is more self-focused unless it relates to, you know, two countries competing or something. But actually another flaw with the NVIDIA argument is at some point, everybody will reach quote unquote AGI or quote unquote RSI. This is not some magic that the Chinese companies can't figure out.

16:17So at some point, you're still 5 % better than the best open models. And then next time, you're only 2 % better because at some point, you will reach diminishing returns. So maybe that's why they want to spend a trillion dollars. That might give them a longer window. But this is all speculative and questionable because when the best open model solves 99 % of the problems, why should someone pay for the super expensive model?

16:47Emily Forlini: That's a very good question. So I just laid out the two arguments, and reasonable people can disagree on whether Anthropic is a buy or a sell.

16:53Allie Garfinkle: Well, that's the sort of thing, too, where a window isn't a guarantee. And I looked at some of those costs of compute, and little old me sitting here who's read a lot of S1s, I thought, wow, okay. Is AI a bubble? AI is absolutely not a bubble. Absolutely not a bubble. No, no, no. The value of AI... Even if Anthropic ends up being overvalued, let's say, it's not a bubble. I think any given company, Chinese or American, can be overvalued at any given moment. And some may actually go out of business so that they create a bubble that didn't self-fulfill. These are all possibilities. But as a whole, as a whole, I think if we look at this technology revolution, it is so different from the previous ones.

17:36Because every technology revolution is a technology becoming commoditized. therefore enabling myriad applications, right? So the PC revolution commoditized compute, internet revolution commoditized information, mobile revolution commoditized presence, and now AI is commoditizing intelligence. So how much bigger is intelligence more than information compute and presence? It is, you know, 10, 100, 1 ,000 times bigger. So this will bring about a whole new way of doing work, of communication, of entertainment. Many companies will go out of business. Many super valuable companies will emerge. Many people will be out of jobs, but many great new jobs will emerge.

18:25So this is the most tumultuous moment that humans have ever reached, where we have the opportunity to reach for the stars, but we might also fall into the abyss.

18:38Allie Garfinkle: To this end, I'm very interested in some of the comparisons to the Industrial Revolution. And you lay it out really clearly in your book. You write at one point, when we speak of history, the first Industrial Revolution is often cited as a turning point for our species. AI could far surpass it in speed, scale, and scope. This sounds like what you're talking about. What did you mean by that? Well, I mean, Industrial Revolution arguably commoditized some muscle. And now AI revolution is commoditizing intelligence. And think about how most of us think about how we differentiate ourselves from other species, animals, et cetera, and how people differentiate themselves from each other.

19:22You know, I'm smarter than you. You're smarter than him. You know, people are smarter than monkeys. Dogs are pretty smart, but ants aren't so smart and so on and so forth.

19:31Emily Forlini: You mean monkeys don't do interviews like this? We don't know. We could. You can try that. in the future with brain-computer interface. Right, right. Yeah. But I really think that commoditizing intelligence has unbelievable ramifications. Not only will it make intelligence free-flowing like electricity and free, and we can create so much value in ways that we can't imagine, it will revolutionize everything digital first, cognitive next, and then physical later. So this has incredible ramifications. But of course, the downside is, can traditional companies survive? Will they get AI fast enough to make it?

20:16And also, what happens to the future of work?

20:19Emily Forlini: Mm-hmm. Yeah, you're actually a very introspective person I got from the book. Like, I didn't expect that, but reading through it, I mean, you do think really deeply. And there was one section where you talked about you were getting your PhD, and you asked a question, what do we do? What's the meaning of being a human when machines are smarter than us? Yes. Do you remember that part? Yes, I did ask that question. And you kind of just got at that. We differentiate ourselves by our intelligence and also through our work. So your professor responded something like, well, why do you think that the meaning of humanity rests on us being the smartest?

20:54Right.

20:55Emily Forlini: Where do you land on that today? I'm completely with him. At the time, he confused me. Because that's something I've always thought was a metric. You can rank IQ of people. But now I see that we've just created a species that's smarter than us. So we need to be humble, but we need to also be ambitious and use it. But are we going to become dependent on it? I mean, even in your book, you're encouraging companies to be AI native. That's the title of the book. Across the organization, adopt AI in every department. Work together to implement it as much and as fast and as smart as possible. So if a company is doing that and it's smarter than them, does the future of the company rely on Anthropics Roadmap, on OpenAI's Roadmap, on the compute coming in as fast as these companies need it to?

21:49Emily Forlini: So is there a dependency issue when we have these machines that we rely on that are smarter than us? Well, because there are multiple choices, right? Closed source, open source. It's not quite like that. I would say it's more like when there's electricity. You can get electricity from batteries, solar, and other ways, an electrical grid. I think it's more about the applications that really matter. And when it comes to enterprises, as I mentioned in my book, one of the biggest problems is as smart as these models are, they really don't understand the company. Because a company is very complex. You know, you and I can use Claude and ask a thousand queries.

22:34It will know enough about us that it becomes almost a personalized agent, which is great. But we are relatively simple compared to a hundred thousand person company, right? What are the organizational charts, decision making processes, special words the company uses and how are people promoted, the workflow, the business plan? what does the CEO want? So many things that if you just can't throw it all at a general model, because there's way too much data to throw. And there's also company secret, you want to send it to them. So new technologies need to be brought in to make AI work for large companies.

23:18And that in my book, I talk about two types of things. Number one is what's called ontology and Palantir, That's a great job. Palantir, yep. And Alex Karp was openly criticizing Anthropic by saying these guys just want tokens, but they don't care about AI understanding your company. Is that ontology?

23:39Emily Forlini: Do you want to define that? Because that's a new term a lot of people don't know. Yeah. So ontology, think of it as everything that's important and true about your company. Described in natural language. think of it as an unbelievably thick employee training manual. But you're not training employee for one job. You're training an employee for every job. So when this IQ300 AI, similar to if you hire the IQ300 person comes to your company, without this manual, they're going to be lost. Their contribution will be very little because they can only answer general questions because they don't know your company.

24:18So the process of creating this ontology semi-automatically, but with human in the loop, with a great tool which Palantir has, which we have at 01.ai, I think is an important first step. The second important step is the hallucination problem.

24:35Emily Forlini: Right. So how do we get AI to learn to tell the truth, not to please us, remove the sycophancy, but more importantly, understand the depth of our question so that it can give us its reasoning chain in a way that's visually understandable. Like here's the five evidence I give you. For example, For example, if I ask a question to my company builds a product called Boss AI. So if I ask Boss AI, of all the meetings in the company, what are the biggest debates that I as CEO really need to know but don't know? So think how hard the question that is to answer. Go through all the Slack messages. Well, and meeting.

25:24And meeting.

25:24Allie Garfinkle: Well, and it's ultimately an exercise in judgment. What's consequential? Right. Right. So you're asking the AI not just to match things, but need to understand what does the CEO need to know. Right. That's not an easy question. The second thing is, what do I not know? So that's also a very challenging question. But our AI can handle it and give a credible answer. But we can't guarantee that answer is 100 percent right. But if you give the evidential chain, it will say, well, the first debate you should be aware of is this, and these are the three reasons I think you should know, and these are the three reasons I think you don't know.

26:07Now, I can say, okay, these are all correct, then must be good advice. Or I'll say, wait a minute, you missed that. I already know this. Then I would give feedback, and then I would self-learn.

26:18Emily Forlini: Right. Or I would say, yeah, what you say is true, but it's not important enough for me to know. Don't mention stuff like that in the future because of whatever reason. So AI is now learning. So AI explains itself. It's very easy interaction. So that means the AI can serve the boss. And that's the powerful thing is that all the companies are doing experiments, giving AI to customer service, legal document checking, and so on. But these are of marginal value. AI's biggest value right now today is to superpower the CEO. And that's what I've built for myself, and we're now releasing it as a product.

26:59So this is like a little angel who flies into every meeting and listens to every meeting, who reads every employee's public docs and also docs that I, as CEO, have access to, and remembers it all. So it's an infinite memory, all-knowing, omniscient, omnipresent chief of staff that whispers in my ears what I need to know and what I should consider doing and how I should make my decisions.

27:27Emily Forlini: That makes me nervous.

27:29Allie Garfinkle: I am sort of to this point, I have been sort of covering AI since a couple of months before ChatGPT came out from the investor perspective. But I've also been covering it from the enterprise perspective. And I feel like there's just been a lot of scrambling. And it feels like it's been going on for years, large enterprises, Fortune 500s, saying, we need to adopt AI, but no one actually being necessarily all that clear about what that means. What mistake do you see over and over again? Number one mistake, which is understandable, is the CEO in the traditional company who's not that technical will see AI as another piece of software, which of course it kind of is, right?

28:12ChatGP is a piece of software. Cloud Code is a piece of software. So as a piece of software, what the CEO has done in the past is number one, ask the CIO to install it. And number two, ask what apps need to be built on top of it. And number three, not use it himself or herself that much, but ask the chief of staff to say, well, find out for me what are the three debates. and then the chief of staff is going through all the paperwork. Or boss AI. Or boss AI, right.

28:44Emily Forlini: Which also makes me nervous, but I'd also want to know. Do you know what I mean?

28:47Allie Garfinkle: Well, I just want my own boss AI. You want your own boss AI.

28:50Emily Forlini: My boss has an AI. I want an AI on my boss. What does my boss care about that I need to know? Well, that's the viral effect. So our product is available for every employee. Right. But if you don't manage a large enough team, the benefit will be smaller because the biggest thing this basically drives a truck through is the information loss in a management hierarchy. When people don't tell the whole truth to the boss and then the boss is surprised by risks that emerge, problems that happen, it's too late to fix and too late to fire someone. And also when the boss wants to do something, his or her direction is pocket veto by one of the employees who nods and doesn't do things.

29:37So now this is Ray Dalio's radical transparency, except it's offered to each person within the scope of what the person can see. So your agent would be able to see every document you can see, which is according to practice.

29:52Emily Forlini: And should my agent be on the org chart? Is it my report or is boss AI or assistant? You know? So the current product - Where do agents go on the org chart? Where do they go on the org chart? Yeah, I would actually love to know. Do they go on the org chart? Currently, these types of decision support agents, which need to be highly intelligent and explainable with their answers and know the ontology of the company, we treat them as decision support. So they don't go autonomously do things because they're working on consequential things. And one of the big problems of AI and agents is they can't have accountability.

30:31So if things mess up, you can't blame the AI or unplug from the wall.

30:37Emily Forlini: But I like the idea of it going on the org chart because you do, as a manager, have some responsibility for your team. So with the agent revolution, dots, muse, instinct, the idea or the promise is that they will take proactive actions. And so somebody has to be accountable for that. And over your team, you have accountability. What is the output of my team? Who is doing what outside of myself? So could it give more accountability and improve the overall system if an agent was on the org chart? Yeah, so the agent should be on the org chart in the following way. The decision support agent should be attached to each manager or executive who uses it at the way the chief of staff is attached.

31:19Right. So that's the decision support agent. Now, some agents are autonomous or semi-autonomous because the tasks they do are not mission critical and autonomous is faster. And when it's proven to be better than humans, let's say customer support or quantitative trading, then we can leave it to run on its own. Now, those still need a manager of human who's basically setting directions, checking answers, approving decisions. And the budget. They need a budget. And the budget. So we call that person the DRI, directly responsible individual. And then I would say the agents report to the DRI in the org chart.

32:05Okay.

32:05Allie Garfinkle: Okay. Well, because it's a thing I wonder about a lot because the buck has to stop with someone and it's not the agent necessarily.

32:12Emily Forlini: Right.

32:13Allie Garfinkle: You know, one thing I would be really curious about to you is there are so many headlines we've seen, so many studies that basically say there are a single digit number of percentage points of companies who try to implement AI actually get real enterprise value out of it so far.

32:29Emily Forlini: Yeah, you're already nodding. You're like, oh, yeah, this is a problem. This is a McKinsey survey. Six percent are seeing ROI. Yeah, it's really low for all the money that's being funneled in. The ones who succeed, what is the thing they did right in your experience?

32:45Allie Garfinkle: Or is there any trend at all? Yeah, so let me first clarify. The McKinsey study was done a little while ago, and the people were using older models, which probably weren't giving that much bang for the buck. So we need to give the CEOs a little more credits for the 6 % number McKinsey gave, which is just because, you know, a year in the AI world is an eternity. So we're really looking backwards like a year ago.

33:13Emily Forlini: But McKinsey thinks this is important too. We went to a couple of us at Fortune went to an event they had. I think it was a month ago. It's very close to here, their office or this event. And they also brought up the 6%. So I think it's something they're still working through as well. I do believe it's - It's enough to make people nervous. It's enough to make you pause and you look at the numbers in the S1 and you're seeing trillions and trillions of dollars and then only 6%, you'd think it might be 96 % of that valuation. Well, it should be 96%, but I think even if it's higher than 6%, it's probably still single digits.

33:44So the McKinsey view is valid. So I think the main problem is the CEO sees it as software, delegates to the CIO, allocates budget for people to use these chatbots, gives them tokens, and also give budget to build agents on top of these on a bottom-up matter, right? The finance department wants half a million dollars to build an agent for reimbursement and customer service, legal, and the CEO approves the budget and thinks this is AI transformation. It is absolutely not. It is AI enablement. It is keeping the company exactly the same and gluing AI here and there, kind of like the early days of the internet when the company didn't change a bit except they told people, well, go on the browser, set up a website.

34:33But that's -

34:34Allie Garfinkle: We do internet here is the energy. Yeah. All right. So at the time when you were at that stage, there was also no ROI for the internet. But now AI native or internet native means if you remove the internet, the business falls apart. The same with AI. And because once internet is integrated fully, everything basically runs around it and then people take it for granted. And then it adds so much to the work. AI is all that plus more because AI has self-learning. So the more you use Boss AI, the more it knows me, the more it knows the company, the better it performs. So what really the CEO is missing, number one, is thinking of it as software, thinking of it as enablement, not thinking of it as AI native.

35:21And then the way to get to AI native is to use it by the CEO. Because finally, then the CEO gets what the AI is good for.

35:30Emily Forlini: Right, it's top down. Then the AI will, the boss will see, oh, these departments aren't functioning. Oh, these departments aren't using AI. And these departments are using AI, but there's no feedback loop to self-improve. Then the CEO can make more and more organizational decisions based on properly using AI and the fact that AI is swarming into the company. And then the organization chart needs to change because you'll need people who manage the AI and traditional companies don't have enough of them. So all of this is a whole series of AI transformation. And one last thing that is psychological is that most middle managers don't want AI native company because it's uncomfortable.

36:14It's new. They feel they're not familiar and they're afraid of change and they think it might shrink their power, salary, team size, etc. All All of the fears are real except that the middle manager has an opportunity to become a DRI. All they have to do is learn how to run these AI agents. And it's not rocket science. It's in fact not engineering.

36:37Allie Garfinkle: To the point of middle managers being afraid, I think a lot of people are. Because it gets at something that you've been very vocal about, which is questions around job loss. Kai-Fu, what industries, what kinds of jobs are in the crosshairs here? Well, I think first is the anything digital. Anything digital AI can do very quickly because the data is already there. The metrics are there. So that would include content companies. Sorry. It's okay.

37:10Emily Forlini: I'd rather you tell me now, you know. entertainment, games, video, movies, and also financial. So banks, insurance, investments, and I'm sure there's many more. So things that are digital are the first to be transformed. And the things I say in my book, the examples I give are not like there's only me who came up with them. I'm quite confident all the top AI companies are implementing this. So traditional companies must put up a defense so that they can have a chance to transform into AI native before the AI native AI companies start invading their territory.

37:55Allie Garfinkle: The other thing about it too is I do wonder a lot about how this reorganizes society. You've talked before a lot about how people put a lot of premium on their jobs. It's a piece of identity. How do you see this evolving the world around us? Well, first, I think we're going to see a continuing wave of increasing unemployment and new jobs will emerge. Old jobs will be replaced, but they're not going to happen at the same pace. The replacement will happen sooner. The new job will emerge later. Sort of like internet led to that, right? That sounds like this will happen even faster. This will happen faster.

38:39But have faith that new jobs will emerge. Like when the internet was first invented, none of us could have predicted the power of e-commerce, right? The power of social, the power of streaming video, the power of grocery deliveries, the power of Uber, right? So internet ended up creating tons of industries and jobs, but it didn't happen the first five years or even 10 years of the internet. So that will happen with AI, just have faith. But in the meantime, there's going to be a big issue for people who feel that not only is AI doing my job better, but any job I can do, AI can probably do better.

39:20So it does make sense for everyone to really think about what are the innately human strengths that will lead to jobs that don't go away and then the new jobs that is being created and then wait for the new jobs that will be created that we don't yet know.

39:39Allie Garfinkle: What jobs do you think will be created? If you put your sort of futurist hat on, what do you think will be here in five, 10 years that isn't right now? So I think this is - What job will I have, Kai-Fu? I think this is the most amazing time to be an entrepreneur, to do a one-person company. So you believe in the one man unicorn? One man. One person. One person company that may not need to be a unicorn. Okay. Because maybe it just earns a living. Maybe it's pivots or maybe it becomes a unicorn. Because then when you have one person or two or three close people, and then with AI doing most of the work, you don't have this friction.

40:19What I describe in this book about the AI transformation. You're starting from the get-go, an AI native company. And as I described, many, many existing companies are very vulnerable. And you go find places that you can attack and win, but also you find things that couldn't be done before. So I think OPC, one person company, entrepreneur is one great sector. The second one is the DRI, which is actually quite similar. is an OPC within a company structure. So find a company that supports empowering DRIs, and then the DRIs become good at running AI. So being able to be the master of AI is clearly the best outcome for any young people or not so young people looking for change.

41:11And programming AI is not as hard as people think. It's closer to writing a contract than writing code. In fact, in our latest tool in my company, there is no code. Code never comes up. You never even see the code. So people who are afraid because, wow, this is computer science. I don't code. It doesn't matter. I don't want to earn Python. You don't have to. So this is really an opportunity open to everyone. But to be fair, maybe only 5 % of the people will become DRIs of some sort. So what about the rest of the world? So I think what is still safe as a job are environments that are open and not necessarily proven with data.

41:55Because AI is so good with anything with data, you can't beat it. But if you can invent a brand new kind of dress, a new style of music, or a new way of doing catering, or Steve Jobs inventing the iPhone, Coco Chanel inventing comfortable and beautiful clothes. So there are lots of ways where going into open spaces where it's not based on data can be something people can do. That's also something to aspire to. But also there are more mundane examples. An assembly line worker who does the same thing, repeat again and again and again, that's clearly replaceable. But a plumber who is almost a detective, right, and every house is different, can have a job that should last 30 years because, you know...

42:48Emily Forlini: Until physical AI comes. I think physical AI may not easily solve that one, but I do think someday all the houses will be rebuilt so that the plumbing doesn't require the... So plumbing is an example, electrician. So there are lots of physical jobs that will be hard and also physical jobs in open spaces, right? Right. Like cleaning any room is a very hard problem because every room is different. But these are lower wage jobs. So it's, I can't really see this playing out. It would be a radically different society than today if we're all service workers. Well, some service can be high value, right?

43:30Emily Forlini: It is high value. I wish I could give all of my money to my cleaning company. I mean, I love what they do. They're amazing people. Plumbers too. I mean, if some leaking pipe, they solve it for you. It's the biggest value ever. But in reality, they don't have the highest paid jobs in society, unfortunately. But that will change because those jobs will be in higher demand. And I think a lot of service jobs will be humans. The other dimensions, I mentioned openness being one dimension. The other dimension really is empathy, compassion, social, human touch, winning trust. So the skills of persuasion, communication, teamwork, compassion, for the ultimate goal of getting trust from other people.

44:19So there's still plenty of those jobs within the company, right? People who do private wealth management, people who do high value accounts, sales, business development, partnership development, negotiating deals, M &A. So there's some high value jobs, but also there are jobs with less economic value, but very high social value so that more people can have jobs, which is important to avoid the despair. So I would say, you know, teachers, nurses, there can easily be more of them. And then also special new jobs, like, you know, a great concierge for the ultra rich who will emerge, a great tour guide, a great storyteller, and a comic, a clown at the parties, someone who does catering, someone who cleans your closet.

45:13So all these service jobs will be in demand because AI will create a lot of money. And then people with money aren't just going to want to talk to the model all day, right? So they're going to want connections with people and people who can create genuine connections that are either worth paying for or at least brings warmth to people's hearts. So the combination of being paid directly for something like that or having a universal basic income, UBI, that taxes the ultra-rich and helps out everyone, I think will more or less balance things in strong AI countries like U.S. and China.

45:55Emily Forlini: It's a very different society. Or just a more dramatic version almost of the society today. But also AI will create new problems and that can create new businesses, which we're seeing with cybersecurity. So there's more hacking instances of exploiting and finding vulnerabilities. And now there will, I think, likely be more cybersecurity jobs. And we're seeing new problems to solve. But there are different cybersecurity jobs. The traditional skills would not be sufficient. You need to train and teach AI to defend rather than the attack.

46:28Allie Garfinkle: Right. All right. I don't know how I feel exactly about the world we're moving towards, but I also feel like there is so much that is still theoretical that we just don't know. And you've devoted your life to this, Kai-Fu. So I want to ask you about what we don't know to wrap here. What is the biggest blind spot you have in AI right now? What is the question you just don't know the answer to? I don't know what RSI will lead to. It's exhilarating and scary at the same time. And I don't know how the companies will figure out how to slow down. I think some slowdown will happen because Anthropical OpenAI will move some model training people to do safety and guardrail, which is a good thing.

47:14So they'll naturally slow down a little. But the people who are in this industry are so intensely competitive and passionate about technology. So slowing down technology development other than nuclear weapons, chemical, biological weapons, we've hardly seen examples of that that's enforceable. And those weapons are single-use, people killing. so it's easier to come to an agreement. And also they're easier to track because you can look for uranium. But how are you going to know who's training a model or using a model to do bad things? And as they get super smart self-learning, we really don't know whether this is going to be Nirvana or Armageddon.

48:07Allie Garfinkle: Kaifu, thank you so much. It's a heck of a place to leave it. Kaifu, Emily, thank you. Thank you. It was great. Okay, good.