# The AI Industry Is Building Modern Empires with Karen Hao

- **Show:** Analyse Asia with Bernard Leong
- **Published:** 9 Dec 2025
- **Source:** https://podcasters.spotify.com/pod/show/analyseasia/episodes/The-AI-Industry-Is-Building-Modern-Empires-with-Karen-Hao-e3bor6c
- **Transcript:** https://podcasts.opennash.com/shows/analyse-asia/the-ai-industry-is-building-modern-empires-with-karen-hao-82ea8b/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

Karen Hao argues that leading AI companies such as OpenAI, Anthropic and xAI behave like modern empires. They take data and labor, control how knowledge about AI is produced, and are driven partly by a quasi-religious belief that they are building an 'AI god'. She says this concentrates economic and political power in a few hands and calls for the public to hold them accountable.

The central tension is between an industry committed to ever larger scale and the evidence that scale is neither necessary nor harmless. Chinese firms, constrained by US chip export controls, found more efficient methods, while Sam Altman's proposed data center buildout would be environmentally devastating. Against that stands the Chilean water activists' example, showing that ordinary people can push back.

## Key insights

1. **AI research is largely funded by industry** Hao says most AI researchers work for or are funded by AI companies, even many university academics. She compares this to climate science bankrolled by fossil fuel firms, which leaves policymakers without independent experts to check company claims.
2. **OpenAI's hybrid structure caused its governance conflicts** Hao says OpenAI recruited some staff as a nonprofit mission and others as a for-profit business, leaving employees confused about what it was. Factions fought over building and deploying technology, and Anthropic was founded by ex-OpenAI executives who disagreed on this premise.
3. **US export controls spurred efficient Chinese AI** Cut off from cutting-edge chips, Chinese companies innovated to do more with less compute instead of simply scaling up. Their open-source models are now being adopted by US companies and academics, which Hao presents as proof that regulation can spur innovation.
4. **Altman's data center plan implies huge energy demand** Hao says Altman has talked about 250 gigawatts of capacity by 2033, roughly four dozen New York Cities of average power demand. She argues the power would mostly come from fossil fuels, adding emissions and air and water pollution.
5. **Closed-model benchmarks cannot be trusted** Because companies do not disclose training data, Hao says outsiders cannot tell whether a model learned a capability or memorized answers. She concludes benchmark results for closed-source models are effectively useless.

## Questions this episode answers

### Why does Karen Hao call AI companies 'empires'?

She says they share four traits with old empires: taking resources such as data and labor, monopolizing knowledge, pursuing a quasi-religious quest for artificial general intelligence, and growing by expanding and consolidating power. She says these companies take from the majority of the world while a few benefit.

### Why was Sam Altman fired from OpenAI, according to Hao?

Hao says the board leaned toward the 'doomer' view that AGI could be devastating, while the company acted like a 'boomer' that expects a utopia. The board was also worried by senior colleagues who described a pattern of abusive, manipulative and untrustworthy behavior by Altman.

### What did the Chilean water activists achieve against Google?

The volunteer group, which Hao names as Mosa Kat, protested a proposed Google data center in 2019 during a historic drought. Google agreed not to use fresh water for cooling, and the Chilean government set up a round table to consult residents on future data centers.

### Is AI scaling hitting a limit?

Hao says scaling the base model is approaching an asymptote, meaning gains are flattening. Companies are now scaling inference compute, which is the computing used when a model answers, and she calls the overall scaling approach scientifically unnecessary.

### Does Hao think AI will replace journalists?

She is not bullish on that idea. She says her reporting depended on empathy, such as a week in Kenya and interviews with OpenAI researchers lasting up to 10 hours, which she believes cannot be automated.

## Facts

- **Karen Hao** [0:11:16]: "Musk and Altman created OpenAI as a nonprofit specifically because Musk in particular really did not like the fact that Google was dominating AI development and Google was a for-profit."
- **Karen Hao** [0:12:04]: "also to recruit a lot of talent because they couldn't compete with Google on salaries, but they could compete with Google on a sense of mission."
- **Karen Hao** [0:12:04]: "And that is, in fact, what allowed them to recruit Ilya Satskever, who became the chief scientist of OpenAI and was like the first major important hire for the company."
- **Karen Hao** [0:12:29]: "They create this for-profit arm within the nonprofit. And that is the heart of all of the governance challenges that OpenAI then faces"
- **Karen Hao** [0:13:04]: "And so to what you were mentioning, Anthropic was founded by former ex-OpenAI executives because they disagreed on this fundamental premise."
- **Karen Hao** [0:14:43]: "the board was much more leaning towards the Doomer ideology, which part of these quasi-religious movements believes that AGI is imminent and the cataclysmic shift that arises will potentially be devastating to humanity."
- **Karen Hao** [0:15:18]: "they felt that he had a pattern of abusive behavior, of manipulative tactics, of being untrustworthy."
- **Karen Hao** [0:18:56]: "because for the last few years, China did not have access to cutting edge computer chips because of U.S. export controls."
- **Karen Hao** [0:19:41]: "that is what's leading to a blossoming of a lot of different, more efficient models that then they are open sourcing"
- **Karen Hao** [0:19:41]: "has led companies in the US as well as academics in the US to actually increasingly turn to Chinese open source AI models instead of American closed source ones"
- **Karen Hao** [0:21:23]: "they say in company blog posts that they want to bring democratic AI rails to countries around the world."
- **Karen Hao** [0:24:49]: "there's this group of Chilean water activists called Mosa Kat that I profiled in the book who, in 2019, they realized that there was this Google data center that was being proposed within their community"
- **Karen Hao** [0:24:49]: "they were completely alarmed by the degree to which this facility could take up their water resources at a time when Chile was suffering a historic drought."
- **Karen Hao** [0:26:03]: "thus far google has agreed with the community that they are not going to use fresh water for cooling the facility anymore which is a major win for that facility"
- **Karen Hao** [0:26:03]: "the chilean government has created this round table to consult with residents and activists for all future data center projects"
- **Karen Hao** [0:28:34]: "OpenAI was considering building supercomputers that could eventually, like within a few years, be powered by five gigawatts."
- **Karen Hao** [0:29:07]: "Altman is talking about building 250 gigawatts of data center capacity, which he estimates could cost $10 1 trillion."
- **Karen Hao** [0:29:07]: "So that is now four dozen New York cities of power because New York is it's 5.5 gigawatts average power demand."
- **Karen Hao** [0:29:07]: "he wants that to happen by 2033."
- **Karen Hao** [0:29:53]: "the energy that would be used to power those facilities will largely come from fossil fuels around the world since there is not enough nuclear capacity."
- **Karen Hao** [0:30:55]: "scaling the base model has seen it approach an asymptote. That's right. But the companies are now shifting to continuing to scale the inference compute."
- **Karen Hao** [0:33:09]: "we literally don't know what is being put into the models. And therefore, any benchmark measuring the capabilities"
- **Karen Hao** [0:36:31]: "I spent a week in Kenya and my interviews with these workers were hours and hours and hours long, sometimes the entire day."
- **Karen Hao** [0:37:06]: "with open AI researchers, there were many who sat with me for 10 hours to recount all of the different things that had happened during their time at the company."

## Transcript

**[0:00:00] Karen Hao:** These empires are amassing extraordinary amounts of resources by dispossessing a majority of the world. That includes the data that they're extracting from people by just scraping it from online or intellectual property that they're taking from artists and creators. Most AI researchers now work for the AI industry and or are funded in part by the AI industry. Even academics that have stayed within universities are often funded by the AI industry. and the effect that it's had on knowledge production is akin to the effect we would imagine if most climate scientists were bankrolled by the fossil fuel industry.

**[0:00:36] Karen Hao:** I cannot stress enough how much they genuinely believe that they are on the path to creating something akin to an AI god and that this is going to have cataclysmic shifts on civilization.

**[0:00:49] Bernard Leong:** Welcome to Analyse Asia, the premier podcast dedicated to dissecting the powers of business technology and media globally. Today, we explore one of the most consequential questions of our era, who owns the empire of AI? With me is Karen Hao, investigative journalist and author of her very famous book now, Empire of AI, a sweeping examination of how companies like OpenAI, Anthropic, and XAI have become modern empires, shaping the future of society, labor, geopolitics, and democracy itself. Karen, welcome to the show.

**[0:01:20] Karen Hao:** Thank you so much for having me.

**[0:01:21] Bernard Leong:** Yes, I think we crossed paths with a couple of friends and I've asked them to all recommend me. So I have to thank Grace and Jin Yang for getting you onto the show. So I actually want to understand how did your career journey begin and what led you from doing mechanical engineering at MIT into journalism and now deep diving into the coverage on artificial intelligence?

**[0:01:42] Karen Hao:** Yeah, it was very much by accident that I got into this career path. I studied engineering thinking that I would go work in the tech industry and have a very traditional trajectory in tech. But then when I started working in San Francisco at a startup and being part of Silicon Valley, I realized that the Silicon Valley model of innovation incentivizes profitable technologies, but not necessarily technologies in the public interest. That's why I went into tech. I wanted to build tech in the public interest.

**[0:02:12] Karen Hao:** And when I didn't really see a pathway forward within San Francisco, I thought maybe I could just switch to a more publicly aligned industry. And that's when I went into journalism and started covering the tech industry. And my first job just so happens that I was assigned to cover artificial intelligence.

**[0:02:34] Bernard Leong:** I think you have covered the tech industry for years. I think you've seen idealism and also its dysfunction as well. What are the biggest lessons you have taken away on reporting on AI within Silicon Valley and the culture that shapes

**[0:02:47] Karen Hao:** AI today? My biggest takeaway is that I think the initial feelings that I had working in tech have largely played out and gotten worse, that this industry has become more and more powerful over time and has wielded that power in ways that continue to distort the ecosystem of innovation towards making technologies that primarily benefit them rather than benefit a broad base of people around the world. And one of the important functions of journalism is to hold power to account.

**[0:03:27] Karen Hao:** And so whereas before, I think tech journalism was seen more as similar to science reporting, where primarily the role is for the journalist to explain a technology to a lay audience. Now I think of tech journalism as more akin to political reporting, where you are trying to hold a core power center of the world accountable to other people.

**[0:03:51] Bernard Leong:** So I think maybe we should just get straight into the main subject of the day. I want to start off with the core of your book. So why do you have came up with the propositions that companies like, say, say OpenAI and dropping X.AI can be understood as empires of AI. I think that's the core part of how this book is built upon.

**[0:04:09] Karen Hao:** Yeah, very, very related to what we were just talking about, which is that these companies have consolidated an extreme amount of not just economic power, but also political power. And the way that they're going about consolidating these powers is very similar in terms of tactics to what I call the empires of old. And there are basically four parallels that I lay out in my book, but you could broadly summarize them as these empires are amassing resources, extraordinary amounts of resources by dispossessing a majority of the world.

**[0:04:45] Karen Hao:** That includes the data that they're extracting from people by just scraping it from online or intellectual property that they're taking from artists and creators, the economic value that they're extracting from labor that rarely sees portionate economic value in return. It also involves this ideological quest that undergirds the AI industry's expansion. So if you look at what the AI industry does purely from a profit perspective, you actually wouldn't get a full picture of why it's doing what it is because the business logic is not 100 % sound.

**[0:05:22] Karen Hao:** But there is this kind of quasi-religious quest that also pushes the industry towards building larger and larger models and consuming more and more resources to do so. It's because they believe that they are ultimately trying to build an artificial general intelligence, an AI system, a theoretical AI system that would be akin to a human or what some people colloquially call an AI god. And so the fusion of capitalism and ideology and the extractivism and the expansionism, like all of these things are essentially very fundamental characteristics of empire.

**[0:05:58] Bernard Leong:** So I think the four traits you mentioned is kind of, if I rewind it correctly, is the resource extraction, the labor exploitation, the knowledge monopolies, and there is this good versus evil narrative that's going on. And at the same time, coupled with this obsession about building AGI, which of these do you think that the world still underestimates most in terms of when they see these AI companies at work then?

**[0:06:23] Karen Hao:** That's a good question. I would say the third one is probably the knowledge monopoly. Yeah, because, yeah, so to, I guess to flesh out that particular parallel more, I don't think most people realize how much AI research as a scientific discipline is no longer truly scientific. Because most AI researchers now work for the AI industry and or are funded in part by the AI industry. Even academics that have stayed within universities are often funded by the AI industry.

**[0:07:00] Karen Hao:** And the effect that that has had on knowledge production is akin to the effect we would imagine if most climate scientists were bankrolled by the fossil fuel industry. Like it just we would not get a clear picture of the climate crisis and we're not getting a clear picture of the true limitations and capabilities of AI and especially these frontier frontier models. And this is one of the things that when I speak with policymakers also is central to their inability to consult with independent experts to put together sensible policy for holding the companies accountable or even just like fact checking the company's claims.

**[0:07:42] Karen Hao:** This is one of the challenges is that most of the experts that they can consult are in some way financially tied to the industry. And so there is just this like echo chamber of self-reinforcing opinions that all lead to just people agreeing with what the AI industry says and its agenda. and yeah that is that is like a very shaky foundation upon which the industry has been allowed to continue doing what it's doing without really having many people challenge

**[0:08:17] Bernard Leong:** what it doesn't says to be told i'm i was working in the ai in the human genome project and actually when i recently looked at all the linkedin with the people i worked with including myself none of worst estate working for a tech company. So I think one of the curious question is, what's the one thing you know about the global AI industry now that very few people do other than just the part where there's the knowledge monopoly side, but also the kind of the cost that the AI has actually inflicted upon society as a whole?

**[0:08:50] Karen Hao:** Different from the previous answer. I guess I would then pick the fourth parallel that I draw, which is this ideological yeah so the good versus evil narrative yeah well the thing i think the thing that people don't really pick up on is the degree to which there are what i describe as quasi-religious movements in silicon valley like this it's it's really hard sometimes to explain the crazy cultural aspects of san francisco in particular to someone who's never been there who's never talked with these people because anything that you say to describe these people sounds crazy if you haven't been exposed to their ideology directly.

**[0:09:36] Karen Hao:** And so, yeah, it's just like the way that these people talk, like they, I cannot stress enough how much they genuinely believe that they are on the path to creating something akin to an AI god and that this is going to have cataclysmic shifts on civilization. And when they, and they've developed an entire vocabulary around this. Like when you listen to them talk, if you have not been exposed to their fanaticism in the past, you wouldn't even understand what they're saying because their vocabulary has become so specialized.

**[0:10:07] Karen Hao:** And like...

**[0:10:08] Bernard Leong:** This is actually due to a lot because of science fiction. But also science fiction also gives you the parallel. Like for example, Foundation, they end up destroying all the robots. And then like in Dune, the same thing happened. They destroyed all the AI as well.

**[0:10:21] Karen Hao:** It is. there is definitely a science fiction element to it, which is that a lot of the people who work in the AI industry and are big fanatics of creating AI were huge science fiction buffs growing up. And there is this kind of tantalizing premise of getting to be part of now the quest to

**[0:10:42] Bernard Leong:** turning these science fiction dreams into reality. So the one thing I did pick up from the book and And also looking at all the different AI books I've read is that the founding of OpenAI, the origin story is actually quite interesting. It was particularly the Sam Altman and Elon Musk dynamics and then that unusual nonprofit for profit structure. How did it create the kind of governance problems that we see today? And also all the offshoots, like, for example, the Anthropic team also came from OpenAI and it broke out and started our own team.

**[0:11:12] Bernard Leong:** And now we have SSI and all the different other companies that evolved from that.

**[0:11:16] Karen Hao:** Yeah. So at the time, Musk and Altman created OpenAI as a nonprofit specifically because Musk in particular really did not like the fact that Google was dominating AI development and Google was a for-profit. And so they thought, let's create something that's the antithesis to Google. that said like in in like after reporting out the whole you know first 10 years inside story of open ai in hindsight i my guess is that there was not high conviction from altman in keeping open ai a non-profit even back then like i think he probably saw there to being like high utility to starting it as a non-profit to accumulate some goodwill create a good positive narrative around the org and

**[0:12:04] Karen Hao:** also to recruit a lot of talent because they couldn't compete with Google on salaries, but they could compete with Google on a sense of mission. And that is, in fact, what allowed them to recruit Ilya Satskever, who became the chief scientist of OpenAI and was like the first major important hire for the company. But of course, then OpenAI, within a very short amount of time, decided they wanted to pursue a path of AI development that was very capital intensive.

**[0:12:29] Karen Hao:** They couldn't retain a pure nonprofit. They create this for-profit arm within the nonprofit. And that is the heart of all of the governance challenges that OpenAI then faces because they recruited some people on the basis that OpenAI is a nonprofit and they recruited other people on the basis that OpenAI is a for-profit. And so it's one of the organizations, the only organization I've ever covered where there's a fundamental confusion among employees about what even the organization actually is just in terms of like, is it actually mission driven or is it profit driven?

**[0:13:04] Karen Hao:** And because of that clash, there are different factions within the company that believed in one or the other and then would have these really big fights over whether a decision about like, should we build this technology or that technology? Should we deploy it faster or slower? Were tied up in their philosophical differences about what they believed OpenAI was. And so to what you were mentioning, Anthropic was founded by former ex-OpenAI executives because they disagreed on this fundamental premise.

**[0:13:40] Bernard Leong:** Yeah. So I guess this was one of the aspects. And then two years back, we had that firing of Sam Altman, which was also quite dramatic in terms of tech history. I think from how you observe it and looking backwards now two years later, what do you think was the board's decision then? And does it actually, on hindsight, if everything had played out then today, does it actually revealed how the governance of AI has actually now accelerated more as compared to it was within control at that point in time?

**[0:14:09] Bernard Leong:** Or maybe all along, this whole AI train is just going to go in the trajectory it should have been?

**[0:14:15] Karen Hao:** I don't think the board make up at that time was in any way effectively holding the company accountable. I mean, clearly it did not because when it tried to hold the company accountable, it failed. And I think part of it is because inherently, even that board was acting in a very anti-democratic way, which is they weren't, you know, broadly consulting lots of different stakeholders about the direction of the company. They were deciding amongst themselves what they should, what the direction of the company should be.

**[0:14:43] Karen Hao:** Just in the same way that Opening Eye's current executives are deciding amongst themselves without actually consulting with the public about what the direction of Opening Eye should be. It's just like the direction for each group. Each group has a different idea of what the direction should be. But yeah, basically what happened was that the board was much more leaning towards the Doomer ideology, which part of these quasi-religious movements believes that AGI is imminent and the cataclysmic shift that arises will potentially be devastating to humanity.

**[0:15:18] Karen Hao:** And opening eyes actions at the time were more boomer leading, which is this belief that the AGI, the cataclysmic shift that comes will be extremely utopic. And so they clashed. The board was deeply concerned about the direction that Altman was taking the organization. And they were also deeply concerned about feedback that they were getting from Altman himself or about Altman himself from his senior lieutenants who were flagging that he was they felt that he had a pattern of abusive behavior, of manipulative tactics, of being untrustworthy.

**[0:15:53] Karen Hao:** and that is what led them to decide that they should oust him. But I don't think we're like, you could say that OpenAI has since then accelerated more because the board is no longer philosophically opposed to the CEO and everyone, executives and the board are now pointing kind of in the direction of Boomer, just pedal to the metal, accelerate the expansion of this technology as quickly as possible. But in terms of like the governance of the org, I don't think we were in a better state before than we are now.

**[0:16:28] Karen Hao:** Like all of it was not a good, robust setup for governing one of the most consequential technologies of this era.

**[0:16:36] Bernard Leong:** And I think now it's a lot more reviewed with the recent deposition from Ilya during the X.AI, OpenAI trial. And then you also have some information from the information which recently talked about actually almost Anthropic. hey, Daryl almost came back to take it back over and all that. So had that alternative history. We will be pretty interested to see what the alternate universe look like. But let me just ask this question. And given that all the founders now left OpenAI or even part of the founding team of OpenAI have now created their own labs to pursue their own idealized versions of what AI should be, are we witnessing more fragmentation or more consolidation of AI power amongst all these

**[0:17:18] Bernard Leong:** players?

**[0:17:19] Karen Hao:** We're definitely seeing more fragmentation in terms of the number of competitors that there are. And it's not a coincidence that every single billionaire has an AI startup. I think this is evidence number one for realizing that the AI race is driven in large part by ego and each billionaire wants to make AI in their own image and has a story that they tell themselves and the public about why their version of AI is the morally superior version as compared to all of the other billionaires' versions of AI.

**[0:17:54] Karen Hao:** But in terms of, I said fragmented, but also I think it would be a mistake for people to view this as somehow a distributed, what's the right word? That somehow power is now being distributed more evenly across the broader public. Because of course, it's still all elites that are consolidating their approach.

**[0:18:17] Bernard Leong:** still actually is now maybe concentrated between two countries, which I didn't mention China and the US. But there is this point of view that, you know, the US policymakers and even the all the American AI builders, they use China as the rhetorical shield to resist regulation, right? I think maybe just, I don't know how much your understanding of the China ecosystem is. How do you compare contrast between say the US and the China AI ecosystems or from your perspective, just seeing how things played out publicly so far?

**[0:18:53] Bernard Leong:** Yeah, I'm definitely not an expert on the China

**[0:18:56] Karen Hao:** AI ecosystem. And I would direct people once again to Grace and Jing. They're my go-to sources as well. But I think broadly just from reading some of their reporting and watching the movements in the industry. Like, I would say that the biggest difference that has happened this year is that because for the last few years, China did not have access to cutting edge computer chips because of U.S. export controls. Many Chinese companies ended up just taking a different path from Silicon Valley, where they are innovating around the, instead of trying to just scale, scale, scale, and build larger models based on the amount of computational resources they have innovated around and found ways to do it much more

**[0:19:41] Karen Hao:** efficiently. And that is what's leading to a blossoming of a lot of different, more efficient models that then they are open sourcing and has led companies in the US as well as academics in the US to actually increasingly turn to Chinese open source AI models instead of American closed source ones, which has been one of the most

**[0:20:04] Bernard Leong:** fascinating trends that's happened. It's quite interesting to see the Airbnb CEO talking about Quinn as the model that they use, and he's not using even the open AI open models, right? Isn't that ironic from that point of view? Yeah, it's super ironic. And it's like,

**[0:20:19] Karen Hao:** I think there's this like really big narrative in the US and Europe right now. I don't know if this is also true in Asia, that regulation is antithetical to innovation. And yet, like this case study shows that with the right regulation, you can spur enormous amounts of innovation. Because when the U.S. placed export controls on these Chinese companies, like that was a form of regulation that hindered their ability to take the same path as Silicon Valley.

**[0:20:50] Karen Hao:** And then within that pressure cooker, they figured out a way around it. So yeah, it's just like incredibly really funny, you could say, that things panned out this way.

**[0:21:03] Bernard Leong:** Yeah, I find it ironic because the most controlled environment is open sourcing it and the most democratic environment is trying to control it. So it seems like they're doing what each other didn't want each other to do. But so for global audience, what are the biggest misconceptions, say, for governments have when they think about how AI is developed, deployed and monetized?

**[0:21:23] Karen Hao:** One of them is that, I don't know if this answers your question, but a common misconception that I've heard is people believing that American companies somehow have American values. So exactly what you just said. They think that Silicon Valley, because it is in the U.S., which is on paper a democracy, that somehow these companies are also ultimately trying to build democratic AI. And OpenAI perpetuates this narrative by, you know, they say in company blog posts that they want to bring democratic AI rails to countries around the world.

**[0:22:02] Karen Hao:** And this is how they're courting a lot of governments. And what I'm trying to help people realize in part through the book is that actually Silicon Valley should not be thought of as part of American democracy. democracy like it is its own power center with its own political agenda that is not in line with the american political agenda most of the time and it's actually like an inherently very anti-democratic place which is why you see them really trying to exert more and more control and develop closed source models rather than the open source brow and i think if policymakers makers understood that more that they are ultimately interfacing with companies that have their own ideological bent, that would lead

**[0:22:52] Karen Hao:** to more sound understanding of who are they actually interacting with.

**[0:22:59] Bernard Leong:** I thought one of the most striking parts of your book, which I read it in, because I teach a class in generative AI in the university here in Singapore, is the part about the real AI supply chain. We talk about the annotation workers, the content moderators or communities that such a hosting data centers. I do tell one of those stories that, you know, the AI that you are using shouldn't be taken for granted because there's a lot of people who put blood, sweat and tears involved in it.

**[0:23:27] Bernard Leong:** What does that actually teach you more about how, what are the things that Silicon Valley actually, the AI companies consistently overlook? I don't know if it's

**[0:23:37] Karen Hao:** overlook more than just like deprioritize or do not care about or are negligent of, which is just the fact that they're perfectly comfortable with the degree to which their technologies are built on labor exploitation and environmental extraction. You know, there's not any confusion among employees about the fact that this happens. And it's more just that they justify it because they use a utilitarian, the ends justify the means kind of logic, where they think that if they're ultimately on the path to AGI, then it doesn't really matter what harms are inflicted along the way.

**[0:24:24] Karen Hao:** And that is the part of the ruthlessness of the empire.

**[0:24:32] Bernard Leong:** And then, of course, the story of the Chilean water activists, right, that resisted a data center. So I think what are the things that the global audience or maybe other parts of the world that can learn from this experience as well, you know, from the book? You covered that. I thought it was quite interesting story to talk about.

**[0:24:49] Karen Hao:** Yeah, so there's this group of Chilean water activists called Mosa Kat that I profiled in the book who, in 2019, they realized that there was this Google data center that was being proposed within their community that could use freshwater resources to cool the facility. And when they started looking into the numbers, they were completely alarmed by the degree to which this facility could take up their water resources at a time when Chile was suffering a historic drought.

**[0:25:15] Karen Hao:** And so they just decided not to accept the project. And they essentially protested in every form possible. They protested in the streets by canvassing their neighbors. They protested to local politicians and really applied a lot of pressure to local politicians to do something to stop the project. And they were so effective at raising the alarm around this particular facility that it escalated all the way to Google Mountain View, which then sent representatives to speak with the communities and then escalated all the way to the Chilean national government, where the government started thinking about how do we actually make sure that other data center facilities that come into our country do not end up engaging in this kind of extraction.

**[0:26:03] Karen Hao:** and they the outcome which is sort of still ongoing is thus far google has agreed with the community that they are not going to use fresh water for cooling the facility anymore which is a major win for that facility and also the government the chilean government has created this round table to consult with residents and activists for all future data center projects to establish a more democratic governance structure around these projects. And to me, this was such an inspiring story of how to actually hold these companies accountable.

**[0:26:39] Karen Hao:** Like you could say that these Chilean water activists have the least amount of power in the global world order. You know, they are in a developing country and in an impoverished neighborhood within that developing country. They all work on a volunteer basis, like they don't have any funding. And yet they're going up against one of the most powerful companies in the world from the most powerful nation in the world. And they were still able to significantly establish guardrails around that company's activities, as well as future activities to come.

**[0:27:16] Karen Hao:** And so that is the kind of spirit that I think everyone, no matter where you sit in society, needs to have to hold these companies and technology development accountable to the broader global public is everyone should feel that they have agency and voice to push back when these companies do things, whether in their local community or their country or the broader global community that they don't actually agree with. And if we all did what the Chilean water activists did, I think we would get to a much, much better place with technology development where we see the innovation ecosystem bend more towards the public interest.

**[0:27:56] Bernard Leong:** I think there's also similar cases that just happened in the last couple of months in the US. I think there was a Microsoft and a Google data center project that was got blocked by the local communities. I think we all know, even as an AI person myself, I know that the environmental demands on large-scale AI, specifically energy, water and compute, are actually really part of that conversation. We talk about it, but people don't realize how important it is.

**[0:28:21] Bernard Leong:** I think maybe my question for you is how severe are these externalities? And also what kind of trajectory are we now on in terms of the natural resource side that's inflated, of course, by where the AI is going?

**[0:28:34] Karen Hao:** Yeah, well, Sam Altman, like when I wrote the book, the numbers that I have in my book were based on reporting that had come out at that time that OpenAI was considering building supercomputers that could eventually, like within a few years, be powered by five gigawatts. And at that time, like five gigawatts is almost equivalent to the power demand, average power demand of New York City. And so they were talking about this one facility that that could take the power demand of this huge metropolis.

**[0:29:07] Karen Hao:** And that in and of itself was already shocking. And now fast forward a year and Altman is talking about building 250 gigawatts of data center capacity, which he estimates could cost $10 1 trillion. So that is now four dozen New York cities of power because New York is it's 5.5 gigawatts average power demand. And so like that is just the like if we and he wants that to happen by 2033. So if we actually follow through with what the industry wants and that's just one CEO, one company, and then think about like all of the other companies that would then try to compete to build similar degrees of infrastructure.

**[0:29:53] Karen Hao:** that would be incredibly environmentally devastating because the energy that would be used to power those facilities will largely come from fossil fuels around the world since there is not enough nuclear capacity. It's not we wouldn't be able to build the nuclear capacity fast enough. And renewable energies are just not sufficient for running these facilities 24-7. And that would be huge amounts of emissions that go into the atmosphere, as well as huge amounts of pollutants that go into people's air and water and lead to an exacerbation of existing public health crises.

**[0:30:30] Karen Hao:** So that, yeah, it is something that like we cannot allow, we cannot afford to allow to pass.

**[0:30:38] Bernard Leong:** But then also, do you feel that now in terms of the frontier model research, I think it's also slowly kind of reaching some kind of asymptotic ceiling and maybe shifting towards, you know, products. So what does that really mean for the AI development in the next five years then?

**[0:30:55] Karen Hao:** So what we've seen, the trend is specifically scaling the base model has seen it approach an asymptote. That's right. But the companies are now shifting to continuing to scale the inference compute. And that hasn't approached an asymptote yet in terms of how Silicon Valley wants to approach things. And just to highlight again, even though we've already talked about this, this scaling is actually completely scientifically and technically unnecessary because that's why we've seen Chinese companies be able to do the same exact thing with significantly less compute resources.

**[0:31:34] Karen Hao:** But from Silicon Valley's perspective, like this is what they know. This is how they've been getting gains over the last few years. It is the most intellectually lazy yet most time guaranteed way of achieving the kinds of gains that they hope to achieve, at least so it seems for now. And so they're just going to continue pursuing that. And that's why, even though we've seen scaling-based models reach an asunto, Altman is still talking about building $10 1 trillion of data centers.

**[0:32:07] Karen Hao:** It's because the overall compute scaling has not reached its end from Silicon Valley's perspective.

**[0:32:16] Bernard Leong:** Yeah, but then there's also the problem with evaluation, right? Not all the AI evaluation is actually, because we don't know what's the training data. whether the open source or closed source model, they don't review their training data. I think the question now, one of the questions even I have myself is, what should policymakers or even public understand the limits of even the current AI benchmarks? Is there a benchmark actually?

**[0:32:38] Karen Hao:** I think what they should understand is that benchmarks are extremely faulty, in part because even determining what benchmark you should use to measure an AI model is more art than science. Like, why do we decide to measure some models capabilities on the LSAT versus the AP biology exam versus like a game versus like complex logic puzzles, you know?

**[0:33:04] Bernard Leong:** Oh, humanities last exam, which is what they are now based their benchmarks on.

**[0:33:09] Karen Hao:** Yeah. And that's why like benchmarks are constantly shifting again and again and again in ways that are not exactly helping improve public clarity. But the other aspect, which you were getting at, is that we literally don't know what is being put into the models. And therefore, any benchmark measuring the capabilities, there's drilling that just started. So any of the benchmarks, the results cannot be relied upon. because literally the essence of the science of deep learning is that you need to know what goes in to know whether the capabilities on the way out are actually learned capabilities or just memorization and regurgitation of the training data.

**[0:33:58] Karen Hao:** And so all of the benchmarks of closed-source models are effectively useless.

**[0:34:04] Bernard Leong:** So if I were to ask you, let's say you've looked from the other side, right, and provide an argument that is in favor of AI development, not for the sake of just trying to still man them, what would that argument be then? If there's a path for AI advancement, what would be your argument for AI advancement then?

**[0:34:23] Karen Hao:** For scaling?

**[0:34:24] Bernard Leong:** Yeah, for scaling.

**[0:34:25] Karen Hao:** The argument, I think their best argument is that this is an approach that has worked, that is the most predictable because it follows scaling laws. And when you are able to advance AI based on certain types of observable measures, then you can know roughly in advance how capabilities might shift and change. And that in and of itself is helpful to figure out what outcomes societal or economic or whatever might come of the next generation of models.

**[0:35:03] Karen Hao:** and the other one being that if like if AGI is going to arrive at the end of this path that it's all going to be worth it yeah so it's I think there's sort of like multiple different arguments and tacks that they usually take to justify why they're doing this thing so what is the one

**[0:35:22] Bernard Leong:** question you wish more people would ask you about your book empire of AI but they don't I don't know

**[0:35:27] Karen Hao:** I feel like most of the questions that like I've been surprised the diversity of questions that I've been asked about the book so I feel like I've often covered almost like pretty much every facet of the book but one of the things that I don't cover as often is the journalistic process that

**[0:35:43] Bernard Leong:** went into producing the book so if I ask you what's the journalistic reducing the book

**[0:35:48] Karen Hao:** can you talk about it now yeah yeah which is which is like and I think this is sort of part of also why I'm like not very bullish on the idea that AI is ultimately going to take away a ton of jobs that and produce somehow produce the same quality of those jobs especially when it comes to information processing because some people say like oh well journalism is ultimately information processing so like why are we not you know automating away yeah just just automating away journalism and and I think the core thing about my reporting process was like the scent the number one tool for me to connect to like ultimately get all of these details and find all these

**[0:36:31] Karen Hao:** stories was empathy. Like it required me traveling to a lot of different places and sitting with people for long amounts of time to slowly like to give them a listening ear and to slowly tease out all of the different experiences, lived experiences that they had as part of grappling with the AI industry. And I don't think that that can be underestimated. Like some of the Kenyan workers, for example, like I spent a week in Kenya and my interviews with these workers were hours and hours and hours long, sometimes the entire day.

**[0:37:06] Karen Hao:** And with open AI researchers, there were many who sat with me for 10 hours to recount all of the different things that had happened during their time at the company. And it was ultimately because I took the time to ask them to build a connection with them and then to listen and just keep listening and listening and listening and give them that sense that they finally were able to talk about something that was on their mind and that they had lived through.

**[0:37:34] Karen Hao:** And so, yeah, I hope that ultimately the book is not just the content is not just a celebration of like the importance of centering people and centering humanity, but also like the process was as well.

**[0:37:50] Bernard Leong:** I think of the book as a reflection of where the acceleration goes and some of the human costs that it bears. And what do we need to do to address the human costs? That's my view after reading the book on that. Even though I'm in the AI space myself as an entrepreneur and even teaching it, but I also do bear those costs in mind that to know that in order for us to reach this, these are the costs that we have to bear.

**[0:38:13] Bernard Leong:** But maybe my last and traditional closing question, what does success mean for you with your book, Empire of AI?

**[0:38:20] Karen Hao:** Sparking that conversation that you just mentioned, like helping people realize that there is another side to the equation that is often not talked about. And hopefully once we have all of the costs and benefits, a more nuanced picture of them, then we can really map out how do we make sure that AI can broadly benefit the global public rather than just a tiny few.

**[0:38:47] Bernard Leong:** That's a good place to end. Karen, many thanks for coming on for the conversation. I actually truly enjoy your book because it also told me the entire weekend to think through some of the things and even brought even to some of the things I teach in the university so that people understand that all these advances that brought about AI, just because you summarize and generate a picture doesn't mean it doesn't come with actual human costs. So in closing, I have two questions for you.

**[0:39:11] Bernard Leong:** Any recommendations that have inspired you recently? I really love the book Hope in the Dark,

**[0:39:16] Karen Hao:** by Rebecca Solnit, which is most people would on the surface think that it's not related to my book, but it really is about how to continue having human agency in a world that often feels like it's trying to extract that agency away and how to build progressive movements that lead to social progress, not just technological progress. And I really enjoy reading Brian Merchant's Substack, Blood in the Machine, as well as this newsletter from Latin America that's in Spanish called La Precisa.

**[0:39:49] Karen Hao:** That is like a Latin American perspective on AI development.

**[0:39:54] Bernard Leong:** So how can my audience find you and your book and your ongoing reporting there?

**[0:39:58] Karen Hao:** I'm on Twitter, Blue Sky, LinkedIn. That's where I post all of my work. And I have a website, KarenDHau.com or EmpireofAI.com.

**[0:40:07] Bernard Leong:** And you can definitely find us in any of the podcast distribution channels. And of course, are subscribed to us at Analyze Asia. And Karen, many thanks for having this conversation. And I'm truly glad to actually be able to talk to the author of the book that I've enjoyed reading. Thank you very much.

**[0:40:22] Karen Hao:** Thank you so much, Bernard.
