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.
Questions
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
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Karen Hao says Elon Musk and Sam Altman created OpenAI as a nonprofit largely because Musk disliked Google's dominance of AI development.
“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.”
0:11:16 · Karen Hao · OpenAI origins
Karen Hao says OpenAI used its nonprofit status to recruit talent because it could not match Google's salaries but could offer a sense of mission.
“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.”
0:12:04 · Karen Hao · Recruitment
Karen Hao says OpenAI's mission-driven pitch helped it recruit Ilya Sutskever, who became its chief scientist and its first major hire.
“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.”
0:12:04 · Karen Hao · OpenAI hiring
Karen Hao says OpenAI created a for-profit arm inside its nonprofit because its chosen path of AI development was capital intensive, and that this is the root of its governance challenges.
“They create this for-profit arm within the nonprofit. And that is the heart of all of the governance challenges that OpenAI then faces”
0:12:29 · Karen Hao · OpenAI governance
Karen Hao says Anthropic was founded by former OpenAI executives who disagreed with OpenAI on whether it was mission-driven or profit-driven.
“And so to what you were mentioning, Anthropic was founded by former ex-OpenAI executives because they disagreed on this fundamental premise.”
0:13:04 · Karen Hao · Anthropic founding
Karen Hao says OpenAI's board at the time of Sam Altman's firing leaned toward the 'doomer' view that imminent AGI could be devastating to humanity.
“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.”
0:14:43 · Karen Hao · Altman firing · 2023
Karen Hao says senior lieutenants told OpenAI's board that Sam Altman had a pattern of abusive and manipulative behavior and was untrustworthy, which contributed to the decision to oust him.
“they felt that he had a pattern of abusive behavior, of manipulative tactics, of being untrustworthy.”
0:15:18 · Karen Hao · Altman firing · 2023
Karen Hao says China lacked access to cutting-edge computer chips for several years because of US export controls.
“because for the last few years, China did not have access to cutting edge computer chips because of U.S. export controls.”
0:18:56 · Karen Hao · Export controls
Karen Hao says Chinese companies responded to chip restrictions by building more efficient models and open-sourcing them.
“that is what's leading to a blossoming of a lot of different, more efficient models that then they are open sourcing”
0:19:41 · Karen Hao · Chinese AI models
Karen Hao says US companies and academics are increasingly turning to Chinese open-source AI models instead of American closed-source ones.
“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”
0:19:41 · Karen Hao · Open-source adoption
Karen Hao says OpenAI says in company blog posts that it wants to bring 'democratic AI rails' to countries around the world, and that this is how it courts governments.
“they say in company blog posts that they want to bring democratic AI rails to countries around the world.”
0:21:23 · Karen Hao · OpenAI messaging
Karen Hao says Chilean water activists she profiled realized in 2019 that Google had proposed a data center in their community.
“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”
0:24:49 · Karen Hao · Chile data center · 2019
Karen Hao says the activists were alarmed because the proposed Google data center could use large amounts of freshwater while Chile was suffering a historic drought.
“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.”
0:24:49 · Karen Hao · Chile data center · 2019
Karen Hao says Google agreed with the Chilean community not to use fresh water to cool the proposed data center.
“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”
0:26:03 · Karen Hao · Chile data center
Karen Hao says the Chilean government created a round table to consult residents and activists on future data center projects.
“the chilean government has created this round table to consult with residents and activists for all future data center projects”
0:26:03 · Karen Hao · Chile governance
Karen Hao says reporting at the time she wrote her book indicated OpenAI was considering supercomputers powered by five gigawatts, close to New York City's average power demand.
“OpenAI was considering building supercomputers that could eventually, like within a few years, be powered by five gigawatts.”
0:28:34 · Karen Hao · Data center power
Karen Hao says Sam Altman has talked about building 250 gigawatts of data center capacity, which the transcript renders as costing '$10 1 trillion'.
“Altman is talking about building 250 gigawatts of data center capacity, which he estimates could cost $10 1 trillion.”
0:29:07 · Karen Hao · Data center capacity
Karen Hao says 250 gigawatts equals about four dozen New York Cities of power, using an average New York City demand of 5.5 gigawatts.
“So that is now four dozen New York cities of power because New York is it's 5.5 gigawatts average power demand.”
0:29:07 · Karen Hao · Data center power
Karen Hao says Sam Altman wants the 250-gigawatt buildout to happen by 2033.
“he wants that to happen by 2033.”
0:29:07 · Karen Hao · Data center capacity · 2033
Karen Hao argues that power for such data centers would largely come from fossil fuels because nuclear capacity could not be built fast enough and renewables cannot run the facilities around the clock.
“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.”
0:29:53 · Karen Hao · Energy sources
Karen Hao says scaling base models is approaching an asymptote, while companies shift to scaling inference compute.
“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.”
0:30:55 · Karen Hao · AI scaling
Karen Hao says closed-source model makers do not disclose training data, so benchmark results cannot show whether capabilities are learned or memorized.
“we literally don't know what is being put into the models. And therefore, any benchmark measuring the capabilities”
0:33:09 · Karen Hao · AI benchmarks
Karen Hao says she spent a week in Kenya interviewing data annotation workers, sometimes for an entire day.
“I spent a week in Kenya and my interviews with these workers were hours and hours and hours long, sometimes the entire day.”
0:36:31 · Karen Hao · Reporting process
Karen Hao says many OpenAI researchers spent up to 10 hours with her recounting their time at the company.
“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.”
0:37:06 · Karen Hao · Reporting process