Book Review: Empire of AI by Karen Hao

This week's review is on Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI by Karen Hao.

Widely anticipated and already a bestseller Empire of AI seeks to take the reader into OpenAI - its history - and examine the growth of AI and its companies within the question of it seeking to "fortify the empire, or...wrest us back toward democracy" (418). These empires are based on a consolidation of power in terms of "knowledge, resources, and influence" (ibid). In connecting these two accounts - the growth of OpenAI from Musk-funded alternative to DeepSeek to company awkwardly caught between capitalism and the promise of open-source generative AI - Hao struggles to create a coherent book but presents an insider and outsider duality within the AI world.

I like the cover!

The main thread of the text follows the history of AI and sets up Sam Altman as a clear boogeyman - a sociopathic and manipulative narcissist using Generative AI to attain greatness. This is clearly framed in the discussion of his ascent to Open AI and the infamous weekend in which he was fired and returned stronger than ever. The back office insights into the conflicts which permeate OpenAI amongst its camps - Research, Applications, and Safety - add to this dynamic illustrating the rocket ship that is OpenAI and the conflicts caused by this growth in the industry.

Hao sprinkles in asides to the world outside OpenAI and while tangentially connected to the company - they often lack direct linkage to the business story. These are usually framed in Marxist terms of corporate power and illustrate the effects of AI on the "global south" - highlighting data workers in South America and Africa; environmental degradation due to water and power consumption; and content moderation issues among workers. There's even a brief side trip into the rise and fall of Effective Altruism and SBF's FTX.

In trying to cover both sides of the coin, it feels like a book caught between causes - is OpenAI a unique threat to the world due to Altman's evil genius? Or has the AI industry as a whole been lulled into a race which will crush the poor and environment in its quest to reach General Intelligence? Hao even drops hints at the limits of AI and calls into doubt whether General Intelligence can ever be achieved given the compute-driven returns so far. Amidst these critiques, her narrative meanders somewhat and seems lost. The sheer number of issues seems to overwhelm the narrative aspects and points to a lack of concrete thesis for the book.

Hao characterizes her book as based on the "only metaphor that encapsulates the future of what these AI power players are: empires...The empires of AI are not engaged in the same overt violence and brutality that marked this history [of colonial empires]. But they, too, seize and extract precious resources to feed their vision of artificial intelligence: the work of artists and writers; the data of countless individuals posting about their experiences and observations online; the land, energy, and water required to house and run massive data centers and supercomputers. So too do the new empires exploit the labor of people globally to clean, tabulate, and prepare that data for spinning into lucrative AI technologies." (16-17)

Her history of OpenAI starts with concerns over alignment and discussions over a "Manhattan Project for AI." The big-dick contest between Musk, Sir Demis Hassabis, and Sam Altman permeates the book and it isn't clear if the race for Generative AI is over who controls it or whether one of these men do not get to control it. But outside of Altman and maybe Ilya Sutskever and Greg Brockman, the other major figures drift in and out of the narrative. We never get a sense of their true motivations or even where they end up as they spin out of OpenAI and Altman's narrative. It is Altman whose "media savvy and dealmaking, erst on his remarkable ability to tell a good story" (33) and it seems that that charisma captured Hao in part.

She highlights the impetus to be first as key to understanding OpenAI's evolution:

Only later would I realize the full implications of this assertion. It was this fundamental assumption - the need to be first or perish - that set in motion all of OpenAI's actions and their far-reaching consequences. It put a ticking clock on each of OpenAI's research advancements, based not on the timescale of careful deliberation but on the relentless pace required to cross the finish line before anyone else. It justified OpenAI's consumption of an unfathomable amount of resources: both compute, regardless of its impact on the environment; and data, the amassing of which couldn't be slowed by getting consent or abiding by regulations. 84

Hao connects this drive to her antihero, Sam Altman: "It was specifically OpenAI, with its billionaire origins, unique ideological bent, and Altman's singular drive, network, and fundraising talent [emphasis mine], that created a ripe combination for its particular vision to emerge and take over" (132) and "It [Annie's story of Altman's abuse] helped me solidify my understanding of how much OpenAI is a reflection and extension of the man who runs it" (328). This conception of empire and extraction and the visionary at the top of it waxes and wanes throughout the narrative.

The story clearly begins to get lost when Hao's access at OpenAI was more limited and before the "excitement" of Altman's ouster and restatement come back in and it is here that the critiques of AI and its role in the modern global economic order begin to take center stage. Hao never deeply discusses regulation as a solution but highlights the Frontier Model Forum, Biden's presidential order, and Altman's currying of politicians. This largely because she understands the entrenchment theory of regulation and the corruptive approach of lobbied and donated politicians in the modern political system.

Some serious critiques are hinted at or ignored because of liberal pieties (pornography cannot be openly critiqued): "The share of pornographic images on the internet was so large that removing them shrank the training dataset enough to notably degrade the model's performance. In particular, it made the model worse at generating faces of women and people of cover due to...a significant share of the online content depicting both groups is sexually explicit. For the same reasons, the researchers left in some other kinds of disturbing images" (238).

Two other major critiques are hinted at but not investigated as thoroughly as the data workers of Kenya and Venezuela. First, Hao highlights the idea that "we now have machines that can mindlessly generate words, but we haven't learned to stop imagining a minder behind them" (254). This hint at the philosophical* implications of AI drifts amidst other references. Likewise, the use of data from children can be safely ignored: "Khan was skeptical but agreed to so in exchange for his platform getting access to the model" (246) highlights the rush to allow access to children in this experimental educational endeavor.

Most of the critique seems aimed at OpenAI in general: "Six years after the m initial skepticism about Open AI's altruism, I've come to firmly believe that Open AI's mission - to ensure AGI benefits all of humanity - has since become a uniquely potent formula for consolidating resources and constructing and empire-esque power structure" (400). It rests on being mission-driven, centralizing resources in an arms-race manner, and allowing iterations of the mission through vagueness.

Like Atlas of AI, it is difficult to tell if Empire is a business book blended with a general anti-capitalist creed or is a serious critique of AI itself, like Snake Oil succeeds at being. Even in presenting stories of the Global South there's a disconnect - her one highlight of positive AI usage is in the use of AI to save a Maori language. Ironically, the Maori are themselves notorious for their empire and eradication of native peoples and resources as part of the Polynesian expansion. Whether artisan AI is actually doable is not seriously considered. In addition to the Marxist conceptions of AI, there is of course references to Surveillance Capitalism by Shoshana Zuboff, anti-black bias in machine learning, struggles with sexist/racist engineers (always lacking specific details), and more to highlight contemporary liberal political philosophies (even a multi-page (270-273) attack on Milton Friedman and the influence of the Chicago economic policies on Chile!).

Hao claims that she sees a future of AI where "Models can be small and task specific, their training data contained and knowable...community driven, consensual, respectful of local context and history...uplifting...inclusive and democratic"(413-414). Outside of her three-page overview of Maori efforts to train an AI on their disappearing language and efforts to unionize and support workers, this vision is not presented. Perhaps Hao's Artisan AI best serves as a counterpart to Altman's own dreams of Generative AI?

A narrative which dealt further into how these tech geniuses confront the difficulties of AI - perhaps following Amodei's Claude efforts and failures, diving deeper into Sutskever's philosophical journey, Radford's jumps in training AI, or presenting genuinely different accounts of AI usage would succeed at uniting these two aspects. But maybe that's only a narrative a Sam Altman could weave.

Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI by Karen Hao.

Rating: 3/5 Stars

Good For: Those looking to learn more about Sam Altman, Open AI, and an intro level conception of the critiques of modern AI capitalism.

Best nugget: Well-written history and explanation of key AI concepts like symbolic learning, p(doom), and more.

*Part of this might be the limited nature of engineer's understanding philosophy, from a former employee: "OpenAI likes to discuss first principles, but only with the people that believe in OpenAI" (385).