Empire of AI - Book Summary
Dreams and Nightmares in Sam Altman's OpenAI
Book by Karen Hao
Summary
Empire of AI exposes how OpenAI transformed from a nonprofit promising to benefit humanity into a $157 billion empire that exploits global labor, extracts resources from vulnerable communities, and concentrates unprecedented power in the hands of Silicon Valley elites, all while claiming to build artificial general intelligence for the greater good.
The Firing That Shocked Silicon Valley
On November 17, 2023, Sam Altman was abruptly fired as CEO of OpenAI by his own board via a brief Google Meet. The company was at its peak - ChatGPT was the fastest-growing consumer app in history, OpenAI was valued at $90 billion, and Altman had achieved global celebrity status. Yet four board members concluded he was 'not consistently candid' in his communications.
The firing sent shockwaves through the company and industry. Over 700 of OpenAI's 770 employees threatened to quit and join Microsoft unless Altman was reinstated. After five chaotic days of negotiations, Microsoft threatening to cut compute access, and massive employee revolt, Altman returned as CEO with a reshuffled board.
Section: 1, Chapter: 1
The Boy Who Would Be King
Sam Altman's rise began early. Born in 1985 to wealthy parents in St. Louis, he learned to fix VCRs at age two and program computers by eight. At Stanford, he dropped out at nineteen to found Loopt, a location-tracking social network that ultimately sold for what investors had put in.
The pattern was set: Altman had remarkable ability to tell compelling stories about the future, raise enormous sums, and emerge from setbacks with enhanced power. Paul Graham famously said 'You could parachute him into an island full of cannibals and come back in 5 years and he'd be the king.' Even at Loopt, senior leaders twice approached the board urging them to fire Altman for self-serving behavior and distorting truth. The board sided with Altman.
Section: 1, Chapter: 1
The False Prophet Of Progress
'I believe the future is going to be so bright that no one can do it justice by trying to write about it now. Although it will happen incrementally, astounding triumphs—fixing the climate, establishing a space colony, and the discovery of all of physics—will eventually become commonplace.'
- Sam Altman, promising utopia while his company's infrastructure drains water from drought-stricken communities
Section: 1, Chapter: 1
The Mentors Who Shaped An Empire
Altman's worldview was forged by two powerful mentors: Paul Graham and Peter Thiel. From Graham, he learned that 'growth fixes all problems' and the importance of network effects. From Thiel, he absorbed monopoly strategy - aim for 10x better technology and be the 'last breakthrough' so no one can catch up.
Both mentors impressed the imperative for scale and the superiority of capitalism over government. As Altman wrote: 'Either you're growing, or you're slowly dying.' This philosophy would drive OpenAI's relentless scaling and aggressive competitive tactics, treating cooperation as weakness and winner-takes-all as natural law.
Section: 1, Chapter: 1
The Empire's Extraction Machine
OpenAI exemplifies modern digital colonialism - seizing data, exploiting labor, and extracting resources globally while concentrating benefits in Silicon Valley. Like historical empires, it projects narratives of civilization and progress to justify dispossession.
The company trained on millions of books without permission, transcribed YouTube videos against terms of service, and hired Kenyan workers for $2/hour to clean up toxic outputs. Meanwhile, its infrastructure demands drove water extraction from drought-stricken communities. The pattern repeats: Indigenous communities in Chile lose water to data centers while tech executives promise AI will solve climate change.
Section: 1, Chapter: 4
The Mythology Of Intelligence
AI's promise rests on the anthropomorphic illusion that neural networks 'learn,' 'understand,' and 'create' like humans. In reality, they are statistical pattern-matching systems that identify correlations in training data. The industry exploits this confusion, arguing AI training is like human 'inspiration' to avoid copyright liability.
This mythology enables dangerous misplaced trust. Lawyers have been sanctioned for citing ChatGPT's fabricated case law. The model described a brain mass as 'brain does not seem to be damaged.' The term 'hallucinations' misleadingly suggests errors are aberrations when they're actually features of how these systems work.
Section: 1, Chapter: 4
The Myth Of AI Progress
The term 'artificial intelligence' was invented in 1956 as a marketing tool by John McCarthy, who originally called it 'automata studies' until realizing he needed something more evocative. This rebranding was the original sin of the field - the promise of 'intelligence' embedded in the name created unrealistic expectations and anthropomorphizing that persist today.
There is no scientific consensus on what intelligence actually is. Throughout AI history, every time a benchmark is achieved - chess, Go, passing exams - the goalposts shift. What was once considered AI becomes mundane. AGI represents this ever-receding horizon of an unknowable objective with no foreseeable end.
Section: 1, Chapter: 4
The Myth Of Technological Inevitability
OpenAI repeatedly justified its actions with the 'inevitability card' - if we don't build AGI, someone else will. But as a Chinese AI researcher observed, OpenAI's approach 'never could have happened anywhere but Silicon Valley.' No other country would fund massively expensive technology without clear vision of what it would accomplish.
The explosive global costs of massive models and the race they sparked could only have emerged from one specific combination: billionaire origins, unique ideological bent, and Altman's singular drive and fundraising talent. Nothing about this trajectory was inevitable - it was the product of thousands of subjective choices by those with power to be in the decision-making room.
Section: 1, Chapter: 5
When Data Becomes Weapons
GPT-2's improved capabilities revealed disturbing patterns lurking in training data. Fed prompts about Hillary Clinton or George Soros, the model quickly veered into conspiracy theories. It generated Nazi propaganda and argued against recycling. One AI safety researcher printed out the anti-recycling rant and posted it above office recycling bins as a warning.
This foreshadowed the 'paradigm shift' to come: Instead of filtering inputs, companies would control outputs through massive content moderation operations. The decision to train on 'data swamps' from Common Crawl - the entire internet's worth of toxic content - created the need for armies of traumatized workers to clean up the mess.
Section: 1, Chapter: 5
The Transformer Revolution
Google's 2017 invention of the Transformer neural network became OpenAI's weapon of choice for scaling. Unlike previous models that looked at words in isolation, Transformers could consider vast context - each word in relation to entire documents. Sutskever recognized their scalability potential immediately.
Alec Radford's decision to use Transformers for text generation rather than translation was pivotal. Training the model to predict the next word forced it to compress the essence of language patterns. This 'intelligence is compression' philosophy drove OpenAI's belief that bigger models with more parameters would inevitably become more intelligent.
Section: 1, Chapter: 5
Scale At Any Cost
OpenAI's core philosophy emerged from Ilya Sutskever's belief that more compute equals more intelligence. He theorized that since biological intelligence correlated with brain size, digital intelligence should emerge from scaling simple neural networks to have more nodes.
This led to 'OpenAI's Law' - the observation that compute use in AI had grown 30 million percent in six years, doubling every 3.4 months. The company committed to matching or exceeding this pace, requiring thousands of expensive GPUs costing tens of millions per model. This scaling doctrine became self-fulfilling prophecy that now dominates the entire industry.
Section: 1, Chapter: 5
The Three Clans At War
Altman identified three competing factions within OpenAI: Exploratory Research (advancing capabilities), Safety (preventing catastrophic risks), and Startup (moving fast to commercialize). Each clan had important values but increasingly came into conflict.
The Safety clan, influenced by Effective Altruism, believed in slowing development to prevent existential risks. The Startup clan pushed for rapid deployment and revenue generation. Exploratory Research split between both. These irreconcilable worldviews - 'Doomers' vs 'Boomers' - created internal warfare that ultimately exploded in Altman's firing and reinstatement.
Section: 2, Chapter: 6
Money Changes Everything
OpenAI's noble nonprofit origins quickly crumbled under financial pressure. Elon Musk's departure in 2018 left a funding gap that forced the creation of OpenAI LP, a for-profit arm nested within the nonprofit. What began as ensuring AGI benefits humanity became raising billions for compute-hungry models.
Microsoft's $1 billion investment in 2019 marked the point of no return. The deal gave Microsoft exclusive licensing rights and priority access to OpenAI's technologies. Despite promises that mission would take precedence over profit, commercial imperatives increasingly drove decisions. The 100x return cap for early investors meant someone investing $10 million could make $1 billion - hardly the 'capped profit' structure it claimed to be.
Section: 2, Chapter: 6
Science In Corporate Captivity
Corporate investments in AI jumped from $14.6 billion in 2013 to $235 billion in 2022, while US government allocated just $1.5 billion to non-defense AI. This funding shift fundamentally altered AI research priorities around commercial rather than scientific merit.
From 2004 to 2020, AI PhD graduates heading to corporations jumped from 21% to 70%. Universities could no longer afford the compute needed for cutting-edge research. By 2020, 91% of the world's best-performing AI models came from industry. The result: AI development now serves profit motives rather than scientific advancement or public benefit.
Section: 2, Chapter: 7