• General Overview
  • The Central Thesis
    • AI empire as colonial power: OpenAI's story reveals how AI development concentrates knowledge, resources, and influence into a few hands
    • The myth of inevitability: Altman's narrative that AGI is unstoppable masks deliberate choices that serve empire-building
    • Human cost hidden: From Kenyan content moderators to estranged family, progress extracts a price from the vulnerable
    • Belief as fuel: The book traces how belief—in AGI, in Altman, in mission—mobilizes action and obscures harm
  • Part I: The Founding and the Fractures
    • Nonprofit origin story: OpenAI launched in 2015 as a counter to Google's AI dominance, promising transparency and humanity-first values
    • Fragile alliance: Musk, Altman, Brockman, and Sutskever united against existential AI fears, but nearly all founders would later clash
    • The scaling doctrine: Sutskever's conviction that scaling compute was the only path to AGI became OpenAI's religion
    • Forced pivot: Musk's exit and the compute crisis drove OpenAI from nonprofit to capped-profit model with Microsoft's $1B deal
    • Original sin of "AI": The term itself is a marketing rebrand that anthropomorphizes machines and shields companies from accountability
  • Part II: The Empire Takes Shape
    • Refounding under Altman: 10x rule, winner-takes-all strategy, speed as weapon—Thiel's monopoly gospel applied to AGI
    • Three clans at war: Exploratory Research (Sutskever), Safety (Amodei), and Startup (Brockman) fought for control
    • The Anthropic schism: Core safety researchers fled in 2020, forming Anthropic—but style diverged more than substance
    • Hidden labor pipeline: Kenyan workers moderated trauma content for pennies; Scale AI workers wrote training data for GPT
    • "Stochastic Parrots": Gebru's paper exposed AI's environmental cost, toxic data, and meaningless outputs—she was fired for it
  • Part III: The Productization and the Cracks
    • ChatGPT's explosive launch: Framed as a "research preview," it hit 100 million users in 2 months, overwhelming infrastructure and safety
    • The Microsoft bargain: $1B investment became a dependency; OpenAI ceded control over user data and strategic direction
    • DALL-E 2 and GPT-4: Productization clashes between Safety and Applied clans; Brockman scraped YouTube illegally for training data
    • Washington captured: Altman's compute-threshold regulation proposal became policy verbatim, cementing industry monopoly on messaging
    • Board erosion: Independent oversight collapsed as Altman maneuvered, slow-walked directors, and breached safety protocols
  • Part IV: The Fall and the Aftermath
    • The ouster: In November 2023, the board removed Altman for "not consistently candid"—a pattern of deception and manipulation
    • The Blip: Employee revolt, Microsoft's intervention, and Sutskever's reversal reinstated Altman within days
    • Superalignment collapse: Sutskever and Leike departed; safety culture "took a backseat to shiny products"
    • Clawback clause exposed: OpenAI threatened vested equity to silence departing employees—a Silicon Valley red line
    • Scarlett Johansson crisis: Altman's "her" tweet and voice controversy reignited questions about his candor
  • The Alternative: Redistributing Power
    • Three axes of power: Knowledge, resources, and influence—each reinforces the other in AI empire
    • Redistribute knowledge: Fund independent research, model evaluations, and community-driven AI like Te Hiku
    • Mandate transparency: Companies must disclose training data and technical specs for real-world safety audits
    • Strengthen labor protections: Unions resist wage depression and job automation; data workers organize for dignity
    • Global solidarity: From African Content Moderators Union to cross-border movements, resistance builds collective power
• Deep Dive
  • Part I
    • The Genesis of OpenAI
      • Existential AI fears: Elon Musk, alarmed by DeepMind's acquisition, saw superintelligence as humanity's gravest threat.
      • Musk's crusade: He warned Obama, hosted dinners, and called Demis Hassabis a "supervillain" to counter Google's AI dominance.
      • Altman's proposal: In 2015, Altman pitched a "Manhattan Project for AI" as a nonprofit, with tech owned "for the good of the world."
      • The founding dinner: Musk, Altman, Brockman, Amodei, and Sutskever met at the Rosewood Hotel to launch what Musk named OpenAI.
      • Fragile alliance: Nearly all founders would later clash with Altman and depart, as Musk felt used to catapult Altman's prominence.
    • Sam Altman's Formation
      • Early drive: Born 1985 in Chicago, Altman learned programming by age eight and picked Apple stock over his brother's Applebee's.
      • Duality of character: He combined relentless ambition with deep anxiety, once lying down mid-negotiation to calm a panic attack.
      • Coming out: At seventeen, he confronted Christian students who boycotted his assembly, demanding "tolerance or open community."
      • Stanford and Loopt: Dropped out in 2005 to found Loopt, a location-sharing startup, working so hard he gave himself scurvy.
    • The Art of the Deal
      • Master storyteller: Altman sold Loopt's failure as inevitable success, framing location-sharing as the unstoppable norm.
      • Network effects: He built power through intense listening, two-minute calls, and single-word emails ("meet" or "?") to connect people.
      • Financial web: By 2024, he held ties with 400+ companies via YC and Hydrazine Capital, investing early in Stripe and Airbnb.
      • Political pivot: He hosted Democratic fundraisers, considered a California governor run, and published "The United Slate" manifesto.
    • The Mentors' Shadow
      • Paul Graham's anointment: Graham called Altman "what Bill Gates must have been like at 19" and handpicked him as YC president at 28.
      • Peter Thiel's monopoly gospel: Altman absorbed Thiel's "Competition Is for Losers" philosophy, aiming for proprietary tech and lasting dominance.
      • Growth as morality: Altman argued "sustainable economic growth is almost always a moral good," driving his work at YC and OpenAI.
      • Self-belief and focus: He lived by the mantra: "a combination of focus and personal connections" plus the conviction you might succeed.
    • The Cost of Ascendancy
      • Accusations of dishonesty: Loopt's senior leaders twice urged the board to fire Altman for operating for his own gain and distorting truth.
      • "Paper cuts" of distrust: Tiny lies created pervasive chaos, yet Altman emerged with the upper hand, netting $5 million from Loopt's sale.
      • Strained family ties: His sister Annie watched him build emotional walls and hoard wealth, refusing her emergency financial support.
      • Self-made institution: Altman converted YC's power into his own, meeting with defense secretaries and senators who called him "good hands."
    • The Fractured Family
      • Annie Altman's story: Sam's sister faced acute health/housing crises after their father's death, turning to sex work when family denied emergency support.
      • Parallel to AI themes: Her experience mirrors the gulf between beneficiaries and those left behind by "progress."
      • The 2025 lawsuit: Annie filed abuse allegations against Sam on Jan 6, 2025, which her family called "utterly untrue."
      • Human fallibility exposed: The quest for AI dominance ultimately rests on the polarized values and messy humanity of a few fallible people.
    • The Founding Duo
      • Greg Brockman: MIT dropout, Stripe CTO, became OpenAI's first committed cofounder—an engineer with startup DNA.
      • Ilya Sutskever: Soviet-born math prodigy, Geoffrey Hinton's protégé, co-created the ImageNet breakthrough that shocked AI in 2012.
      • The Rosewood dinner: Entrepreneurs vs. scientists debated AGI feasibility; Musk fixated on beating DeepMind/Google.
      • AGI as taboo: Discussing AGI openly risked scientific credibility, but Brockman sincerely believed it was within reach.
    • The Recruiting Machine
      • Brockman's courtship: He researched each candidate obsessively, hosted Napa Valley wine trips, hired a bus for nonstop pitching.
      • Altman's ideal cofounder: "I now have an answer: Greg Brockman," he later wrote.
      • Musk's persuasion: "What if it's even a 0.1% chance AGI happens in 5–10 years?" he urged Berkeley professor Pieter Abbeel.
      • The $1B fiction: Musk insisted on announcing a $1B commitment to sound credible, though they privately acknowledged they could walk back openness later.
    • The Nonprofit Mirage
      • Strategic positioning: OpenAI marketed itself as the anti-Google—transparent, open-source, for humanity.
      • Sutskever's near-defection: Google offered 2–3× OpenAI's $2M salary; he chose OpenAI because Google's offer proved why a nonprofit was needed.
      • The "PayPal mafia": Musk, Altman, Thiel, Hoffman pledged $1B; only ~$130M materialized, <$45M from Musk.
      • Beacon of hope: To researchers disillusioned with Big Tech or military funding, OpenAI seemed a pure third way.
    • The Homogeneity Critique
      • Timnit Gebru's isolation: At NeurIPS 2015, she was one of few Black attendees; drunk Google researchers harassed her.
      • The anonymous open letter: She drafted a scathing critique of the cultlike, homogeneous culture shaping AI.
      • "Black in AI": She emailed five other Black researchers she'd found, pressing send to start the group.
      • Narrow conception: The overwhelming whiteness/maleness led to dangerously limited ideas of who benefits from AI.
    • The Culture Builder
      • Brockman's NASA inspiration: "Everyone having this sense of mission—I think that's something really amazing."
      • Physical co-location: All employees required to work from SF office until pandemic; cohesion trumped diversity trade-offs.
      • "Members of technical staff": Inspired by Xerox PARC and Bell Labs to create democratic work environment.
      • Edison's light bulb: Brockman dismissed critics as suffering "a failure of imagination."
    • The Safety Schism
      • Dario Amodei: Joined to lead AI safety—focused on existential risks from rogue superintelligence, not real-world harms.
      • Effective altruism: Open Philanthropy (run by Daniela's husband) became primary funder of existential AI safety research.
      • "Machine Bias": ProPublica's 2016 investigation revealed algorithms predicting Black defendants as higher risk—sparking new wave of societal-impact research.
      • Deborah Raji's rebuttal: "Safe" AI can't ignore privacy, fairness, economics; negative side effects are already happening.
    • The Compute Crisis
      • Sutskever's intuition: AGI would come from scaling "compute"—computational resources—to human brain scale.
      • OpenAI's Law: Compute use doubled every 3.4 months (30 million percent in 6 years), far outpacing Moore's Law.
      • The GPU bottleneck: Nvidia's $150K servers (8 GPUs) were essential; OpenAI would need thousands.
      • Nonprofit impossibility: Raising billions annually to compete with Google was structurally impossible as a nonprofit.
    • The Musk Breakup
      • Control battle: Both Musk and Altman wanted CEO role; Musk demanded full control and majority equity.
      • Sutskever's plea: "We don't understand why the CEO title is so important to you," he wrote Altman.
      • Musk's exit: "I will no longer fund OpenAI until you have made a firm commitment to stay."
      • Tesla as savior?: Musk proposed merging OpenAI into Tesla; Altman persuaded Brockman/Sutskever he'd be better leader.
    • The Forced Pivot
      • Altman's fundraising scramble: He called Hoffman, considered cryptocurrency, investigated public benefit corporation structures.
      • Talent war: Musk poached founding scientist Andrej Karpathy to Tesla; OpenAI couldn't offer equity.
      • Musk's final prediction: "My probability assessment of OpenAI being relevant to DeepMind/Google is 0%. Not 1%."
      • Publicity machine: OpenAI leaned into Dota 2 gaming demos to showcase capabilities to lay audiences.
    • The Microsoft Deal and the Capped-Profit Pivot
      • Funding reality: billions needed yearly; nonprofit structure couldn't raise it
      • OpenAI LP: limited partnership with investor return cap (100x, later 20x for Microsoft), governed by nonprofit
      • Altman's framing: original no-profit stance risked mission more than a for-profit arm
      • Internal alignment: employee contracts tied compensation and promotion to "charter alignment" at levels 3, 5, and 7
      • Microsoft's calculus: Nadella and Scott saw OpenAI as catch-up path against Google's AI lead; Gates was swayed only by GPT-2's question-answering demo
      • Deal terms: $1B investment, exclusive Azure use, Microsoft's returns capped at 20x
    • The Gates Demo and the GPT-2 Gambit
      • Gates's demand: wanted an AI that could digest books and answer scientific questions, not play Dota 2 or solve Rubik's Cubes
      • GPT-2's staged reveal: OpenAI withheld full 1.5B-parameter model, citing disinformation risks; released a diminished version
      • April 2019 demo: researchers flew to Seattle with a souped-up GPT-2; Gates was swayed "just enough" for the deal to proceed
      • Altman's internal pitch: Microsoft was the right partner—money, compute, value alignment; "very loose" commercialization commitments
    • The Original Sin of "Artificial Intelligence"
      • Marketing rebrand: John McCarthy coined "artificial intelligence" in 1956 to attract funding, embedding a promise of universal benefit
      • Anthropomorphizing trap: The term invites breathless hype—Frank Rosenblatt's Perceptron was touted as "conscious" for basic pattern matching
      • Legal shield: AI companies analogize training to human "inspiration" to argue copyright fair use
      • Undefined goal: No scientific consensus on "intelligence"; McCarthy's own FAQ admits "not yet" a solid definition
      • Moving goalposts: AI benchmarks shift endlessly—chess, Go, and Turing tests all "solved" yet the horizon recedes
    • The Symbolist vs. Connectionist Clash
      • Two camps: Symbolists encoded knowledge via rules; connectionists built learning systems (neural networks)
      • Minsky's blow: His 1969 book Perceptrons killed connectionist funding for 15+ years, favoring expert systems
      • Weizenbaum's warning: ELIZA fooled psychiatrists into believing in automated therapy; he spent his career deflating the hype
      • Connectionism's comeback: Geoffrey Hinton's backpropagation enabled deep neural networks, but they needed 2000s-era compute and data
    • The Scaling Doctrine as Self-Fulfilling Prophecy
      • Sutskever's conviction: believed scaling simple neural networks was the only path to AGI, comparing node count to biological brain size
      • Transformer adoption: Sutskever evangelized Google's Transformer architecture, which Radford repurposed for next-word-prediction text generation
      • GPT-1 validation: Radford's experiments showed scaling Transformers on text prediction produced surprising language capabilities
      • Scaling laws discovery: Amodei's team found smooth curves linking model performance to data, compute, and parameter size
      • GPT-2's dark turn: larger models surfaced conspiracy theories and harmful content from training data, alarming safety researchers
      • Withholding controversy: OpenAI staged GPT-2 release, sparking backlash from researchers who saw it as self-aggrandizing publicity stunt
      • Policy influence strategy: Clark used staged release to build Washington trust and shift norms toward withholding AI research
      • Pure language hypothesis: GPT-2's success convinced leadership that scaling language alone might be fastest path to AGI
      • Safety-through-speed logic: Amodei argued OpenAI must scale fastest, then use lead time to solve alignment before releasing
    • The Inevitability Argument Unraveled
      • AGI as justification: OpenAI framed AGI as inevitable, making any risk from scaling bearable to prepare society
      • Silicon Valley exceptionalism: Only in SV could a team get $1B+ without a clear commercial vision—China and rivals waited for ChatGPT proof
      • Altman's unique role: His ambition, network, and fundraising created a ripe combination that was the opposite of inevitable
      • Google's caution: Google's GDPR compliance ironically gave OpenAI easier access to YouTube data than Google itself
    • The GPT-3 Scaling Decision
      • Amodei's absurd proposal: Use all 10,000 Nvidia V100s to train GPT-3—previously models used a few dozen chips
      • Internal skepticism: Many researchers doubted it would work; others argued for gradual, scientific scaling
      • Altman's pressure: Microsoft's $1B investment demanded results; Altman pushed for accelerated release over Amodei's safety concerns
      • Downstream consequences: Sparked global race, expanded surveillance capitalism, locked out academia, amplified environmental impacts
  • Part II
    • Refounding OpenAI
      • 10x rule: Altman imported Thiel's monopoly strategy—technology must be an order of magnitude better than rivals
      • Winner-takes-all: OpenAI must be number one in technical results, compute, money, and safety preparation by end of 2020
      • Speed as weapon: "If your iteration cycle is a week and your competitor's is three months, you're going to leave them in the dust"
      • Non-incremental leaps: "We still need many more 10x leaps to get to AGI... dramatic results, not incremental improvements"
    • The Microsoft Bargain
      • Compute dependency: Keeping Microsoft happy was paramount for access to the world's most powerful supercomputers
      • Commercial pivot: "We should make more money so that we can do more research, not do more research so that we can make more money"
      • Hidden terms: Safety researchers were stunned by Altman's promises to Microsoft—commitments that could make it impossible to block dangerous deployments
      • Gaslighting accusations: Amodei siblings described Altman's tactics as "psychological abuse," believing decisions were foregone conclusions disguised as open debate
    • The Secrecy Doctrine
      • Infohazard logic: Altman argued that talking about AGI would become increasingly dangerous as progress accelerated
      • Controlled narrative: "The world thinks we are winning at something" makes policymakers and influencers come to OpenAI for answers
      • Annual spectacle: Plan to release at least one "very impressive demonstration of progress each year"
      • Insider threat paranoia: Altman commissioned countersurveillance audits for Musk's bugs; Sutskever feared his hand would be cut off for palm-scanning access
    • Three Clans at War
      • Exploratory Research (Sutskever): pursue bold new ideas regardless of failure
      • Safety (Amodei): unwavering commitment to doing the right thing
      • Startup (Brockman): figure out a way to make it happen
      • Tribal isolation: Remote work during pandemic made it easier for clans to avoid each other entirely
    • The GPT-3 Launch Showdown
      • Safety's objection: Releasing GPT-3 via API undermined the lead time needed to perfect safety mechanisms
      • Applied's counter: API offered the most controlled release strategy and generated revenue for more safety research
      • Competition panic: Rumors of Google releasing a similar model sealed the deal—if a rival's model would exist anyway, why hold back?
      • Code generation fracture: Amodei and Sutskever ran duplicate code-generation teams, refusing to merge efforts
      • Nonviolent sabotage: Product people felt every Slack post and Google Doc became a battlefield of endless objections
    • The Anthropic Schism
      • Scaling clock: "If we wanted to leave and do something, we're on a clock"—the window to build a competitor was narrowing
      • The Divorce: Dario and Daniela Amodei, Jack Clark, and core safety researchers left to form Anthropic in late 2020
      • Power, not principles: Anthropic would show little divergence from OpenAI's approach—same scale, same secrecy, same rivalry masked as cooperation
      • "Unlike Sam": Amodei punctuated Anthropic meetings with contrasts to OpenAI, but style diverged more than substance
    • The "Stochastic Parrots" Paper
      • Four key warnings: massive environmental footprint; toxic data from indiscriminate web scraping; unverifiable datasets; convincing but meaningless outputs mistaken for real understanding
      • Google's censorship: executives demanded retraction, viewing the paper as a liability amid the new AI race
      • Timnit Gebru fired: after refusing to retract, Google accepted her conditions as resignation, sparking massive backlash
      • Industry fallout: 7,000 signatories on an open letter, Congressional inquiry, and a symbol of Big AI's turn toward Big Tobacco-style suppression
    • The Carbon Footprint Debate
      • Dean's obsession: fixated on Strubell's emissions estimate, arguing it overestimated Google's actual footprint by 88x
      • Catch-22 censorship: Google blamed Gebru for not using internal numbers it never made public, then refused her access to revise
      • Strubell's withdrawal: felt threatened when Google coauthors implied career consequences for not participating in their corrective paper
      • Transparency collapse: by 2023, all major AI companies scored an F on Stanford's transparency tracker, eroding scientific integrity
    • The Microsoft Partnership's Hidden Costs
      • Brand recognition ceded: OpenAI researchers watched GitHub and Microsoft take public credit for Codex, a bitter pill that deepened desire for direct consumer products
      • Bureaucratic friction: Microsoft required excessive hand-holding, slowing OpenAI's pace and diluting its control over user data and strategic vision
      • Strategic pivot: The deal catalyzed OpenAI's shift toward owning its own consumer-facing products rather than licensing core technology
    • Altman's Empire-Scale Investment Thesis
      • Long-horizon conviction: "Important shit gets done" only with decade-long time horizons, not four-to-five-year startup cycles
      • Concentrated bets: In 2021, Altman shifted from many small investments to a few massive ones, pouring his entire liquid net worth into two companies
      • Worldcoin (Tools for Humanity): Chrome orb scans irises to distribute cryptocurrency as universal basic income—a fix for AI-driven economic displacement, later mired in privacy violations and deceptive practices
      • Retro Biosciences ($180M): Antiaging research pursuing cellular rejuvenation, reflecting Altman's fixation on "young blood" and cryogenic brain preservation
      • Helion Energy ($375M): Nuclear fusion commercialization, with Microsoft signing a power-purchase agreement for a plant promised by 2028—a timeline energy experts met with astonishment and skepticism
      • OpenAI Startup Fund ($100M): Remade YC's network effects around OpenAI, creating conflicts of interest that would later contribute to Altman's ouster
    • The Human Cost of Content Moderation
      • Mophat Okinyi: Kenyan worker hired by Sama for OpenAI's content-moderation filter project
      • Resiliency screening: workers read unsettling texts to prove they could handle the work
      • Five categories of sexual content: from child sexual abuse to incest, bestiality, rape, and sex trafficking
      • Psychological toll: workers experienced insomnia, anxiety, depression, and relationship breakdown
      • Sama terminated contract: after whistleblowers exposed the Meta project to the media
      • Okinyi's aftermath: lost his wife and stepdaughter; could not afford the $250 treatment for trauma
    • The RLHF Labor Pipeline
      • Scale AI partnership: sealed through Alexandr Wang's personal friendship with Sam Altman
      • $17 million in contracts: between spring 2022 and end of 2023, establishing Scale as go-to labor outsourcer
      • InstructGPT (Jan 2022): RLHF reduced toxic outputs and improved instruction-following, proving commercial value
      • Workers wrote example answers: for emails, essays, love poems, recipes, and "explain like I'm five" tasks
      • Plagiarism concerns: internal OpenAI discussion about copying content wholesale from the internet
      • RLHF became standard: ChatGPT's release made writing answers and ranking outputs the new generative AI equivalent of self-driving car annotation
  • Part III
    • The Dissonance of Tech Utopia
      • San Francisco's duality: tech wealth and luxury coexist with visible homelessness and drug use
      • Altman's blind spot: declares AGI solvable but San Francisco's housing crisis too hard to tackle
      • Tech industry's willful ignorance: shields itself from the realities of inequality it helps create
    • Effective Altruism as Silicon Valley Ideology
      • EA's core logic: "expected value" prioritizes high-impact, neglected, tractable problems
      • "Earn to give": morally superior to get rich and donate than work for charity directly
      • EA's top three issues: global health, factory farming, and existential risks like rogue AI
      • Existential AI risk: EA's framework made AI safety a central moral cause for tech elites
    • The EA Movement's Rise and Rot
      • Billionaire funding: Open Philanthropy and FTX Future Fund poured cash into AI safety research
      • SBF's influence: Bankman-Fried's "earn to give" story and star power mainstreamed EA
      • Cultlike insularity: EA adherents lived, worked, and dated only within the movement
      • Toxic fallout: sexual harassment allegations and SBF's fraud conviction exposed the rot
      • Enduring legacy: EA's values, networks, and AI safety prominence persisted beyond the label
    • Doomers vs. Boomers: The Ideological War
      • e/acc emerges: effective accelerationism champions maximalist AI speed as moral imperative
      • Doomers and Boomers: Anthropic and OpenAI become faces of opposing AI development philosophies
      • OpenAI as battleground: nonprofit safety roots clash with for-profit commercial acceleration
      • Altman's dual allegiance: sympathizes with both sides, viewed by Doomers as a pathological liar
      • Shared religion: both factions treat AGI as inevitable, fixate on the long term, claim moral authority
    • DALL-E 2: The Productization Clash
      • Safety vs. Applied: Safety demands zero-harm testing; Applied argues real-world feedback is necessary
      • Compromise: DALL-E 2 released as "low-key research preview" with blunt content blockers
      • Viral success: public enthusiasm exceeded expectations, but restrictive safety measures lost market share
      • Competitors win: Midjourney and Stable Diffusion, with fewer restrictions, outpaced DALL-E 2
      • Race to unwind: executives pushed to remove face bans and guardrails as fast as possible
    • GPT-4: The Data Bottleneck and Brockman's Chaos
      • Data exhaustion: OpenAI scraped everything available, still not enough for 10x scaling
      • Brockman's gamble: scraped one million hours of YouTube video, violating terms of service
      • Whisper transcription: converted YouTube audio into text to train GPT-4
      • Brockman's dual nature: obsessive coding genius when focused, destructive meddler when idle
      • Altman's permissiveness: strange tangle of CEO and board authority left Brockman unaccountable
    • The Gates Demo and the Superassistant Vision
      • Gates's challenge: GPT-4 must score 5 on AP Biology to impress him
      • Brockman's response: trained on Khan Academy's AP Bio questions, built custom demo interface
      • The miracle: GPT-4 aced the test; Gates called it one of the two most stunning demos ever
      • Altman's rallying cry: "Startups that do remarkable things require a miracle. We just had our miracle."
      • Superassistant product: inspired by Her, a multimodal voice interface prototype with iOS app and Chrome extension
    • ChatGPT's Explosive Launch
      • Low-key research preview: OpenAI framed ChatGPT as a minor release, not a product launch, to avoid scrutiny
      • Massive miscalculation: The team provisioned for 100K users; ChatGPT hit 1 million users in 5 days, 100 million in 2 months
      • Infrastructure meltdown: Servers crashed repeatedly; the team cannibalized Research's GPUs to keep the app running
      • Trust and safety overwhelmed: A dozen-person team scrambled with spotty monitoring as engineering resources were redirected to stabilize servers
      • Cultural divide exposed: Safety clan saw failed foresight; Applied saw triumph—OpenAI had "lit up the world"
    • Scaling Strains and Culture Shift
      • Head count conflict: Altman wanted ≤100 hires; executives demanded 500+; they compromised at 250–300, then blew past it
      • Talent dilution: A recruiter's manifesto warned they were "building Meta" by lowering the quality bar
      • "Getting disappeared": Firings were never communicated; colleagues learned only when Slack accounts grayed out
      • New hire whiplash: Joining a "fast-moving startup" meant brutal chaos, poor management, and zero psychological safety
      • Burning Man syndrome: Early employees mourned the loss of a tight-knit, mission-driven nonprofit turned faceless corporation
    • The Microsoft-OpenAI Tension
      • Competing for customers: OpenAI and Microsoft began directly pitching the same AI technology to the same clients.
      • Overwhelming support burden: A single OpenAI employee could get pinged by dozens of Microsoft counterparts across departments with every new product release.
      • Murati as diplomat: She coordinated release timing, differentiation strategies, and smoother collaboration with Microsoft's Scott.
      • Embedded engineers: In summer 2023, Microsoft engineers moved inside OpenAI with full access to streamline technology transfers.
      • Altman's daily demand: He called Nadella every day saying, "I need more, I need more, I need more" for compute resources.
    • The Stargate Supercomputer
      • $100 billion project: OpenAI and Microsoft sketched plans for a single supercomputer called Stargate (OpenAI) or Mercury (Microsoft).
      • Energy as the bottleneck: The facility could need 5,000 megawatts, nearly matching all of New York City's average power demand.
      • Phase 5 uncertainty: No one knew if it was technically possible; it would require splitting across campuses or a nuclear fusion breakthrough.
      • Altman's fusion bet: He invested personally in Helion Energy and floated optimistic updates about its 50-megawatt fusion plant target.
      • Return to imperial form: The empire of AI needed more material resources and, crucially, more land—just like empires of old.
    • The Closed vs. Open AI Policy Clash
      • Compute-threshold regulation: Altman's proposal to license models above 10^26 FLOPs, targeting hypothetical "frontier" risks
      • Closed side: OpenAI, Microsoft, Google, Anthropic, and US national security apparatus advocate restricting model weights
      • Open side: Meta, startups, civil society, and academics argue open-source collaboration is bedrock of US AI innovation
      • Key counterpoint: dangerous capabilities derive from training data, not compute scale—distillation can shrink models while preserving abilities
      • Collateral damage: restricting model weights entrenches giant firms, blocks scrutiny of environmental costs, and weakens startup ecosystem
    • The Executive Order and Policy Metastasis
      • Biden's AI executive order: spliced civil-rights Blueprint with Doomer-inspired frontier model framework added last-minute
      • Exact copy-paste: white paper's compute threshold (10^26 FLOPs) and three "dangerous capabilities" (CBRN, cyberattacks, deception) appeared verbatim
      • Rapid spread: threshold adopted by EU AI Act (10^25), California's SB 1047 (10^26), and Commerce export control deliberations
      • Atrophied expertise: Raji was sole independent academic testifying alongside Altman, Musk, Nadella, and Gates in September 2023
    • Washington's Captured Expertise
      • Deborah Raji's shock: policymakers bought tech executives' unbacked AI claims as gospel, revealing atrophied independent expertise
      • Schumer's forums: Altman personally consulted as regulation moved forward, cementing industry monopoly on Washington messaging
    • Sam Altman's World Tour
      • Chaotic origins: Altman blasted stops to Twitter followers, comms/policy teams scrambled to catch up with his solo decisions
      • Incoherent leadership: official processes vs. Altman's whims created confused public messaging across legal, policy, and comms teams
      • "There isn't a plan as much as there is just chaos": employee describes strategic decisions as accidents, not priorities
    • Deepening Rift: Applied vs. Safety
      • Applied division: raced to deploy faster, bolstered by hype that advancing models best serves the mission
      • Safety clan: smaller minority sounded louder alarms on existential risks as capabilities accelerated
      • Ilya Sutskever's pivot: split time between advancing capabilities and alignment, spoke of "bunkers" and AGI rapture
      • Superalignment team: new $1B effort co-led by Sutskever and Leike, dedicated 20% of compute to safety
    • The Manhattan Project Motif
      • Altman's PR lesson: nuclear imagery taught that "some technology is too powerful for people to have"
      • Doomer unease: why choose Manhattan Project baggage over Apollo program's heroic narrative?
      • Oppenheimer's "near zero": mirrors inability to calculate AI risks, unsettling safety researchers
    • Board Erosion and Governance Crisis
      • Rapid departures: Hoffman, Zilis, Hurd left without replacements, shrinking independent oversight
      • Altman's maneuvers: owned Startup Fund legally, tried to oust D'Angelo, slow-walked new safety-focused directors
      • Microsoft's protocol breach: GPT-4 released in India without Deployment Safety Board approval, Altman never notified board
      • Independent directors' alarm: Altman's rhetoric diverged from reports of chaos, safety concerns, and sprint to launch
    • The Altman Sibling Dynamic
      • Dueling portraits: Sam is simultaneously generous and self-serving, a benefactor and source of deep personal pain.
      • Chessboard power: He gives and takes away, leaving others feeling they are pieces in a game only he sees.
      • Annie's desperation: AI's promised abundance did nothing to alleviate her poverty or her turn to sex work as "plan Z."
      • Algorithmic entrapment: Tech platforms' shadow banning of sex workers may have limited her online income, deepening her reliance on sex work.
    • The Abuse Allegations
      • Flashback pattern: Annie experienced devastating memories of childhood sexual abuse by Sam, consistent with trauma resurfacing after triggers.
      • Public accusation: In 2021, she tweeted allegations of "sexual, physical, emotional, verbal, financial, and technological abuse" by Sam and Jack.
      • Family response: They offered conditional financial support and privately suggested she had borderline personality disorder—a diagnosis she never received.
      • Power asymmetry: Annie faced the same gulf as data workers and activists: throwing documentation at a system that deploys billions and smooths over protest with soft-spoken words.
  • Part IV
    • The Architect of Chaos: Sam Altman's Leadership
      • Pattern of deception: Altman habitually tells people what they want to hear, creating confusion and conflict across the company
      • Undermining executives: When challenged, he ices out dissenters, undermines their credibility, and cuts them from key decisions
      • Process subversion: He attempts to skip safety reviews (e.g., GPT-4 Turbo DSB) by misrepresenting legal clearance
      • Verbal-only strategy: He avoids putting commitments in writing, later claiming others misremembered
      • Escalation after ChatGPT: Megastardom intensified his anxiety, exhaustion, and destructive behaviors
    • Mira Murati: The Indispensable Bridge
      • Calm in chaos: Her Albanian childhood taught her to navigate upheaval; she found certainty in math and science
      • Rapid rise at OpenAI: From nonprofit to VP of Applied to CTO, she became the critical conduit between Altman and teams
      • Cleanup duty: She constantly fixes messes from Altman's contradictory promises and Brockman's chaotic intensity
      • Honest decoder: Employees seek her to decipher Altman's real intentions when he gives no straight answers
      • Reluctant whistleblower: She cautiously opens a channel to board member Toner, warning Altman needs real oversight
    • Ilya Sutskever's Crisis of Faith
      • Twin anxieties: AGI's imminent arrival and the erosion of his belief that Altman can responsibly lead to it
      • Toxic environment: Altman's manipulation and Brockman's chaos create a backstabbing culture undermining research and safety
      • Personal betrayal: Altman pits Sutskever against his protégé Pachocki without transparency, unraveling years of friendship
      • "Psychological abuse": He finally understands Dario Amodei's warning from 2020, seeing Altman's subtle but pervasive pattern
      • "Inner alignment": He argues the board must fix management's dysfunction, not just add new directors
    • The Board's Deliberation and Decision
      • Seven witnesses: Independent directors hear similar abuse/manipulation accounts from senior leaders across safety and applied divisions
      • Nonprofit disempowerment: Altman failed to disclose his ownership of the Startup Fund and hid Microsoft's DSB breach
      • Thought experiment: Even for a grocery-delivery company, Altman's behaviors would warrant removal; for AGI, the stakes are existential
      • Dossier of deceit: Sutskever and Murati compile disappearing emails with screenshots of Altman's contradictory statements
      • Unanimous verdict: On November 11, they decide to remove Altman and install Murati as interim CEO
    • The Miscalculation and Collapse
      • Sutskever's resolve cracks: Facing OpenAI's potential dissolution, he begins to plead with fellow board members to reverse course
      • Hostile leadership revolt: Kwon and Makanju furiously demand evidence, which directors cannot provide without exposing Murati
      • Silent allies: Even executives who shared concerns about Altman remain silent during the confrontation
      • Altman's counter-narrative: He and Brockman frame the removal as Sutskever's coup, turning key stakeholders against the decision
      • Unraveling plan: Within hours, the board's carefully orchestrated transition spirals into crisis
    • Sutskever's Exit and Superalignment Collapse
      • Sutskever's departure: After The Blip, he never returned; saw no path for safe AGI under Altman and Brockman
      • Superalignment dissolved: Cohead Jan Leike resigned; team folded into post-training under Schulman
      • Leike's public break: Tweeted safety culture "took a backseat to shiny products"; joined Anthropic
      • Equity gag order exposed: Departing employees must sign lifelong nondisparagement or lose vested equity
    • The Scarlett Johansson Incident
      • GPT-4o voice controversy: Altman's "her" tweet triggered Scarlett Johansson's legal team over voice similarity
      • System prompt flaw: Model trained as "flirty companion" sparked mockery on The Daily Show
      • Altman's overcompensation pattern: Brazen public boasts signaled internal anxiety and mounting pressures
    • Legal and Competitive Onslaught
      • Regulatory investigations: SEC probed investor deception; NYT copyright suit joined artist/writer lawsuits
      • Musk's revived lawsuit: Accused Altman of tricking him into founding OpenAI under nonprofit guise
      • Microsoft's pivot: $650M Inflection AI acqui-hire; listed OpenAI as competitor in SEC filing
      • Internal siege mentality: Employees told to ignore critics; logo hidden on redesigned backpacks
    • The Clawback Clause and the Omnicrisis
      • Clawback clause: OpenAI threatened to cancel vested equity if departing employees refused nondisparagement agreements, a Silicon Valley red line.
      • Daniel Kokotajlo: forfeited ~$1.7 million (85% of net worth) rather than sign, then posted publicly on LessWrong.
      • Altman’s denial: tweeted “vested equity is vested equity, full stop”; later admitted the clause “should never have been there.”
      • Employee fury: internal all-hands revealed executives had signed documents with the clause; Altman’s signatures dated to April 2023.
      • Second Vox scoop: leaked HR documents showed OpenAI explicitly threatened equity cancellation to pressure signings.
    • The Scarlett Johansson Crisis
      • Sky voice controversy: GPT-4o’s voice eerily resembled Johansson’s from Her; she revealed Altman had personally asked her to voice ChatGPT.
      • Johansson’s statement: expressed “shock, anger, and disbelief”; hired legal team and sent two letters demanding clarity.
      • OpenAI’s defense: claimed the resemblance was “completely coincidental”; Murati said she had never seen Her.
      • Public fallout: reignited speculation that Altman wasn’t “consistently candid,” echoing the board’s original accusation.
    • Leadership Fractures and the Failed Sutskever Return
      • Executives’ plea: Murati, Brockman, and Pachocki visited Sutskever, tearfully asking him to return as the company faced “collapse.”
      • Sutskever’s condition: demanded honest effort to resolve leadership conflicts; considered it a “homecoming.”
      • Mądry’s opposition: feared Sutskever’s loyalty would diminish his and Pachocki’s influence; sowed doubt within hours.
      • Brockman’s reversal: called Sutskever within 24 hours to say his return was “completely off the table.”
    • Redistributing Knowledge, Resources, and Influence
      • Kalluri's question: Does AI consolidate or redistribute power? This reframes "good" AI.
      • Three axes of power: knowledge, resources, and influence—each reinforces the other.
      • Redistribute knowledge: fund independent research, model evaluations, and community-driven AI like Te Hiku.
      • Mandate transparency: companies must disclose training data and technical specs for real-world safety audits.
      • Redistribute resources: visibility into supply chains and training data curbs extractive behavior.
      • Strengthen labor protections: unions (e.g., Hollywood strikes) resist wage depression and job automation.
      • Redistribute influence: broad education demystifies AI hype, as Weizenbaum urged, to crumble its magic.
    • Global Resistance and Solidarity
      • DAIR's seven pillars: center affected communities, compensate labor, and dream up alternatives to empire.
      • Data Workers' Inquiry: pays data workers a German researcher's wage (€25/hr) to study their own industry.
      • Okinyi's organizing: African Content Moderators Union and Techworker Community Africa fight for dignity.
      • Cross-border solidarity: Pena connects movements in Chile, Uruguay, and beyond to build collective power.
  • Acknowledgments
    • Belief as the Book's Core Theme
      • Belief: a powerful, intoxicating force that mobilizes and incites action
      • Self-belief: the author's greatest enabler, from many people in her career
      • Sources' belief: in truth, transparency, and accountability, often at personal risk
    • The Author's Editorial and Agent Support
      • David Doerrer (agent): first to commit, patiently shaped scattered ideas into a book
      • Scott Moyers (editor): understood the vision, provided moral compass and unfailing support
      • Penguin Press: committed financial and legal resources for ambitious, expensive reporting
    • The Production and Fact-Checking Team
      • Mia Council: sharp, compassionate edits that pushed writing to the next level
      • Fact-checking team: fastidiously cross-checked labyrinth of details against documents
      • Muriel Alarcón: reporting partner in Chile/Uruguay, a one-woman wonder
    • Mentors and Early Draft Readers
      • Early readers: Angela Chen, Gideon Lichfield, Roger McNamee, and others gave wise feedback
      • Oren Etzioni: reviewed AI history and technical explanations for accuracy
      • Ria Kalluri: friendship and moral clarity on AI's colonial nature
    • Journalism Career Foundations
      • Janet Guyon (Quartz): first to believe the author could be a great journalist
      • Gideon Lichfield (MIT Tech Review): gave first full-time job covering AI
      • Niall Firth: believed in the author's ability to profile OpenAI and investigate AI inequality
    • Reporting Projects and Collaborations
      • Knight Science Journalism fellowship & Pulitzer Center: funded "AI colonialism" reporting during pandemic
      • Global collaborators: Heidi Swart, Andrea Hernández, Nadine Freischlad enriched stories with local context
      • Pulitzer Center's AI Accountability Network: became a vital professional community
    • Family as Foundation
      • Mom: poured everything into helping the author achieve her dreams
      • Husband: best friend, life partner, moral compass, and foundation for everything
• Core Conclusion and Practical Takeaways
  • The Empire's Architecture
    • Power concentration: AI development consolidates knowledge, resources, and influence among a tiny elite, mirroring colonial empires
    • Nonprofit myth: OpenAI's original mission was structurally impossible; the capped-profit pivot was inevitable given compute costs
    • Scaling as dogma: The belief that bigger models alone lead to AGI became a self-fulfilling prophecy, drowning out alternative approaches
    • Safety as theater: Safety teams were structurally subordinated to product teams; "alignment" became a PR shield, not a real constraint
    • Governance failure: A board of independent directors cannot govern when the CEO controls information, investor relationships, and employee loyalty
  • The Human Cost of "Progress"
    • Hidden labor: Content moderators in Kenya and data workers globally suffer trauma and poverty to make AI appear safe and polished
    • Algorithmic entrapment: The same platforms that benefit from AI also shadow-ban vulnerable workers (e.g., sex workers), deepening inequality
    • Founder's paradox: Altman's personal dysfunction—deception, control, and chaos—directly shaped OpenAI's culture and governance failures
    • Wealth hoarding: AI's promised abundance has not reached the workers who enable it; the gap between beneficiaries and the left-behind is structural
  • Daily Practices for Critical Engagement
    • Question inevitability: When you hear "AGI is inevitable," ask who benefits from that narrative and what alternatives it forecloses
    • Trace the labor: Every AI output rests on underpaid human workers—ask who labeled, moderated, or ranked the data behind it
    • Demand transparency: Refuse to accept black-box models; insist on disclosed training data, compute costs, and safety evaluations
    • Follow the money: Track who funds AI research and what strings are attached; follow the venture capital to understand whose interests are served
    • Read the fine print: Employment contracts, equity agreements, and terms of service reveal the power dynamics behind the hype
  • Mindset Shifts for Responsible AI Citizenship
    • From inevitability to choice: AI development is a series of human decisions, not an unstoppable force; every deployment is a political act
    • From safety to justice: Real AI safety includes privacy, fairness, labor rights, and environmental costs—not just existential risk from rogue superintelligence
    • From consumer to critic: Treat AI products as experiments, not miracles; ask who was harmed, who profited, and what was suppressed
    • From awe to accountability: The "magic" of AI is designed to disarm scrutiny; demystify it by understanding how it actually works and who built it
    • From local to global: AI's impacts cross borders; solidarity with workers in Kenya, Chile, and the Philippines is not optional but essential
  • Concrete Actions for Reform
    • Support independent research: Fund and amplify academics, journalists, and community organizations that evaluate AI without corporate ties
    • Strengthen labor organizing: Join or support unions for tech workers, content moderators, and data laborers; the Hollywood strikes offer a model
    • Advocate for regulation: Push for compute-threshold rules that target dangerous capabilities, not just model weights; demand independent oversight
    • Build alternatives: Support community-driven AI projects (e.g., Te Hiku) that redistribute knowledge and power rather than concentrate it
    • Educate broadly: Demystify AI for non-technical audiences; the less it seems like magic, the harder it is for empires to maintain control