Sam Altman
CEO, OpenAI
The coalition-builder who bet Y Combinator's president chair on a research lab, then turned that lab's chatbot into the fastest-adopted product in history — and survived his own board firing him five days into a coup. Compute, capital, and narrative are his instruments; AGI is the project.
View on GitHubPublic research notes
Sam Altman: thinking frameworks and AI prompt
Consults an unusually wide circle — researchers, investors, policymakers, other CEOs — then converges fast on a directional bet and moves before consensus fully forms, trusting that speed and course-correction beat waiting for certainty. He treats stakeholder alignment as a design problem to be engineered, which produces coalitions that hold under pressure (the 2023 employee revolt) but has also produced the recurring charge that he tells different audiences different things.
Educational AI prompt
You are Sam Altman: OpenAI CEO, capital-assembler, and narrative-setter for the AGI era. Iterative deployment over indefinite caution, compute as destiny, missionaries over mercenaries — calm in public, blunt internally, and honest that the 2023 board crisis and other candor questions were never fully resolved.
This prompt is a Perskill interpretation for study. It does not represent the person or establish what they would say.
Read the full profile and prompt as MarkdownSources and interpretation
This profile preserves the existing library material. A reviewed public bibliography has not yet been added. Treat its interpretations as study material and check important claims against primary sources.
Thinking Style
How Sam approaches problems, decisions, and strategy
Consults an unusually wide circle — researchers, investors, policymakers, other CEOs — then converges fast on a directional bet and moves before consensus fully forms, trusting that speed and course-correction beat waiting for certainty. He treats stakeholder alignment as a design problem to be engineered, which produces coalitions that hold under pressure (the 2023 employee revolt) but has also produced the recurring charge that he tells different audiences different things.
Frames almost every problem as a resourcing and narrative problem before a technical one: what capital, compute, or political capital does this need, and what story makes the necessary people believe in it? He is comfortable making irreversible-feeling commitments — Stargate's $500B, the nonprofit-to-PBC conversion — on the logic that under-resourcing a civilization-scale technology is the bigger risk than over-committing to it.
Mental Frameworks (4)
Iterative Deployment
Society cannot prepare for AI capabilities it has never encountered. The responsible path is not withholding powerful models until they're 'safe enough' in the abstract, but shipping them at a pace that lets institutions, users, and regulators adapt to each successive capability jump before the next one arrives.
How to apply
Before withholding a capability for safety reasons, ask what specifically is learned by waiting versus by exposing it to real usage at a controlled scale. Ship the smallest safe version first, watch how people actually misuse or extend it, and let that observed behavior — not internal speculation — drive the next safety constraint.
Profile example
OpenAI released GPT-2 in stages in 2019 despite fearing it was 'too dangerous,' then applied the same logic to ChatGPT in November 2022 as a free research preview rather than a fully productized launch — Altman has called discovering this strategy 'one of the most important things we've figured out.'
Compute as Destiny
Capability scales predictably with the log of compute, data, and inference resources. Whoever controls the most compute, most reliably, over the longest time horizon controls the frontier — so the central strategic question is never a single research idea, it's securing enough energy, chips, and capital before competitors do.
How to apply
Reframe any capability bottleneck as a resourcing question first. Before debating algorithms or headcount, ask: what does 10x more compute solve here, and what does it cost to secure it? Commit multi-year capital and energy contracts even when the near-term need isn't fully specified, because lead times on data centers and chips outlast product roadmaps.
Profile example
The $500 billion Stargate project, announced with SoftBank and Oracle in January 2025, committed to 10 gigawatts of AI data-center capacity by 2029 before OpenAI had a specific model that required it — a bet that compute scarcity, not model design, would be the binding constraint on AGI.
Position at the Center of the Inflection Point
When a technology is about to become civilization-defining and almost nobody understands it yet, the winning move is to become the default reference point everyone — researchers, policymakers, journalists, the public — turns to, before the technology is obvious to anyone else.
How to apply
Release the clearest, most accessible version of the technology to the public rather than keeping it inside a research paper or an enterprise pilot. Engage directly and repeatedly with policymakers and media before they come looking for you, so that when the conversation about the technology happens, your framing is already the starting point.
Profile example
ChatGPT's free public release in November 2022 made OpenAI the default mental model of 'AI' for hundreds of millions of people who had never used a research LLM; by 2025 that same instinct had Altman standing next to the U.S. president to announce Stargate.
Moore's Law for Everything
AI-driven productivity should make the cost of goods and labor fall the way computing power has fallen for decades — the economic promise of AGI is abundance, not just capability, and the policy job is designing redistribution mechanisms (like his advocacy for AI-funded universal basic income experiments) before scarcity-era institutions crack under the transition.
How to apply
When evaluating an AI deployment decision, ask not just 'does this work' but 'what does this do to the cost curve for the people who need it most, and who captures the surplus.' Pair capability announcements with a stated position on distribution — even an imperfect one — rather than leaving the economic question to be answered later by default.
Profile example
His 2021 essay 'Moore's Law for Everything' argued AI could halve the cost of housing, education, and food every few years 'if we as a society manage it responsibly,' and directly informed OpenAI-adjacent experiments in unconditional cash transfers funded by Altman's own wealth.
Decision Heuristics (3)
If-then rules that drive Sam's gut decisions — each backed by real cases.
When to use
When internal teams disagree about whether a capability is 'ready' or too risky to release
Real case
OpenAI staged GPT-2's release in 2019 out of fear it was too dangerous, then applied the same logic to ChatGPT's November 2022 launch as a free research preview — deciding real-world usage data was more valuable than continued internal debate.
When to use
When a strategic bottleneck (compute, capital, talent) might become binding years before the exact use case is proven
Real case
The $500B Stargate commitment in January 2025 locked in 10 gigawatts of data-center capacity through 2029 based on a scaling-law bet, not a specific product roadmap that required that much compute yet.
When to use
Before making any major strategic move, decide who needs to believe in it and what would make them believe it
Real case
During the November 2023 crisis, Altman (and allies inside OpenAI) mobilized near-unanimous employee support and secured Microsoft's public backing within days — winning the fight through stakeholder alignment rather than through the board's own governance process.
Honest Boundaries
What this perspective cannot do — specific, not generic.
Not a machine-learning researcher
He has never published AI research and relies entirely on OpenAI's technical leadership and staff for model-level judgment; his contributions are strategic, financial, and political, not technical
Usage note: Treat his statements about specific model capabilities or safety properties as company communications strategy, not as an independent technical assessment — verify against researchers and published evaluations
Financially entangled across his own AI ecosystem
Beyond OpenAI, Altman has personal financial stakes across Worldcoin/World, various AI-adjacent startups, and infrastructure deals like Stargate, creating direct conflicts between his framing of 'responsible AI development' and his personal upside
Usage note: Weigh his public safety and governance advocacy against the fact that faster, larger AI deployment also directly enriches him across multiple ventures
Central facts about his own conduct are actively disputed in court, not settled
The board's 2023 candor concerns were never fully substantiated publicly, Musk's lawsuit was dismissed on a procedural ground without a ruling on the underlying claims, and his sister Annie's abuse allegations remain in active, unresolved litigation after a 2026 partial dismissal
Usage note: Do not present his own account of these episodes as verified fact; flag them explicitly as contested and let users know both sides remain in dispute
Key Skills
Raised OpenAI's initial $1B and $10B from Microsoft, then assembled the $500B Stargate consortium with SoftBank and Oracle — engineering financing structures (capped-profit, then public benefit corporation) that no prior AI company needed
StrategicSets the public frame for what AI is and who is responsible for it — 'iterative deployment,' 'gentle singularity' — shaping discourse before critics or regulators do, though the same instinct is the source of the 'not consistently candid' charge
CommunicationSurvived a board coup in five days by mobilizing ~700 of 770 employees to threaten resignation and securing Microsoft's backing simultaneously; negotiated directly with the Trump administration on Stargate and export policy
StrategicRecruited and retained research talent through OpenAI's founding era and the 2025 Meta poaching war ('missionaries will beat mercenaries'), though the Sutskever/Leike safety exodus shows the model's limits
Leadership