Altman, Sam

CEO of OpenAI

Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI

Lex Fridman Podcast · Mar 18, 2024

• Sam Altman
• OpenAI board crisis
• Ilya Sutskever
• Elon Musk lawsuit
• Sora video generation
• GPT-4 and GPT-5
• AI safety and governance
• AGI definition
• compute as currency
• power and control

Core ideas

Altman describes the November 2023 board crisis as "the most painful professional experience of my life" — chaotic, shameful, and upsetting, but paradoxically dominated by a feeling of love rather than hate, with one sharp exception: the moment Sunday evening when the board appointed an interim CEO, which he calls the low point of the whole weekend. He recounts going through rapid emotional stages — initial confusion and rage giving way, within hours, to genuine excitement about starting a focused new AGI research effort once he'd mentally accepted OpenAI as "dead" — before two board members called Saturday morning asking him to return. He frames the episode as revealing a structural flaw rather than pure interpersonal betrayal: a nonprofit board, unlike a normal corporate board, answers to no shareholders and effectively no one, and while the board legally had the authority to fire him, the fact that in practice it "didn't quite work" (because employees and stakeholders revolted) is itself, in his own words, "its own kind of governance failure." He states plainly that the experience made him less of a "default trusting person" — a shift he doesn't like — and describes the new board (including Bret Taylor and Larry Summers, chosen "on the battlefield" under time pressure that weekend) as prioritizing board experience and governance expertise that the prior nine-to-six-member board lacked. On Ilya Sutskever, he affirms deep respect and love, denies Ilya has "seen AGI," and frames Ilya's quiet withdrawal as consistent with a person who takes AI safety and societal impact with unusual long-term seriousness — while sharing a small human detail (Ilya being "silly," playing with a puppy at a dinner party) to counter Ilya's public reputation for gravity.

On the Elon Musk lawsuit, Altman characterizes OpenAI's early years as genuine improvisation ("fumbling through the dark") before language models or productization were even conceivable, with structural decisions patched repeatedly as circumstances changed, arriving eventually at something that "does look eyebrow-raising" in hindsight even though each step seemed reasonable at the time. He states Musk wanted total control to "turn OpenAI around" after predicting its failure, wanted Tesla to run its own AGI effort, and at various points proposed OpenAI be acquired by or merged into Tesla under his control — proposals OpenAI rejected, at which point (per Altman) "Elon chose to leave." Altman explicitly says he'd pick a different name than "OpenAI" if he had an oracle in hindsight, but defends the substance of "open" as meaning free/low-cost distribution of powerful tools to the public as a mission-driven public good, distinct from open-sourcing weights (which he supports selectively, e.g., for smaller local models, but not as an absolute policy). He calls the lawsuit "unbecoming of a builder," says it makes him and others sad given how much people "looked up to" Musk, and states he'd strongly discourage any other startup from adopting a nonprofit-to-capped-profit structure, denying OpenAI's structure sets a viable precedent.

On Sora, Altman argues these video-generation models understand more about the world than they're given credit for (citing successful object permanence through occlusion as one example) while cautioning against over-claiming a genuine 3D world model, and predicts the approach "will go surprisingly far" with scale despite persistent glitches (a cat sprouting an extra limb). He confirms substantial human labeling effort underlies the training pipeline alongside self-supervised learning on internet-scale data, and acknowledges deepfakes and misinformation as serious release-blocking concerns alongside the more mundane bottleneck of compute efficiency at scale. On copyright, he affirms that creators of "valuable data" deserve compensation and that artists should be able to opt out of style-mimicking generation, without offering a concrete mechanism yet, and is broadly unworried about human creative motivation surviving AI's rise (drawing the analogy to photography's disruption of painting) — while reframing the "AI will replace X% of jobs" question as instead "what percentage of tasks, across five-second to five-day time horizons, will AI perform," letting humans operate at higher levels of abstraction rather than being cleanly replaced.

On GPT-4 and iterative deployment, Altman self-deprecatingly calls GPT-4 "kind of suck[y]" relative to where the field is headed, predicting the GPT-4-to-GPT-5 gap will mirror the GPT-3-to-GPT-4 gap, and frames OpenAI's strategy of releasing numbered, incremental models (rather than building in secret to some final system) as deliberately designed to let the world adjust and build governance structures before being surprised — while admitting that despite this stated goal, users (including Lex) still perceive discrete "leaps," suggesting OpenAI may need to release even more incrementally going forward. He credits both the underlying base model and the post-training/RLHF "product" layer as equally critical to ChatGPT's breakthrough impact, distinguishes today's chat interface from a hypothetical future with much longer context windows (drawing an analogy to Bill Gates once being unable to imagine gigabyte-scale memory) that would let a model "get to know you" cumulatively across a lifetime of interactions and correspondence, and insists such memory should be entirely opt-in/opt-out at the user's discretion with real transparency, not the ad-driven data practices of past tech eras — a design choice he ties to OpenAI's aversion to advertising as a business model, since he doesn't want the company's incentives to require nudging users toward sponsors' products.

He declines to reveal details of the "Q*" reasoning project beyond confirming better reasoning is an important, unsolved research direction, and describes a "slower thinking" paradigm — allocating more compute to harder problems, akin to a system "talking to itself" — as plausible future architecture without committing to specifics. On compute economics, he reframes the viral "$7 trillion" claim (which he says he never personally tweeted, only memed about) into a substantive point: compute, unlike smartphone chips, isn't capacity-limited by a fixed population of buyers, since cheaper intelligence simply gets used more (reading your email all day vs. only curing cancer), making compute "the currency of the future" and "the most precious commodity in the world" — with energy (he specifically bets on nuclear fusion, name-checking Helion, and separately laments the stalled fission industry) as the central bottleneck. He is candid about existential-style personal risk from public backlash to AI ("I don't know what the percent chance is that I eventually get shot, but it's not zero") and about AI likely becoming a partisan left/right battleground, while arguing dramatic "theatrical" risks (that make good movie climaxes) get outsized attention compared to slow-burn harms (drawing the same comparison Lex raises: air pollution deaths vastly outnumber nuclear accident deaths, yet nuclear plants draw disproportionate fear).

On competition, Altman lists the standard pro-competition benefits (faster/cheaper innovation) against the risk of an unsafe "arms race," places himself in the "short timeline, slow takeoff" quadrant as the safest desired scenario, and voices open disappointment in Musk's conduct specifically (contrasting Musk's stated AI-safety concern with actions like the lawsuit and a widely-noted tweet mocking Jeff Bezos), while still calling him one of the great builders of the era and hoping for reconciliation. On Google/search, he dismisses "build a better Google" as an uninteresting goal, framing OpenAI's ambition instead as synthesizing or directly producing answers rather than ranking ten blue links, while confirming interest in combining LLMs with search "well" as an unsolved, exciting problem; he expresses a personal aesthetic distaste for advertising as a business model (preferring users pay directly so the product isn't shaped by advertiser incentives) while not ruling out some future unbiased ad mechanism. Addressing Google Gemini's "Black Nazis" controversy, he proposes that AI companies should publish detailed, concrete behavioral specifications (not vague high-level "principles," but explicit "when asked X, respond Y" rules with edge cases) so that bad outputs can be cleanly categorized as either bugs to fix or deliberate policies open to public debate — and states OpenAI has, in his view, been comparatively insulated from San Francisco's broader ideological "culture war" pressures, attributing this partly to the company's singular unifying focus on the AGI mission itself.

On safety more broadly, he predicts an increasing share of the entire company (not one isolated "safety team") will need to think about alignment, state-actor infiltration/model theft, societal and economic impacts, and technical robustness together, treating GPT-4's launch as already requiring company-wide coordination on these fronts. Pressed on AGI's definition and timeline, Altman rejects the term's usefulness as a single threshold, preferring to discuss specific capabilities, but concedes that by decade's end (or somewhat sooner) the world will likely have systems "we look at and say, wow, that's really remarkable" — while insisting ChatGPT/GPT-3.5, however historic, did not itself "change the world" in the way he thinks true AGI should (using the test: does the global economy feel materially different than before GPT-4 launched? He says no). For him, a meaningful AGI marker is a system that measurably accelerates the rate of scientific discovery, since he believes essentially all real economic growth stems from scientific and technological progress; asked what he'd ask a nascent AGI, he doubts the very first such system could answer grand questions (a unified theory of physics, alien life) at all, suspecting it might instead specify what new instruments or measurements are needed first. On power, he states unambiguously that he does not want, has never wanted, and does not currently hold super-voting control over OpenAI, framing the events of November 2023 as proof that even nominal board authority over him functioned imperfectly in practice — itself a governance failure worth taking seriously — and argues no single individual or company should control the path to AGI, favoring eventual government-level rules of the road even at the risk of being accused of seeking "regulatory capture." He states existential AI-takeover risk (an AI itself escaping human control) is not currently his top personal worry, calling excessive focus on "escaping the box" scenarios a partly "theatrical" distraction that has crowded out attention to other significant, more mundane AI risks, while affirming ongoing technical work on the problem continues regardless.

The conversation closes on lighter and more speculative territory: Altman explains his lowercase-only writing style as inherited from early instant-messenger culture and a general belief that capitalization conventions are historically fluid and somewhat arbitrary, while still using full capitalization for formal documents; on simulation theory, he says Sora's success modestly raises his credence in simulation-type hypotheses but isn't among the top handful of reasons to take the idea seriously, offering the "square root of negative one" as an example of how a simple, almost childlike operation can suddenly open onto a wholly new conceptual reality; on aliens, he says he "deeply want[s] to believe" intelligent civilizations exist and finds the Fermi paradox genuinely puzzling. He closes reflecting that what gives him hope for humanity is not any single genius but the cumulative societal "scaffolding" (language, science, institutions) that lets an individual today vastly outperform ancestors with barely any biological change — a fundamentally collective inheritance he finds hopeful for AGI's future too — and, asked about his own mortality (echoing his earlier remark about the nonzero chance of being shot over AI backlash), says he would mostly just feel grateful for the life he's had.

Excerpts