What this covers
This conversation between Sam Altman and Lex Fridman covers the economics, governance, and trajectory of artificial intelligence development, with Altman presenting a framework for understanding both the technical path forward and the institutional arrangements required to navigate it responsibly. The discussion moves across three interconnected claims: that power concentration in AGI development demands distributed governance structures answerable to broad constituencies rather than single actors; that compute will function like energy—a commodity with elastic demand whose scarcity and allocation will define economic hierarchy; and that capability improvements follow an Exponential Curve, making today's systems appear primitive within years.
Altman spends significant time on how human psychology distorts AI risk perception, arguing that Theatrical Risk Bias causes people to fear dramatic escape scenarios while underweighting slower harms—much as they fear nuclear reactors more than coal plants despite the latter's documented deaths. He contests the framing of jobs displacement, proposing instead a Tasks vs Jobs Framework that asks what granular capabilities AI can perform across different time horizons. On governance specifically, Altman unpacks the structural failures revealed by OpenAI's board crisis, arguing that nonprofit boards hold immense power while answering to no one, and that Balance of Power systems require democratic answerability. The conversation also addresses Elastic Demand for Compute, world models embedded in generative systems, product design's role in ChatGPT's adoption, memory systems that grow with individual users, compensation mechanisms for artists whose work trains models, and whether human creativity persists when machines excel at the same tasks. Throughout, Altman emphasizes that progress is continuous rather than punctuated, and that governance frameworks must be built now for a world that will arrive faster than institutions typically adapt.
Altman argues that the path to AGI will inevitably be a high-stakes power struggle requiring robust governance no single person controls, that compute will become the world's most precious commodity, and that progress is continuous and exponential rather than a series of discrete leaps.
- AGI development concentrates power, so governance must distribute it and no one person should control AGI
- Compute behaves like energy with elastic demand and will be the future's defining commodity
- Capability gains are continuous and exponential, so today's models will look poor in retrospect
Robust governance with distributed power and external oversight is essential; no single person or company should control AGI decisions alone.
- No single person should have total control over OpenAI or over AGI; a robust governance system with balance of power is needed, and ultimately governments must put rules of the road in place because no company should be making these decisions alone. Notably, although the board legally had the ability to fire him, in practice it did not work, which is itself a governance failure.
“it is important that I nor any other one person have total control over OpenAI or over AGI. And I think you want a robust governance system.”
- The structural flaw exposed by the board crisis is that, unlike typical corporate boards answerable to shareholders, the board of a nonprofit has very large power and answers to no one but themselves unless other rules are put in place, whereas what OpenAI really wants is for its board to answer to the world as a whole.
“the board of a nonprofit has, unless you put other rules in place, quite a lot of power. They don’t really answer to anyone but themselves.”
- Board members should be hired in slates rather than as individuals one at a time, because while an executive needs to do one role well, a board must collectively cover a whole range of governance competencies—nonprofit expertise, company-running experience, and legal and governance expertise—which is best optimized for as a group.
“we want to hire board members in slates, not as individuals one at a time.”
- The road to AGI will be a giant power struggle, and the board crisis was a preview of escalating high-stress conflicts that will intensify as the stakes get higher, which is why building a resilient organization and robust governance structures matters.
“The road to AGI should be a giant power struggle. The world should… Well, not should. I expect that to be the case.”
- For some roles, track record (the Y-intercept) should be ignored and one should look only at slope—rate of growth and trajectory—whereas for a board member the Y-intercept matters much more because experience is very hard to replace.
“there are some roles where I totally ignore track record and just look at slope, ignore the Y-intercept.”
AI capabilities advance exponentially, making today's tools look primitive in retrospect and requiring teams to think years ahead.
- Progress is on an exponential curve, so just as GPT-3 now looks 'unimaginably horrible' compared to GPT-4, the delta from GPT-4 to GPT-5 will be similar to the delta from 3 to 4, and it is the team's job to live a few years in the future and remember today's tools will look bad in retrospect, which is how they ensure the future is better.
“now we have GPT-4 and look at GPT-3 and you’re like, “That’s unimaginably horrible.” I expect that the delta between 5 and 4 will be the same as between 4 and 3”
- There is no single big unlock behind GPT progress; rather, as an Ilya Sutskever quote puts it, OpenAI 'multiplies 200 medium-sized things together into one giant thing,' with distributed constant innovation across teams that come together in surprising ways, requiring people to keep the whole picture in their heads even while working in the weeds.
“We multiply 200 medium-sized things together into one giant thing.”
- You should always just follow the exponential of technology and trust that we will find ways to use better technology, illustrated by Bill Gates being unable to imagine needing gigabytes of memory when early computers had kilobytes—context windows that seem absurdly large today (billions, then trillions of tokens) will be used and we won't want to go back once we have them.
“you always do just need to follow the exponential of technology and we will find out how to use better technology.”
- By the end of this decade, and possibly somewhat sooner, we will have quite capable systems that, if we could look at them now, we would say are really remarkable—though by the time we arrive we may have adjusted our expectations.
“I expect that by the end of this decade and possibly somewhat sooner than that, we will have quite capable systems that we look at and say, “Wow, that’s really remarkable.””
OpenAI distributes increasingly powerful AI as a public good through free access, embodying a meaningful form of openness distinct from open-sourcing code.
- OpenAI deploys iteratively—talking about GPT-1, 2, 3, 4 rather than building in secret to GPT-5—because AI and surprise don't go together: the world's people and institutions need time to adapt and to put governance in place before being under the gun, and the goal is explicitly to avoid shock updates to the world.
“part of the reason there is I think AI and surprise don’t go together. And also the world, people, institutions, whatever you want to call it, need time to adapt and think about these things.”
- ChatGPT was the moment much of the world went from not believing to believing in AI, but this was driven less by the underlying base model than by the interface and product—the post-training RLHF step that tunes the model to be helpful—which, despite being a comparatively small 'wrapper' of compute on top of the base model, was a hugely important and largely novel achievement.
“That was more about the ChatGPT interface. And by the interface and product, I also mean the post training of the model and how we tune it to be helpful to you”
- One of the most important things OpenAI does is put increasingly powerful technology in the hands of people for free as a public good, without ads or other monetization of the free version, because giving people great tools lets them build an incredible future for each other—this is the meaningful sense of 'open' for the mission, distinct from open-sourcing the code.
“putting powerful technology in the hands of people for free, as a public good. We don’t run ads on our free version.”
- Building a better copy of Google search (ranked blue links plus ads) would massively understate what's possible; the exciting opportunity is a fundamentally better way to help people find, act on, and synthesize information—creating answers in some cases, synthesizing in others, pointing to sources in yet others—which ChatGPT already does for some use cases.
“the thing that’s exciting to me is not that we can go build a better copy of Google search, but that maybe there’s just some much better way to help people find and act on and synthesize information.”
Humans will remain economically and creatively valuable even as AI surpasses them at individual tasks because creating and achieving status are hardwired.
- Even when AI surpasses humans at a task, people will still fundamentally care about other humans—as seen in chess (humans still play despite AI dominance) and running races (despite cars being faster)—so humans will keep creating and society will find ways to reward it, because the drive to create, be useful, and achieve status is hardwired.
“humans are going to do cool shit and society is going to find some way to reward it. That seems pretty hardwired.”
- The right framework for AI's economic impact is not what percentage of current jobs AI will replace, but what percentage of tasks across different time horizons (five-second, five-minute, five-hour, five-day tasks) AI can do, because AI is a tool that lets people operate at higher levels of abstraction; at some point this shifts from a quantitative efficiency change to a qualitative change in the kinds of problems people can hold in their head.
“The way I think about it is not what percent of jobs AI will do, but what percent of tasks will AI do on over one time horizon.”
- The most promising glimpse of GPT-4's future value is not coding or translation but its use as a creative brainstorming partner—helping name things, think about problems in different ways—and as something that can break a problem into multiple steps and execute some of them (search the internet, write code, put it together); when that longer-horizon capability works, which is not very often, it is magical.
“there’s something about the kind of creative brainstorming partner... that I think gives a glimpse of something I hope to see more of.”
- A person's capability and knowledge today is dramatically different from their great-great-great grandparents' not because of biological or genetic change but because of the collective scaffolding humanity built—no one person could build the iPhone or discover all of science, yet each individual gets to use it—and this raises the open question of whether AGI will be like a single brain or like that connective scaffolding between all of us.
“what you have is this scaffolding that we all contributed to built on top of. No one person is going to go build the iPhone. No one person is going to go discover all of science, and yet you get to use it.”
Compute will become the world's most precious commodity because demand for AI intelligence is elastic and grows with capability.
- Compute will become the currency of the future and possibly the most precious commodity in the world, because intelligence behaves like energy rather than like discrete consumer goods: demand is elastic to price, so at low prices the world will use vast amounts of compute for everyday tasks and at high prices only for the most valuable problems, meaning total demand will be enormous and hard to reason about.
“I think compute is going to be the currency of the future. I think it’ll be maybe the most precious commodity in the world.”
- There will be a new paradigm enabling slower, sequential 'thinking,' where an AI can think harder and allocate more compute to a harder problem while answering an easier problem more quickly—rather than spending the same amount of compute per token—analogous to human thought, and proving something like Fermat's Last Theorem would require more compute than answering today's date.
“I think there will be a new paradigm for that kind of thinking.”
Companies must transparently specify model behavior and take public input on safety and bias tradeoffs rather than operating ambiguously.
- To handle bias and safety disputes, companies should write out the desired behavior of a model—concretely specifying how it should respond to specific prompts, including edge cases—make it public, and take input on it, so that when a model misbehaves it is clear whether that is a bug to fix or intended behavior to debate at the policy level, rather than being caught ambiguously in between.
“it would be nice to write out what the desired behavior of a model is, make that public, take input on it, say, “Here’s how this model’s supposed to behave,” and explain the edge cases too.”
- The goal for AI memory is not merely remembering factoids and preferences but a model that gets to know you and grows more useful over your life, integrating the lessons of your experiences and reminding you what to do differently—and the right answer to the resulting privacy/utility tradeoff is simple user choice, with high transparency, letting users strike anything from the record.
“I want it to integrate the lessons of that and remind me in the future what to do differently or what to watch out for.”
- These models hallucinate some of the time, and people are more sophisticated users of technology than they're given credit for—they understand that if something is mission-critical they must check it—and while grounding in truth is an area of intense interest that will improve with upcoming versions, it won't be fully solved soon.
“people seem to really understand that GPT, any of these models hallucinate some of the time. And if it’s mission-critical, you got to check it.”
- A paid, ad-free business model is preferable for AI because users know the answers they get are not influenced by advertisers; while an unbiased ad unit for LLMs may be possible, ads create dystopic incentives (e.g., ChatGPT nudging product purchases) that get worse, not better, in a world with AI, and Altman believes OpenAI can build a great business covering its compute needs without ads.
“I like that people pay for ChatGPT and know that the answers they’re getting are not influenced by advertisers.”
The board crisis exemplified how high-stress escalating conflicts will define the path to AGI and how governance structures must be resilient.
- During the four and a half day board crisis, Altman didn't sleep or eat much yet had a surprising amount of energy, illustrating a weird thing one learns about adrenaline in wartime-like high-stress situations.
“there was this period of four and a half days where I didn’t sleep much, didn’t eat much, and still had a surprising amount of energy. You learn a weird thing about adrenaline in wartime.”
- Elon Musk thought OpenAI was going to fail and wanted total control to turn it around, including at various times wanting to make it a for-profit he could control or merge it with Tesla under Tesla's full control; OpenAI declined, and Elon decided to leave.
“He also wanted Tesla to be able to build an AGI effort. At various times, he wanted to make OpenAI into a for-profit company that he could have control of or have it merge with Tesla.”
- Any startup considering starting as a nonprofit and adding a for-profit arm later should be heavily discouraged from doing so; OpenAI would have structured itself differently from the start had it known what would happen, and laws make exploiting this path for tax savings difficult, so it won't set a precedent.
“I would heavily discourage any startup that was thinking about starting as a nonprofit and adding a for-profit arm later. I’d heavily discourage them from doing that.”
- The truest measure of a leader is not how they act in dramatic crisis moments but how they show up in the normal day-to-day drudgery—the quality of decisions they make on a boring Tuesday at 9:46 in the morning—because most of the work and value is created meeting by meeting.
“a thing I really value in leaders is how people act on a boring Tuesday at 9:46 in the morning and in just the normal drudgery of the day-to-day.”
- Altman has lived by a philosophy of high default trust—not worrying about edge cases or paranoia, accepting getting occasionally screwed in exchange for living with his guard down—but the board crisis so caught him off guard that it changed how he thinks about default trust and planning for bad scenarios, a change he dislikes.
“I’ve always had a life philosophy of don’t worry about all of the paranoia. Don’t worry about the edge cases. You get a little bit screwed in exchange for getting to live with your guard down.”
- Generative models like Sora understand more about a world model than people give them credit for—evidenced by behaviors like correctly maintaining an occluded object after a person walks in front of and away from it—even though obvious failures (like a cat sprouting an extra limb) make it easy to dismiss them as fake; the truth is some of it works and some doesn't, and it will improve with scale.
“all of these models understand something more about the world model than most of us give them credit for.”
- OpenAI is not really about the one dramatic board weekend but about the other seven years of work.
“that’s not really what OpenAI is about. OpenAI is really about the other seven years.”
People misperceive AI risks by overweighting dramatic scenarios while underweighting diffuse harms unfolding over time.
- Humans are wired to weight risks that make a good climax scene of a movie far more heavily than risks that are very bad but unfold slowly over a long period—just as people fear living next to a nuclear reactor more than a coal plant despite air pollution killing far more people—and this same theatrical bias distorts how people perceive AI risk, crowding out attention to less dramatic but significant harms.
“the ones that make a good climax scene of a movie carry much more weight with us than the ones that are very bad over a long period of time but on a slow burn.”
- Loss of control of the AI itself is not Altman's top current worry; the dramatic 'escape the box' scenario is a theatrical risk that a group of well-meaning AI safety researchers got super-hung up on—arguably without much progress—and which pushed many other very significant AI-related risks out of the space of discourse.
“there’s a big group of very smart, I think very well-meaning AI safety researchers that got super-hung up on this one problem... but super-hung up on this one problem.”
- Among the quadrants formed by timeline-to-AGI (short vs long) and takeoff speed (fast vs slow), the short-timeline, slow-takeoff quadrant is the safest and the one Altman would most like to be in, and he wants to ensure the takeoff is slow.
“I think short timeline, slow takeoff is the safest quadrant and the one I’d most like us to be in. But I do want to make sure we get that slow takeoff.”
AI systems will advance toward meaningful capability thresholds like accelerating scientific discovery rather than a singular defined AGI moment.
- Some simple-sounding operators can tip you into a whole new realm of knowledge: the square root function is explainable to a child, but asking for the square root of negative one forces a whole new kind of number and a different reality; many such 'psychedelic' gateway insights exist, and AI may serve as that kind of simple gateway to new realms of understanding.
“once I come up with this easy idea of a square root function... then you can ask the question of “What is the square root of negative one?” And this is why it’s a psychedelic thing. That tips you into some whole other kind of reality.”
- The question 'when will we build AGI' is poorly formed because people use wildly different definitions; it is more useful to ask when systems can do specific capabilities X, Y, or Z. AGI is not an ending but closer to a beginning—a mile marker—and to Altman a key threshold is when a system can significantly increase the rate of scientific discovery, since most real economic growth comes from scientific and technological progress.
“I think it makes more sense to talk about when we’ll build systems that can do capability X or Y or Z, rather than when we fuzzily cross this one mile marker. AGI is also not an ending. It’s closer to a beginning”
Creators of valuable data deserve compensation mechanisms and control over their work's use in AI training.
- There is a real place for open source models, particularly smaller models that people can run locally where there is huge demand, and the ecosystem will end up with a mix of some open and some closed models, not unlike other technology ecosystems.
“there is definitely a place for open source models, particularly smaller models that people can run locally, I think there’s huge demand for.”
- People who create valuable data deserve some way to be compensated for its use in training AI; specifically artists should be able to opt out of having art generated in their style, and if their style is used anyway there should be an economic model that pays them, even though the exact mechanism (analogous to the CD-to-Napster-to-Spotify transition) is not yet known.
“do people who create valuable data deserve to have some way that they get compensated for use of it, and that I think the answer is yes.”