Sam Altman Expects the Road to AGI to Be a Power Struggle

On Friday afternoon, November 17, 2023, OpenAI's board dismissed Sam Altman. He calls it the most painful professional experience of his life — chaotic, shameful, upsetting — and also, in hindsight, a thing he is glad happened relatively early. Between founding the company and building AGI, he had assumed something crazy and explosive would occur. This one nearly destroyed the company. The road to AGI, he tells Lex Fridman, should be a giant power struggle. Not *should* as in ought, he corrects: he expects that to be the case.

He could not get much in the way of answers, but he accepted that the board gets to do this. After a short stretch of confusion and rage he started looking for the blessing in disguise. Running a decently sized company had never been the part he liked most; working with the researchers had. By Friday afternoon he was excited about a focused AGI research effort. It did not occur to him that any of it might be undone. Friday night the executive team said they were going to fight. He went to bed still excited. He barely slept or ate for four and a half days and still had a surprising amount of energy. Adrenaline, he learned, teaches you strange things.

Saturday morning two board members called: they had not meant to destabilize things, could they talk about him coming back? His first answer was no. Then he thought about the people, the partners, the shareholders, and named his conditions. The rest of the weekend was the most painful stretch — the team trying to hold OpenAI together while the world tried to pull it apart, everyone told they were almost done, just a little more time. Sunday evening the board appointed a new interim CEO. That was the low point. Other than that one moment, he would not describe his emotions as anger or hate. His phone blew up with messages from colleagues and from people he had not talked to in a decade. It felt like going to his own eulogy. The dominant emotion of the weekend, he says, was love.

For a month, maybe forty-five days after, he drifted through a fugue state, down, wanting to crawl into a cave while still having to run the company. One morning he woke up and chose: he could feel like a victim forever, or he could say this was the most important work he would ever touch and get back to it. He still wakes in the night thinking about it. He does not spend a lot of time on it now. Things are too busy. He does not think it was the last high-stress moment.

Boards are supposed to have a lot of power. In most companies they answer to shareholders. The OpenAI nonprofit board, unless you put other rules in place, answers to no one but itself. There are ways in which that is good. What they really wanted, Altman says, was a board that answers to the world as a whole, as much as that is practical. He thinks the members were far-and-well-meaning people who made suboptimal decisions under time pressure. The company will need a board and a team that can operate under that pressure.

The old board shrank over about a year from nine people to six, then could not agree on whom to add. It also lacked experienced board members. Larry Summers and Bret Taylor were chosen in the heat of that weekend, a battlefield decision with no time to design a rigorous process. Summers was one of the old board's own suggestions. Taylor had been approached earlier, been busy, and then was needed. Altman did not think he could work with the old board in the same configuration; he is grateful Adam stayed. They aimed for a board of three in a short window. The current board is not final. They will add more.

Going forward they have criteria, and they hire in slates, not one at a time — a Bret Taylor line Altman likes. A board needs nonprofit expertise, people who have run companies, legal and governance skill. Technical savvy matters for some seats, not every seat. Fridman asks whether they look at track record or just talk. For most executive hires Altman ignores the y-intercept and watches slope. For a board member he cares much more about the intercept. Experience is hard to replace.

People saw Mira Murati in the chaos of that weekend, which is how leaders usually get judged. What Altman says he values is how someone acts on a boring Tuesday at 9:46 in the morning: how they show up in a meeting, the quality of the decisions in the day's ordinary drudgery. The dramatic weekend is what everyone wants to talk about. OpenAI, he says, is really about the other seven years.

Is Ilya Sutskever being held in a secret nuclear facility? No. A regular secret facility? No. A nuclear non-secure facility? Neither. Altman loves Ilya, has tremendous respect for him, and hopes they work together for the rest of his career. He will not speak to Ilya's plans. The meme that Ilya saw AGI, and that the seeing is what drove the board drama, gets a flat no. Nobody has seen AGI. They have not built it. What Altman does say is that Ilya takes AGI and its social impact seriously, and that over the last couple of years he is one of the people Altman has spent the most time with on what this will mean and how to get it right. Ilya did not see AGI. In how much he thinks and worries about getting it right, Altman calls him a credit to humanity.

He will not explain the silence. He thinks of Ilya as always on a soul search, in a good way. At a recent dinner Ilya was playing with a puppy, very silly, very endearing — not the side the world sees most. Altman feels great about the new board. One of its jobs is to look at OpenAI's structure and make it more robust. He has no super-deep theory of the weekend beyond a perfect storm of weirdness, a preview of higher stakes, and a shockingly painful lesson in governance.

Did it make him more hesitant to trust people? Yes. His philosophy had been not to worry about paranoia and edge cases: you get a little screwed in exchange for living with your guard down. This caught him so off guard that it changed how he thinks about default trust and planning for bad scenarios. He does not like the change. He is not worried about becoming cynical; he thinks he is the extreme opposite. He is worried about becoming less of a default-trusting person. Fridman is not sure which mode is best for someone building AGI. On the team he works with every day, Altman would give himself a very good grade: enormous gratitude, trust, and respect.

Altman says he does not know what Elon Musk's lawsuit is really about. He can tell the history. OpenAI started as a research lab with no idea where the technology would go — before language models were a big deal, before an API, before selling chatbot access, before any plan to productize. With fundamentally new things you fumble through the dark and most of the assumptions turn out wrong. Then it became clear they would need to do different things and raise huge amounts more capital. They patched the structure, patched it again, and ended up with something that looks, at least, eyebrow-raising. With an oracle he would do it totally differently. He did not have the oracle at the time.

Musk, in Altman's telling, thought OpenAI was going to fail and wanted total control to turn it around. He wanted Tesla to be able to build an AGI effort. At various times he wanted OpenAI made into a for-profit he could control, or merged with Tesla. Altman's memory of the proposal is an acquisition with Tesla in full control. They did not want that. Musk chose to leave, which is fine. If Altman could go back with an oracle he would pick a different name. What "open" means to him now is putting powerful technology in people's hands for free as a public good: no ads on the free version, no other monetization, increasingly capable tools as part of the mission. Open source some things and not others. That has become a religious battle line where nuance is hard. He thinks nuance is the right answer.

Musk said change the name to ClosedAI and he would drop the lawsuit. Altman reads that as a comment on how seriously Musk means the suit — an astonishing thing to say. Grok, he notes, had not open-sourced anything until people pointed out the hypocrisy, then announced an open-source dump. He does not think open versus closed is what this is really about. The whole thing, he says, is unbecoming of a builder, and he respects Musk as one of the great builders of our time, which makes the tossing extra sad. He said in an interview that he missed the old Elon and got a flood of messages saying that exactly encapsulated how they felt. He hopes years from now they have an amicable relationship. For other startups he is blunt: he would heavily discourage anyone thinking of starting as a nonprofit and adding a for-profit arm later.

Fridman asks whether Sora understands the world more or less than GPT-4. Altman thinks all these models understand more of a world model than critics admit, and that the failures make it easy to call the whole thing fake. Watching Sora, he was impressed when a person walked in front of something, occluded it for a few seconds, and walked away with the object still there. Some sequences hold physics over many steps. He will not go as far as Fridman: it is doing something that deals with occlusions really well. Claiming a great underlying 3D model of the world is more of a stretch. The 2D-patch approach looks like it will go surprisingly far. He does not want to speculate about which limits it will surmount.

Cats still sprout an extra limb at random points in a video. Is that a fundamental flaw or a scale problem? Yes to both: something about the approach feels different from how we think and learn, and it will also get better with scale. Training uses visual patches the way language models use tokens. Without specifics on Sora, they use lots of human data. How many humans? More than three people work on labeling. Fundamentally there is a lot of self-supervised learning; the technical report's phrase is internet-scale data. Before release they still have to get efficiency to the scale people will want. Deepfakes and misinformation do not take much thought.

On copyright, the question behind the question is whether people who create valuable data deserve some compensation. He thinks the answer is yes. If he were an artist he would want to opt out of people generating art in his style, and if they did it anyway he would want an economic model attached. Artists panicked at photography too; photography became a new art form and people made money taking pictures. He does not think we have seen the last evolution of how the economic system works. The better frame for jobs is not what percent of current roles AI will replace, but what percent of tasks — five-second, five-minute, five-hour, even five-day — it can do, over what time horizon. That is a tool that lets people work at a higher level of abstraction. Most YouTube videos may use AI in production and still be driven by a person. Humans, he agrees, care about other humans a lot. We still run races. Cars are faster.

Fridman calls GPT-4 historically impressive. Altman says he thinks it kind of sucks. GPT-3 looked like a marvel of technology until GPT-4 made it unimaginably horrible. He expects the delta between five and four to be the same as between four and three. The job is to live a few years in the future and remember that the tools we have now will look bad looking backwards. That is how you make the future better. The glimmer he likes most is a creative brainstorming partner: a name, a different angle, a thing he is not sure how to do. Another small glimpse is longer-horizon work — break a problem into steps, search, write code, put it together. When that works, which is not very often, it is magical. Ten-step problems on their own still fail a lot.

ChatGPT was the moment a lot of the world went from not believing to believing, and that was more about the interface, the product, and the post-training than the base model. RLHF and the rest of the wrapper are a small compute layer on top and a huge amount of work. They had to invent the underlying technology and figure out how to make it a product people would love — alignment and usefulness, not just shipping. Scaling was a known hard problem other companies have had to solve. The two things that had never been done before were the achievements he wants counted.

They deploy iteratively — talk about GPT-1, 2, 3, and 4 rather than hide until GPT-5 — because AI and surprise do not go together, and because the world needs time to adapt. The goal is not shock updates; the opposite. The fact that people still feel leaps makes him think they may be missing the mark and should consider releasing GPT-5 differently. When is it coming? He does not know. That is an honest answer. They will release an amazing model this year. He does not know what they will call it. Before a GPT-5-like model, named that or not, they have a lot of other important things to ship. The unlock is never one thing. An original Ilya line he butchers: they multiply 200 medium-sized things together into one giant thing.

Most people do not need 128K of context most of the time. In a distant future he imagines context lengths of several billion: feed in all your information and history, and the system gets to know you better and better. He recalls a clip of Bill Gates talking about early computer memory — 64K or 640K, most of it the screen buffer — unable to imagine gigabytes, let alone terabytes. You always find out how to use the exponential. Once you have it, you will not want to go back. The most interesting GPT-4 use case he sees is not a single trick but younger people who start every knowledge-work task there: code, search, vision, editing a paper. Hallucinations remain an intense research problem. He thinks users are more sophisticated than critics allow, and that they know to check when it is mission-critical. He does not put journalists' lazy GPT use at the top of his long list of complaints about journalists.

Memory in ChatGPT is an early exploration. What he wants is a model that gets to know him and gets more useful over a life, across many models. Not just factoids and preferences. Fridman teases him about remembering November. Altman is serious: he does not only want it to remember the weekend. He wants it to integrate the lessons and remind him in the future what to do differently or what to watch out for. If context ever feels like trillions of tokens, every conversation and every email in the window every time you ask a question would be pretty cool. Privacy, for him, is user choice: anything he wants stricken from his AI agent's record, he wants out; if he wants it to remember nothing, that too. Different people will pick different points on the privacy-utility tradeoff. The answer is easy user choice, with more transparency than companies have historically offered when they collected data for ads.

Each token still gets about the same compute. Fridman asks about slower, sequential thinking. Altman thinks there will be a new paradigm for that. Architecturally it might be a layer on top of LLMs, or many other implementations; the spiritual point is that an AI should think harder about a harder problem and answer quickly on an easy one. A proof of Fermat's Last Theorem should take more compute than today's date, unless the proof is already memorized. Whether that looks like an LLM talking to itself, they do not know. Q* sits in the same lore bin as the nuclear facility. They are not ready to talk about it. They have said for a while that better reasoning is an important direction. They have not cracked the code.

He never tweeted that OpenAI was raising $7 trillion. Once the misinformation was out he memed it. Underneath the joke is a serious claim: compute will be the currency of the future, maybe the most precious commodity in the world, and the world should invest heavily to make a lot more of it. Phone chips do not work this way. Eight billion people, six or seven billion phones, upgrade every two years, and the market is about three billion SoCs a year. Make thirty billion and you will not sell ten times as many phones. Intelligence is more like energy. The only numbers that make sense are: at price X the world uses this much compute, at price Y that much. Cheap, he would have it reading his email all day, offering suggestions, and trying to cure cancer. Expensive, maybe only cancer. The world is going to want an amount of compute that is hard to reason about right now.

Energy is the hardest part, then data centers, supply chain, and fabricating enough chips. He believes the energy answer is nuclear, including fusion. He thinks Helion is doing the best work and is happy there is a race. Fission is also amazing; it is sad how that industry ground to a halt, which he puts down to mass hysteria. He worries about the same wiring for AI. Some things are going to go theatrically wrong. He does not know the percent chance he eventually gets shot, but it is not zero. Air pollution has killed far more people than nuclear reactors; most people still worry more about living next to a reactor than a coal plant. Risks that make a good movie climax carry more weight than slow burns. He thinks AI will have tremendously more good consequences than bad ones, and it will have bad ones, some of them not theatrical at all.

Competition with Google, Meta, xAI, and others has an obvious pro: better products, faster, cheaper. The con is an arms race he is nervous about. He feels that pressure in some ways. If you draw a quadrant of short versus long timelines and slow versus fast takeoff, he wants short timelines and a slow takeoff — the safest quadrant — and he wants to make sure they get the slow takeoff. Musk says he cares a great deal about AI safety; Altman assumes he is not going to race unsafely. Collaboration on safety would help. It is not what Musk is most known for. Someone reminded Altman that the day Musk passed Jeff Bezos as richest, he tweeted a silver medal at him. Altman hopes for less of that as people work toward AI. He thinks the world needs Musk to step up as a leader through this next phase.

Google has owned the world's access to information for twenty years. Will OpenAI take that on? If the question is building a better search engine, Altman finds it boring. Ten blue links and thirteen ads are one way to find information. What excites him is a much better way to help people find, act on, and synthesize it. ChatGPT is already that for some use cases. The world does not need another copy of Google. The intersection of LLMs and search, nobody has cracked. He would love to go do that. Ads he kind of hates as an aesthetic choice. They had to happen to get the internet going; it is a more mature industry now. He likes that people pay for ChatGPT and know the answers are not influenced by advertisers. Dystopia is easy to imagine: you ask ChatGPT something and it tells you which product to buy. Ad-supported products get worse, not better, in a world with AI. He thinks they can have a great business that pays for compute without ads.

Gemini 1.5 generated Black Nazis and Black founding fathers — theatrical, and a warning about a human safety layer importing a company's ideology. OpenAI has made its own mistakes and will make others. Someone inside had an idea he likes: write out the desired behavior of a model, make it public, take input, explain the edge cases. Then, when the model misbehaves, you can tell whether it is a bug to fix or a policy to debate. High-level AI principles do not work. You have to say: when you ask it to do X, it is supposed to respond in way Y, including concrete cases like Trump versus Biden. San Francisco is an ideological bubble. He feels lucky that OpenAI does not have the culture-war problems he hears about at other companies. Every company has an ideology; theirs is about AGI, and it pushes some others out. Subtle infiltration, sure. Not the obvious kind.

Safety, broadly defined, will eventually be what most of the company thinks about: technical alignment, societal and economic impact, state actors trying to steal the model. It will not be one safety team. They are trying. He will not go into accents or further details. As GPT-5 gets smarter, the special thing is that it gets better across the board, not better in one area and worse in others. Programming in five or ten years will still be a lot of programming, in a very different shape; some people may work entirely in natural language, the way almost no one uses punch cards. The last one percent of the craft will still want multiple tools, including writing C when needed. Humanoid robots: at some point. It would be depressing if AGI arrived and the only way to get things done in the physical world was to make a human go do it. OpenAI will return to robots in some way. They will not turn themselves into robots.

When will humanity build AGI? He used to love speculating. He now thinks the question is poorly formed because everyone uses a different definition. It is more useful to talk about systems that can do X or Y than about fuzzily crossing one mile marker. AGI is closer to a beginning than an ending. By the end of this decade, possibly somewhat sooner, he expects quite capable systems that we look at and say, wow, that is really remarkable. GPT-3.5 would have looked like AGI to many experts in the 1990s. It did not change the world as an artifact. It changed expectations, which matters. His own bar includes whether the global economy feels different. After GPT-4, for him, it does not. A personal marker: a system that significantly increases the rate of scientific discovery, because he believes most real economic growth comes from scientific and technological progress.

If someone handed him a laptop and said this is the AGI, he finds it surprisingly hard to know what to ask. He does not think the first one will explain a grand unified theory of physics. If it could answer yes or no, those would be the first questions: does such a theory exist? Are there other alien civilizations? He does not actually expect it could answer even those. A better scene is the system saying you want this physics answer, so build this machine, make these five measurements, and bring me the data. Whoever builds AGI first gets a lot of power. Does he trust himself with that much power? He still believes that neither he nor any other one person should have total control over OpenAI or over AGI. You want a robust governance system.

He can point to the board drama: at first he did not fight it, it was the will of the board even though he thought it was a really bad decision; later he clearly did fight, and he can explain why that was okay. As many people observed, the board had the legal ability to fire him and in practice it did not quite work. That is its own kind of governance failure. It makes it harder to look Fridman in the eye and say the board can just fire him. He continues not to want super-voting control. He never had it, never wanted it, and after all the craziness still does not want it. No company should be making these decisions alone; governments need to put rules of the road in place. He knows Marc Andreessen and others will call that regulatory capture. He is willing to be misunderstood. Losing control of the AGI itself — the model escaping the box — is not his top worry as he currently sees things. It has been at times; it may be again. It is a theatrical risk that took over a lot of well-meaning safety research, he argues, without much progress, and pushed other significant AI risks out of the discourse. They still have to work on it super hard.

Fridman spends a surprising amount of time on why Altman tweets without capital letters. The shift key works. He grew up a very online kid, chatting on instant messenger, and never capitalized there, as he thinks most internet kids did not. Old English capitalized random nouns; he finds the leading capital a bit of a dumb construct. Strategy memos and formal long messages still get proper case. Short-form has drifted toward how he texts friends. If it felt like a sign of respect he could force himself. He does not think about it. Fridman's punchline: he is the only CEO of a company who does not capitalize tweets. Altman doubts even that is true.

Sora raised his belief that we live in a simulation somewhat, not as one of his top five pieces of evidence. He was already certain we would be able to generate worlds; it happened faster than he thought. A friend recently offered what he found a profound, psychedelic insight: the square-root function. Square root of four is easy; square root of two needs a new kind of number; square root of negative one tips you into another reality. A child-simple operator opens a realm. He thinks there are a lot of those operators for why people find simulation hypotheses more likely than before. For him, Sora working is not in the top five. AI at its best will be those kinds of simple gateways to another way of seeing.

Does he think there are intelligent alien civilizations? He deeply wants the answer to be yes and finds the Fermi paradox very puzzling, and scary — intelligence may not be good at handling powerful technologies. What gives him hope is the past: huge problems, deep flaws, a lot to be ashamed of, and on the whole inspiring. He wonders whether AGI will look more like a single brain or like the scaffolding between all of us. You have not had much genetic drift from your great-great-great-grandparents, and what you can do is dramatically different. That is not biology. It is the layer we all contributed. No one person builds the iPhone. No one person discovers all of science. You still get to use it. That was a collective thing, and it fills him with hope. If he got shot tomorrow he would be sad to miss what happens next, in a curious time, and mostly grateful for his life.