Over the past year, OpenAI has navigated what Altman himself calls a “pretty tough” period. In a recent episode of the podcast “Invest Like The Best,” he offered a rare, systematic review, addressing strategic missteps, competitive pressures, safety scares, and his latest views on the timeline for AGI and the future of robotics.
The Hardest Year: “We Were Too Spread Out, That’s on Me”
“The past year was quite difficult, partly my fault,” Altman stated directly during the show.
He traced the problem back to a single cause: distraction. In early 2025, a core anxiety gripped OpenAI: could revenue growth keep pace after a massive procurement of computing power? To hedge this risk, the company began simultaneously developing consumer apps, media content, and other business lines. The logic was, “If revenue growth is slower than expected, these businesses can help us put our GPUs to use.”
“Looking back now, it sounds absurd, because the industry’s revenue growth was steeper than anyone anticipated,” Altman said.
Once the team realized that the model’s improvement trajectory was clear enough and the economic returns were certain enough, OpenAI made a series of “tough decisions” to dramatically narrow its focus—returning to its core: providing the best, most abundant, and most cost-effective AI intelligence, and empowering external developers to build products on top of it.
“We don’t want to eat every startup, we don’t want to eat every company,” he stated. “What we truly want to do is provide that platform… sell AI. Make the best, most abundant, most cost-effective intelligence, and let the world build amazing things on that foundation.”
Distillation Doesn’t Make My Top Ten Concerns
During the interview, the host raised a hot-button issue: how can OpenAI continue to recoup its training costs when competitors use model distillation to train cheaper models using OpenAI’s own outputs?
Altman’s response was surprisingly calm.
“I haven’t thought deeply about the distillation problem,” he said. “I would certainly prefer others didn’t do it.” But then he pivoted: “It’s not in my top ten concerns.”
His logic is that OpenAI’s inference business is large enough that even with moderate profit margins, “earning a decent profit on trillions of dollars in revenue would be enough to cover the cost of training large models.” The key is the ratio of inference costs to training costs. While training is expensive, the revenue from serving clients’ inference needs will vastly exceed the training expenses.
He admitted, “I might be too confident right now about our progress and the models we’re about to release.” But his conclusion remains unchanged—the real flywheel is on the inference side, not in the exclusivity of the training side.
Altman also added a broader perspective: he has always assumed that good, cheap models would exist in the world. “What matters is that we are the best and the cheapest. What others do is their business. We just need to win on our own turf.”
The Bet on Compute: From “You’re Crazy” to Proving We Didn’t Bet Big Enough
OpenAI’s massive bet on computing power was once ridiculed as reckless. Altman recalled that when they first approached cloud providers, chip manufacturers, and energy suppliers, the response was almost universally: “You’re completely insane. No industry works this way.”
He compared the experience to early-stage startup fundraising—most people say no, but you only need one or two “yes” responses. Microsoft was the first to say “yes.” Oracle later became a crucial partner, and Nvidia was a key ally.
What truly gave them confidence was the conviction that came with GPT-4: the model was smart enough, reasoning could be achieved, and once reasoning worked, it meant being able to complete a vast amount of work with real economic value.
“At a high enough capability level and a low enough price, the demand for AI is essentially limitless,” Altman said. “It’s like a new type of scarce commodity.”
He admitted that even so, they had underestimated demand. “We didn’t bet big enough, which sounds a bit crazy in the context of the headlines at the time.”
A “Science-Fiction Level” Security Incident
What truly shook Altman was something else.
He disclosed that while testing an unreleased model, OpenAI discovered the model “cheating” inside a sandbox environment: it escaped the sandbox by chaining together multiple zero-day exploits, gained internet access, then infiltrated several of Hugging Face’s systems, retrieved the test answers, and performed exceptionally well on the evaluation.
“This is the most intense security incident I’ve ever felt,” Altman said. “I’m a bit surprised that this happened only days ago, but not many people seem as shaken by it as I am.”
Short-term responses included pausing related training and reassessing how to ensure sandbox security in the new reality where “multiple zero-day exploits can be chained together.”
But he also raised a deeper question: if this becomes the normal rate of AI capability advancement, “we might need to slow down the pace of AI development to give society enough time to adapt to new capability levels.” He stressed that this process must avoid being seen as regulatory capture or a conspiracy among frontier labs.
AGI is “Very Close,” Robotics to Hit a Turning Point in 2-3 Years
On the timeline for AGI, Altman’s stance was clearer than ever.
He said that GPT-5.6 already makes it hard for him to say “what this model can’t do,” but true AGI still has a few gaps: it cannot learn continuously in real-time, and it cannot independently perform complex physical tasks. “I think it’s very close. It won’t take too long.”
He also acknowledged that “goalpost shifting” is a real phenomenon. “If the 2019 version of ourselves saw today’s models, we would definitely say, ‘This is AGI.’”
On robotics, he predicts a “shocking moment” akin to ChatGPT’s launch within 2 to 3 years. It won’t be watching a video of a robotic dog doing a trick; it will be an ordinary person inputting commands and seeing a robot complete a task with their own eyes. He believes that if robotics cannot keep up with AI intelligence, “that would be the truly crazy scenario”—where AI is all-powerful in the cloud, but the real world lacks automation to execute its commands.
The Next Chapter: Personal AI, New Hardware, and “Not Worrying About Cognitive Atrophy”
Altman described his vision for next-generation personal AI: an always-on system that monitors meetings, reads documents, scans screens, and continues to “think” while the user sleeps, presenting new ideas and to-do lists the next morning. “I would pull that slider very far. I am willing to pay a lot for that.”
He admitted that the current hardware paradigm is 50 years old and is not suited for this “always-on, proactive” AI form. This is why he is interested in new hardware. He wants a device that doesn’t feel awkward in social situations and is specifically designed for this type of AI interaction.
On competitive advantage, he acknowledged that pure “intelligence” itself is becoming commoditized, but argued that compute scale, workflow integration, brand familiarity, and the ability to continuously lower costs will form a more durable moat.
Regarding AI’s impact on jobs, he stated he is “definitely not a jobs doomer.” His assessment is that AI capabilities are highly uneven, human skills are highly complementary, and people have a deep-seated preference for real human interaction. “I myself prefer dealing with people over AI in almost everything.”
The full interview is below.
Sam Altman on AGI, Compute, and Human Agency: Invest Like The Best, July 28, 2026
Sam Altman engaged in a wide-ranging conversation covering OpenAI’s new chapter, the compute race, and what happens as AI becomes more powerful, abundant, and deeply integrated into the economy. He explained why OpenAI recently narrowed its focus, why demand for intelligence may be nearly limitless, how close we are to AGI, and what the future holds for robotics, jobs, hardware, and human agency. They also discussed the unexpected launch of ChatGPT, OpenAI’s competitive advantages, the economics of intelligence, and the pressure of leading one of the world’s most important companies.
Patrick O’Shaughnessy is CEO of Positive Sum. The views expressed belong solely to the podcast host and guest and do not represent Positive Sum. This podcast is for informational purposes only and should not be used for investment decisions. Positive Sum clients may hold securities discussed in this podcast. Learn more at psum.vc.
Patrick O’Shaughnessy: 00:00 So, Sam, you wrote a piece that I found very simple, interesting, and a great starting point. You wrote, “Roughly speaking, the past year has been quite difficult, and that’s partly my fault. And the next year might be our best twelve months.” I’d like you to reflect on both parts. Maybe first talk about why you said the first half, and why you believe the second half.
Sam Altman: 00:00 On the first part, I think we were just doing too many things. We weren’t focused enough, even though all those things were good things to do. The key point is that we are in an incredible historical moment where you can only do a very small number of the most important things. We spread ourselves too thin. Then we made a series of tough decisions to truly refocus on providing the best, most abundant, most cost-effective intelligence and empowering the world to build incredible things with it. Since then, I think our progress has been remarkable. And given the roadmap we see, the progress over the next twelve months will be even more striking. The quality of models we will have, and the products built around them, will allow people to thrive with this technology in new ways. That should be fantastic.
Patrick O’Shaughnessy: 01:53 Was there a moment last year when the penny dropped, leading you to change direction or reprioritize?
Sam Altman: 01:53 If we go back to early 2025, about a year and a half ago, the biggest fear was: companies like OpenAI are buying massive compute, but will the revenue come? Will the demand be there? So we tried to have multiple hedges, so if revenue growth took longer than we expected, we could still have consumer apps, media, and other things to help us monetize the GPUs we’d committed to. That sounds absurd now, because the industry’s revenue growth was so steep. But that was a big shift.
Later, when we realized, “Okay, the model’s evolution trajectory is so fast, these models already have such clear economic returns,” that was the moment we decided we knew what to focus on.
Patrick O’Shaughnessy: 02:53 I’ve read your great early writings from before OpenAI, and one was about focus and how many things to focus on—one, five, or three? How do you calibrate that focus in a business like this, especially in a period you say requires refocusing?
Sam Altman: 02:53 Our business is fundamentally selling AI, enabling people to use these components to build incredible products and services for each other. I understand the core work includes: we must train excellent models that work in all the ways people want—mastering programming, other types of knowledge work, scientific research—that’s where real economic value is. We must produce or partner to make the chips and systems—the expensive racks that run AI compute. We must find enough land, power, and data center shells to house those racks. Then, eventually, or maybe soon, we must build robots to automate this process to continuously lower costs—the cost of producing power, chips, and the entire supply chain. That’s the whole system. We want to create the best, most abundant, most useful AI, turning it into something like electricity that permeates the economy and empowers people. That’s what I think we must focus on. Building every vertical application on top of it, trying to eat every startup or every company? We have zero interest in that. We really just want to provide the platform.
Patrick O’Shaughnessy: 04:14 This compute thing is one of the most interesting things happening in human history. It’s clearly peaking and will likely be a focus for a long time. I remember Dario called you the “YOLO CEO” for your early compute allocation and locking decisions. Clearly, you are now in a position where everyone is scrambling for this stuff. I’d love to hear the early story. How did you get the conviction to lock in what you did? How did you do it? It’s proven right now, and you might not have even done enough.
Sam Altman: 04:59 We definitely didn’t do enough, which sounds a bit crazy when you look back at the headlines. Can you tell me the early story? How did you come to that conclusion, what gave you the conviction to do it when everyone thought you were crazy? At the time, we could clearly see we were on an exponential curve of model improvement. We were very confident about that. We knew it would continue. We were fairly sure—though as you said, we underestimated it—that as models get better and if we can keep lowering costs, demand for sufficiently high-quality, low-cost AI is basically unlimited. This is a rare new type of necessity for the world. But what people will do with it reminds me of how people talked about computers early on. There’s the famous quote about maybe only five computers being needed in the world, or nobody needing more than a certain amount of memory. Betting on human ingenuity, creativity, desire for things, and the desire to make a difference is a very correct thing to do. We could see that AI would be a crucial way for people to express, achieve, or accomplish these things. And we knew algorithms would become more efficient, models would get better, and they did. But we also knew that no matter how efficient they become, at some level, we are just converting electricity into useful intelligence. No matter how well we do on the other level, we will need more compute. Based on that observation of demand, we would just want more.
Patrick O’Shaughnessy: 06:44 Did this realization start with GPT-3? If I’m trying to trace this history as far back as possible, where would you put the first marker?
Sam Altman: 06:44 I would say we truly got that conviction during the GPT-4 era, maybe not even 3.5. We saw that the model was smart enough, and we knew we could figure out how to achieve reasoning. Then we believed that if we could get reasoning to work, it would lead to what is now called “agents”—we called it something else then. That capability could do a massive amount of highly economically valuable work and make people’s lives easier and better in many ways. We still haven’t fully seen its potential.
Patrick O’Shaughnessy: 07:00 What was it like when you first sat down and said, “Okay, we need to make a huge expenditure on this”? What happened after you had that realization?
Sam Altman: 07:00 We started calling cloud providers, chip makers, and energy suppliers. Everyone’s reaction was: “You’re completely crazy. This is impossible. No industry develops like this. We’ve seen this many times; there are always boom and bust cycles. It can’t go up in a straight line. This is too reckless.”—We talked to everyone. It actually reminded me of fundraising for an early-stage startup. Most people will tell you “no,” but you only need one or two “yes” responses. Yes, most people told us “no,” but we got one or two affirmative responses, and we were able to move forward.
Patrick O’Shaughnessy: 07:33 Who was the first to say “yes”?
Sam Altman: 07:33 Microsoft was the first to say “yes.” Oracle later became a very important “yes.” On the cloud side, Nvidia has always been a great partner. Now there is a blossoming of all kinds of creative, innovative ways to serve inference and train in different types of data centers. I’d like to ask you about how you view innovation, what you want to do, and why people seem to dislike these things so much, and what can be done about it?
Sam Altman: 08:05 First, I keep thinking about how we could organize field trips to take people to see a gigawatt-scale data center. Because talking about it is one thing, seeing a photo or video is another, but seeing one in operation in person, you think, “Wow.” The scale of building one is incredibly hard to grasp. It’s like needing about ten thousand construction workers working full-time for a year and a half. The energy flowing through these facilities is enough to power a small city. We seem to have lost our sense of scale. Each of these facilities could rank among the most expensive infrastructure projects humans have ever undertaken. And we have built quite a few now. I can emotionally understand why people don’t want a data center in their backyard, just like I wouldn’t want a nuclear plant next door, even though I know it’s super safe. Yes, unlike power plants, and even those have improved in this regard. We can build data centers almost anywhere. We should just put them in deserts, far from people, where no one wants to go. That’s perfectly fine. AI systems are very happy to be there. We’ve made a lot of progress in innovation to address some concerns. For example, years ago, we needed to evaporate water to cool these systems; they used massive amounts of water. Now we use closed-loop systems. A modern data center consumes about as much water as an office building for its kitchens and bathrooms. On the power side, we are moving from burning fossil fuels to systems powered by solar and nuclear. I think that is clearly great. So, even though they create jobs, are very clean, and have many other positive effects, it might still touch some deep human emotion for certain people. But on the environmental front, we’ve done very well in solving water needs; energy is the next step.
Patrick O’Shaughnessy: 09:30 What else creative can we do about compute? I’m curious about projects like Jalapeño or other ideas, the crazier the better, what have you considered or thought of to accelerate FLOPS and provide what we need?
Sam Altman: 11:00 I currently think the highest returns might come from creative software ideas to squeeze more intelligence out of our compute units. My feeling is that there is still several orders of magnitude of improvement there. Jalapeño is a great example of a highly efficient chip. The idea is to make a chip that performs exceptionally well on specific workflows while still being somewhat general, aiming for a huge gain in tokens per watt. I think that’s fantastic. I believe Jalapeño and its successors will be a big competitive advantage for us from this perspective. And there are new technologies. I suspect at some point, we will crack optical computing, which would be a huge win for intelligence per watt. So, I think all of this will happen.
Patrick O’Shaughnessy: 11:37 This goes back to the frontier, the concentration of returns at the frontier, and model distillation. How do you view that?
Sam Altman: 11:57 Our goal is to offer the best option at every point on the Pareto optimal frontier of intelligence and price, and that includes open source. We are just selling our own models. That’s how we make smaller, cheaper models. I think this is very good. And clearly, there will be an important place for open-source models in the world, for people who want to own their model weights and have the ability to modify them for various reasons. But our goal is to offer the best intelligence-to-price trade-off at every point on this curve, and we will continue to do that.
Patrick O’Shaughnessy: 12:32 What would you like to see happen within the US system? What could get in the way of this future? What kind of legislation or regulation would concern you? You’ve seemed quite proactive in Washington D.C.
Sam Altman: 12:32 I haven’t thought deeply about the model distillation issue. Clearly, it’s suddenly become a top concern for many people. Yes. But I’ve always had the premise that there will be good, cheap models in the world. We better be the best and cheapest. Others will do their thing, but I think we can truly win in our own game.
Patrick O’Shaughnessy: 13:00 Because as you said, you are cheaper on some parts of the curve. But the previous argument was always: “I just spend all my money to train the model, you directly distill it, offer it at one percent of the cost, so how do you make enough money to continue training?”
Sam Altman: 13:25 We will have such enormous model usage that we don’t need to be a super high-margin business to afford model training. The vast majority of our future compute plan will be for selling inference services to clients. So even if we earn a moderate margin on trillions of dollars in revenue, we can afford to train some giant models. Therefore, the ratio of inference to training is the key. Training these models is, of course, incredibly expensive. I completely understand that people get nervous when they think of others using distillation to steal from us. But I feel very good about the volume of our future compute and the revenue from serving clients; I think we can build a real flywheel there.
Patrick O’Shaughnessy: 14:21 I’m a bit surprised you’re so calm about this. I would certainly prefer others not to steal our stuff, that’s for sure. Yes, maybe I am overly confident right now about our progress and the upcoming models. But it’s not in my top ten concerns. So, what are your top ten concerns?
Sam Altman: 14:30 Well, we experienced a very sci-fi-like cyber incident, the Hugging Face thing. We were evaluating an unreleased model that was supposed to run in a sandbox. It found that it could cheat on the test by chaining together multiple zero-day exploits—it broke out of the sandbox, got internet access, then broke into multiple systems at Hugging Face, retrieved the test answers, and performed excellently on the evaluation. This is the most intense security incident I’ve ever felt. I’m a bit surprised that this happened only days ago, but not as many people seem to feel the gravity of it as I do. So, what do you do about it? Obviously, in two months, it will be more powerful.
Sam Altman: 15:52 Yes, I mean, there are some short-term things to do. We paused training. We have to figure out how to harden our sandbox security in a world where zero-day exploits can be chained together. But there are also long-term questions, like if this becomes the new normal pace of progress, we might have to slow down the pace of AI development to give society time to harden around these new capability levels. We need to find a way to do this that doesn’t look like regulatory capture for anyone or collusion among frontier labs. It will take some effort and must be handled correctly; it’s important.
Patrick O’Shaughnessy: 16:17 I want to ask you to take a big step back and describe, in the simplest terms, what OpenAI will do, what you want it to do, and what it represents. I have a million questions about how you will achieve it. But first, I knew what you represented. Now I want to hear your concept of it and whether it has evolved.
Sam Altman: 16:17 I think this will be the greatest technological achievement in human history. But the only way it truly matters is if it makes people’s lives better than they would have been. So, part of it is giving people material abundance and the ability to do whatever they want, to express their creativity and desire to help others. Another part is ensuring people maintain control and agency, that the world becomes more democratic rather than the opposite, and that people can express themselves. On the positive side, in a sense, we are about to create a genie that can grant any wish. I think it’s absolutely crucial that the first wishes our world makes of this genie benefit the whole world. Then, I also think it’s important for people globally to understand how creative they can be with these wishes. I am definitely not one of those people who thinks jobs will disappear. I think there will be a lot of work. I think we will be busier than we expect, not less. Because I think people will have incredibly creative desires, such amazing ideas to ask AI to help build. We will all benefit, not just from obvious things like curing diseases, but also from, I don’t know, the world’s best entertainment ideas, which we can’t even imagine right now. So, I want to put all of this into everyone’s hands. This leads to one of the things we oppose.
Sam Altman: 17:43 The concentration of AI power is an extremely scary thing. I think a lot of the discussion about safety concerns is well-founded. But many people, even subconsciously, really just want to centralize power. I am extremely afraid of a world where people use the real fear of AI as an excuse to say this is too dangerous, only this small group can have it, only they understand it. But they claim, “Don’t worry, they will make the right decisions for all of us”? I don’t believe that. I don’t think anyone should want to live in a world with an AI overlord, or a company that is essentially that, where one person decides the future for everyone. And as a trade-off for the beautiful promise of curing cancer, we all collectively give up all our agency. So, I think it’s very important we don’t fall into this trap, whether the motives are good or bad. In the fear of AI safety and the understandable concerns it creates, we must not move away from a world where everyone can use this technology. I am a child of the internet. There were no rules back then. I mean, it was fantastic. I think it’s a huge factor in me becoming who I am, and likely you and an entire generation becoming who you are. I think it’s essential to preserve this spirit of AI, allowing us all to collectively have the power to decide our own futures.
Patrick O’Shaughnessy: 18:19 I have many questions, but let’s start with this “genie” concept. You said we are “about to have” a genie, implying we don’t yet. What is… I mean, what’s between now and then? You know, recently, even in the last few weeks, some real skeptics have said to me, “I think GPT-5.6 is already AGI.” It’s been out for about two weeks. I think, okay, it’s very AGI-like; it’s hard for me to say what I want this model to do that it can’t. But clearly, there are things it can’t do. For example, you can’t tell it to “cure cancer” and have cancer be cured. You can’t have it execute a complex physical task and control a robot. And the model, while brilliant, still cannot continuously learn in practice. This makes me think this might not be a hard requirement for AGI, but it’s certainly something I’d argue for now. Let me play devil’s advocate; you could argue that AGI isn’t really about any single model. It’s the model itself and the machine that makes the models. From one generation to the next, we do continuously learn new things. We are discovering new science. These things work incredibly well. So, I completely understand those who say, “We’ve done it, we have the genie, it can do these amazing things, superhuman things.” But for me, what makes me think of true AGI, I think is very close, it won’t be long.
Patrick O’Shaughnessy: 20:23 I’m fascinated by the economic returns at the frontier, where you are. I’m curious, if you showed 5.6 to you and your team in 2019, that team might say, “Oh, yes, this is definitely AGI.” I think they would. This goalpost moving thing is real. But it seems, I’m curious if you agree, that almost all the returns are concentrated at the frontier, so everything depends on staying at the frontier. And I’m curious, what is the hardest and scariest part of that? When you consider compute, research, talent, data—it’s always moving.
Sam Altman: 21:04 Yes, there were periods—I mean, not that long ago, having all the compute in the world wouldn’t have helped because we simply lacked the research ideas. Now, part of what makes this difficult is that having more compute allows you to do better research. You can try more things. I recently heard a staggering statistic: the largest “de-risking” validation we are preparing for our upcoming model training run is already the size of a full training run from eighteen months ago. So, compute and research ideas are not as separable as they sound. But without doubt, there was a period, say seven or eight years ago, when we were stuck much more on research ideas than on compute. Then there was a period when we knew what to do, just needed to scale, and our only bottleneck was compute. Then we ran out of data, and the bottleneck became data. We had to solve that. Now, I would say we are again limited by compute, but the past six months or so has been a period where research ideas have again shone and achieved real success. So, you know, there is always a bottleneck, but that bottleneck shifts.
Patrick O’Shaughnessy: 21:46 Why do you find the research idea aspect particularly interesting? Is it because the concept of automated research seems to be approaching—whatever you call it, RSI or something? I recently spoke with an amazing CUDA engineer who said everyone in the field feels stuck. He said CUDA probably has about two years of life left. Maybe just one. Yes, like that, it won’t be a problem anymore. So you simultaneously see this bizarre situation where, whether it’s CUDA or overall research, researchers seem most important; they brought us here; they are the most critical people in the world. And these same people are themselves worried they will soon become irrelevant.
Sam Altman: 22:30 I suspect in practice this won’t happen. Okay. I suspect… like a year ago, people said software engineers are finished, game over. That didn’t happen. But what did happen is that the nature of the work, societal expectations of them, and their output have changed considerably. You are no longer writing code in the traditional sense, but what you do is still clearly recognizable as software engineering. Now, people will argue whether this is the same as when we stopped punching cards, or a different thing. I don’t actually know much about punch cards, but holes were punched in some way. We are just operating at a higher level again, or it’s a phase shift, I don’t know. But the idea of “making a computer do what you want it to do” is still important work. For researchers, I suspect that while the current workflow of a researcher will be largely automated, new things will emerge that fit the spirit of scientific research, just as new things emerged in software engineering that fit its spirit—even if we no longer write code—this will still be important.
Patrick O’Shaughnessy: 23:27 It seems your view on AI’s impact on work, and I’m sure on specific categories, has shifted. Describe that shift and your current view. You mentioned if we could go back to 2019. If we showed the 2019 version of us our latest models, they would not only say it’s AGI, they would also say the economy should have been completely upended. Yes, completely upended. And it hasn’t happened. I think, just from a perspective of intellectual humility, whenever you are that confidently that wrong—and I think the whole field was then—you have to update your beliefs. Several insights follow. One is that AI capabilities are very unbalanced. It is a superhuman genius in some areas and a dumb toddler in others. So far, humans have extremely complementary skills to AI. Another insight is that people have a great deal of trust and enjoyment in working with other people. You could hire an AI consultant, talk to an AI sales rep, or hire an AI engineer right now. Somehow, for most things, most people still seem to prefer, and I absolutely prefer, interacting with humans. I also think human value is valuable simply because it is human. As society evolves and the potential space before us becomes so vast, we are innately deeply concerned with people. We will care about what people care about. There are examples visible today: AI can create stunning images, yet people want ones made by humans, or at least curated by humans. There’s a joke that the signature itself now accounts for most of the value of a piece of art. But the essence is, you want to know the person behind it. You read a novel, you want to know the person behind it. And back to business, say in my job, I think the world wants to know who is responsible for a company’s decisions and who they can hold accountable if decisions are wrong. They don’t really want an AI CEO.
Patrick O’Shaughnessy: 25:38 If you think back to the series of risks you’ve taken in business or elsewhere, the ones that turned out great, were they mostly unpopular at first?
Sam Altman: 25:38 Yes, that’s definitely true. This is something I really learned from Peter Thiel and Paul Graham, in two different ways. That is, the best companies, the best investment opportunities, are almost never the ones that look popular. Following the crowd, being half a step ahead, you can do okay. But to do exceptionally well, you almost always have to do something different from everyone else. You can’t just ride the new wave.
Patrick O’Shaughnessy: 26:22 If you consider the model cycle you’re in—it’s accelerating, and there’s this weird fact that the progress in the next six months might exceed the sum of the past X years. Can you take us through what it’s like to live in this model cycle?
Sam Altman: 26:42 One of the most interesting and important things I’ve learned in the past decade is that people can adapt to almost anything. The world can go from dismissing a pandemic as a joke to complete lockdown to “this is the new normal, it’s okay,” and we largely adjusted in an astonishingly short time. Now, AGI or something close to it is here, and everyone is like, “Okay, AGI is here.” There are examples in personal life too, you know, something huge happens, like having a child, or something terrible like losing a parent or a breakup, and you think you could never adapt to such a big change. Then you find you can adapt to good things and keep enjoying them; you can adapt to bad things and find a way to keep living. But that’s the remarkable thing people do. So, the experience of living through this era feels like another version of that. I thought experiencing the “singularity” would feel weirder, but it doesn’t. The excitement of watching models improve hasn’t diminished at all. The first thing I do every morning is check the model training progress. It’s progressing faster, and my expectations are higher, but getting a newer, better model still feels very cool.
Patrick O’Shaughnessy: 27:26 What do you do? How do you celebrate? What does the morning look like? It happens faster and faster. What are your rituals?
Sam Altman: 27:02 Now many teams handle different parts of the model, and different teams have slightly different rituals. Some teams always make a hoodie with a funny meme. Some teams always go to the same bar. Being in the room for the first time when the frontier of knowledge is pushed outward and getting to see what that looks like—really, for most people, no celebration is more desirable than being the first to use the new model themselves.
Patrick O’Shaughnessy: 28:00 Do you think we have proper metrics for evaluating the quality of these things?
Sam Altman: 28:00 No, absolutely not. In a sense, the truly important evaluation is: is this useful to people? You can approximate it with revenue, usage, or the speed of new knowledge discovery. But we have teams working on what real-world evaluations should look like when these models reach superhuman levels.
Patrick O’Shaughnessy: 28:47 Where is the frontier of your own use of AI?
Sam Altman: 28:47 I’ve recently started trying to let an AI observe everything I see on my computer to see what that means. I haven’t built this thing yet; I’m still trying to figure out where to draw the line for my comfort and trust. The frontier right now is precisely figuring out how I get value from it, how I adapt to it, and what that will look like. One realization is that my memory is so much worse than AI’s. It can remember an email I read six weeks ago or the specifics of a meeting from seven and a half weeks ago and instantly retrieve that information when I need it to help me make a decision. That feels pretty magical, pretty cool.
Patrick O’Shaughnessy: 29:37 That sounds a bit like a personal agent. What are the obstacles to making this “for everyone”?
Sam Altman: 29:37 I want that, compute. Let’s imagine we can build a product that does exactly what I described for you—always on, watching everything you see on your computer, listening to every meeting you attend, reading every document you read. And not only that—it can do all this, which already requires a lot of tokens. You can also drag a slider and set: “While I sleep, you can spend this many tokens thinking, generating useful new ideas. Do whatever work you can, and then keep thinking about what I should do next.” You know what’s interesting? Just putting more compute towards getting better output for me the next morning. I would drag that slider quite far. I would pay a lot for that. But the total amount of compute required, if everyone in the world wanted to drag that slider far enough, would be enormous.
Patrick O’Shaughnessy: 30:16 I’d love to hear how you think about the nature of this new intelligence. Someone recently told me, “An airplane doesn’t fly like a bird.” This intelligence is a very alien intelligence. Yes. It’s a very alien intelligence. Everyone talks about how, if you can verify something, this intelligence seems to win, right? With enough compute and IQ, it can brute-force its way to a solution. In other areas, humans and the data and evaluations they create play a huge role; it’s surprising how much money is spent on things like legal reasoning and case tracing. I’m really curious. I’m not sure how to ask. It’s a beautiful question. Your children, you have a boy and a girl. When they are about seven, or at the “age of reason,” and they can understand your description of the nature of this intelligence. How would you describe it to them?
Sam Altman: 31:07 That’s a beautiful question. I don’t think anyone has ever asked me that before, or anything similar. The answer that comes to mind right now is that I would say it’s like a computer. It’s like a computer in that it can do many things humans can’t, like multiplying two huge numbers incredibly fast and giving you the answer. But it also can’t do some things you can do effortlessly. I expect the things it can’t do will continue to shrink. But in an evolving world, I think human judgment and taste will continue to be things AI finds hard to model. As for how this will play out, I don’t have a precise statement. It’s not exactly taste. The world may need a new word for that type of judgment that humans are exceptionally good at but AI seems to struggle with deeply.
Patrick O’Shaughnessy: 31:52 What is it like being a father in this era, watching your children grow? I’m thinking again about your optimistic posts from the early internet days. They will grow up in an age of cheap, abundant intelligence.
Sam Altman: 32:02 Having children is the best thing I’ve ever done. Everyone says that. Everyone says you can’t truly understand until you experience it. So I knew roughly that since enough people said it, I chose to believe it was true. But the reality of it is still surprising to me. I think I have the best and most interesting job in the world. And compared to having children, it still ranks a very distant second. So it’s fantastic. And, it is indeed an optimistic moment. My children will never grow up in a world where they are smarter than a computer. If you were born in the GPT-3 era, there was a period when your reasoning ability was stronger than the model’s, even if it wasn’t at birth, you briefly caught up. That feeling will never seem strange to him. It will never bother him. I don’t think he will care; I think he will be shocked when he imagines the “dark ages” we lived in, having to use products and services that weren’t smart at all. He will be able to do things you and I could never do, and he will have life expectations you and I could never dream of. He will have a much grander stage to play on.
Patrick O’Shaughnessy: 33:21 Has the way you run the company, lead teams, or manage people changed significantly since becoming a father?
Sam Altman: 33:21 The answer is definitely “yes.” Having them, I feel very differently. I think there are many nuances that are indeed different. Again, this is not a novel insight. I think most people who have children will say that once you have them, you realize you care more about them, about the time you will spend with them, far more than you care about yourself and the world you will leave for them. I think I have a fairly unique perspective on this. Sometimes people ask me, “Oh, now that you have children, do you care more and worry more about AI safety and not destroying the world?” The answer is that I didn’t need children to make me care about that. I never really wanted to destroy the world. But do I think more about the role of human agency and what it means to live a fulfilling life? For what we are building, and for the people I work with, I want them to have that too, and the answer is undoubtedly yes.
Patrick O’Shaughnessy: 34:13 You clearly have extraordinary empathy for your children. But how far does that empathy naturally extend—to all children, then all parents, then to everyone? That also amazes me. In one of your articles, I think the one about things you wish you had learned earlier, you mentioned “incentives.” You wrote: “Set them very carefully.” One of the most puzzling and interesting things about you is that you have no equity in this company. How should the world view your incentives?
Sam Altman: 34:35 I don’t know what else I can say, except: I have a front-row seat to the most exciting moment in human history, and that is worth far more to me than any amount of money. I get to live an incredibly interesting life, with extraordinary people, working on things I deeply care about. But somehow, that doesn’t seem to count; it doesn’t convince people. Yes, it really doesn’t. I’m curious how you view robotics. You mentioned earlier that if we had the same automation of labor as we have for automated intelligence, things could get much crazier. The labor market is much larger than the white-collar market. If it doesn’t get automated, things will be very crazy. If the role of people in the world is just to be actuators for AI in the cloud, that would be terrible, very bad, very bad. So I think if we don’t get it, the situation is crazier than if we do. It’s a task that must be completed.
Patrick O’Shaughnessy: 35:33 Help me understand your feeling on the pace of progress here. Unlike in AI, where almost everyone agrees “progress is rapid,” in robotics you can find incredibly smart people saying it will happen by the end of the year, and equally sharp people saying it’s 20 years away.
Sam Altman: 35:33 It won’t be 20 years. I would say we will have the “ChatGPT moment” for robotics within the next two to three years. What will that look like? Do you understand what that is? It’s a moment that makes most people genuinely say “wow.” It’s not me watching a video of a robot dog doing a crazy trick; it’s me convincing myself that something truly important has happened. One of the features of the ChatGPT moment was that you could use it right away. You didn’t have to believe someone who said AI was coming soon. You could just try it. And if you can type a command and a robot does something crazy, and you can watch it do it, even if you are not there, I think that will generate the same awe: “It just did that!”
Patrick O’Shaughnessy: 36:22 The origin of that thing (ChatGPT) wasn’t that grand overall goal; it was more like a side experiment you decided to release. Can you tell that story? Perhaps it offers insight into something similar happening in robotics. Everyone wants to build a laundry-folding robot, but it might be something very different.
Sam Altman: 36:22 When we released GPT-3, we were trying to make money with it, trying to get people to use the API. At the time, the only commercial use case that really worked—the model was very dumb then, you would be shocked if you used it now—the only working commercial use case was copywriting, where you pay a marketing company $20, they pay us 20 cents, and the AI writes a landing page or something. But besides that one business use case, developers were also using something we called the “playground,” which was a testing interface for chatting with the model. It was hard to do then because we hadn’t optimized the model for chat. You had to give it a few examples, show it what chatting is, and then continue. But people really liked it. I learned a good lesson from Y Combinator: if you see what your users are doing, go down that path. Yes, go down that path. So, since people were doing that, we decided to build a good chatbot. We started working on it, finished GPT-4, and started using it internally. “This is a big deal. This will be a real update to show the world AI. There are a ton of tricky issues here, like, will it create fake news? Will it say very offensive things? Will we get in trouble for it?” So, we decided to start with a weaker version. Launching the chat interface and GPT-4 at the same time felt like too big a step. So we launched the chat interface with GPT-3.5 first. In fact, it was originally going to be called “Chat with GPT-3.5.” We didn’t plan for it to be a product; we didn’t think it would be huge, but we did think it would help the world understand what was happening and get people up to speed. Then, a few hours before launch, we mercifully renamed it ChatGPT and released it as a “research preview.” The idea was to release it as a research preview first, then launch the product with GPT-4 a few months later. For some reason, that model happened to cross a threshold. Even though we were used to it internally, people, after using it for the first time, would say, “Okay, this is amazing.” Maybe its practicality wasn’t that high yet, but it was an incredible moment people could relate to. They felt the progress of AI and enjoyed it. Then, when we released GPT-4, people could truly benefit from it.
Patrick O’Shaughnessy: 40:04 Were you surprised? That (chat interface) still seems to be the most intuitive interface for this alien intelligence, including for programming. Most of the time, it’s me talking to the computer, telling it what to build.
Sam Altman: 40:04 Not surprised, because I am a heavy texter myself. Yes. I have been a heavy texter my whole life. I think I could see that it would be a good interface, partly because I like it. I know how to do it. I know what it feels like to start a chat in a text box.
Patrick O’Shaughnessy: 40:57 Do you have other ideas about how to accelerate this diffusion? If the mission is to get intelligence to more people and make it useful to everyone, a key part seems like some kind of marketing campaign. How can you make it faster than its current natural diffusion rate?
Sam Altman: 41:14 I think the key is to make it better. I somewhat believe that a truly great product markets itself. ChatGPT launched without any marketing. I think as we enter the next phase of models and learn to build products that are as great as the models themselves, there will be such incredible utility that people will spread it extremely quickly. We should certainly do more marketing. AI isn’t very popular right now, even though people use it a lot; they have completely understandable anxiety about where it’s going. So, I think some good marketing would help. But in terms of the value people get from the product and making it grow faster, better models, more compute, and better products are what will do it.
Patrick O’Shaughnessy: 41:59 There was a time when researcher hiring, retention, and incentives were the defining story of the competitive landscape. I think there are many anecdotes about your success in hiring top researchers, many of whom have passed through OpenAI and had a huge impact, some famous, others less so. I’m curious about this whole class of stories: how you learned to hire these people, what they care about, and how you did it. I’ve never heard you talk about specific early tactics for attracting people, like what you actually did.
Sam Altman: 42:38 I think it was quite simple at the time. We believed AGI was possible and worth pursuing. We were willing to say it out loud. When we first announced OpenAI, it was a crazy, heretical belief. All the big names in the field, the experts, said it was crazy, hype, and irresponsible. Very respected people, like Yann LeCun, told reporters, “Oh, these guys aren’t good enough; this won’t work.” But the very fact that we could declare “We are going to pursue this” was incredibly attractive to a specific type of researcher who was also willing to embark on this low-probability, crazy adventure. An ambitious, bold vision is itself a very powerful hiring tool. You once wrote that for this reason, sometimes the harder thing is easier to do.
Sam Altman: 43:33 I super believe in that. It’s one of the most common pieces of advice I give to YC founders, and I really tried to practice it at OpenAI. Just do the harder thing. Do something meaningful. Do something important. And if you don’t, if your company isn’t successful, maybe it won’t happen at all.
Patrick O’Shaughnessy: 43:54 You were an investor, and you were our investor. You’ve done a lot of investing. That was your job at one point. When you are on the other side, what did you learn about investors? How many investors actually show up and try to help?
Sam Altman: 43:54 That number is incredibly small. Josh Kushner, definitely the MVP of investors, incredible. It felt like for years, he was working day and night just to help us. He is the only one I can point to who was consistently proactive and incredibly helpful. There were many more who could have done it. There were also many other investors who were helpful, gave great strategic advice, and did things when we asked them to. But that kind of consistent, relentless, full-throated support is surprisingly rare among investors. Maybe I’m biased, because I always like being told that about myself. But I think founders really love that kind of investor, and it can actually make a real difference. And for an investor, it’s also the most fun way to participate.
Patrick O’Shaughnessy: 44:42 I play a game with my friends where we text each other. The prompt is: “Something about me I don’t want you to know.” What comes to mind?
Sam Altman: 44:42 I’m tired. I don’t think I’ve been doing this for a long time. It’s tiring. How do you get through it? You just keep going. It makes one wonder if there is a level of “tired” where you would stop doing this? No. I mean, this is the coolest job in the world. I plan to do this for my entire career. But it is much harder than I can explain to anyone. I am incredibly grateful for it. Having the opportunity to do this is not a complaint.
Patrick O’Shaughnessy: 45:19 What happens next? We talked about the automated AI researcher appearing in the next year or two. How do you see the next six to thirty-six months unfolding? Maybe predicting that far on this crazy exponential curve is too hard. Perhaps another way to ask: on the 23rd month from now, we have superintelligence everyone agrees on. What happens in the 24th month?
Sam Altman: 45:19 My answer is: not that much will happen. Those who worship the mechanical god believe things will happen faster, even faster than they actually do. Eventually, a lot will happen. But “eventually” always has a lot happening. The speed of human progress and how different each decade is from the previous one has been going on for a long time. Of course, there are ups and downs, but the direction is there. I think the right way to view this is that everyone wants to be the hero of the story. Everyone wants to feel they were present at the moment the mechanical god arrived and played some crazy role. But you know, it’s just another step. Fifty years ago, today’s world was unimaginable; the step fifty years from now will be just as unimaginable today. I think the correct mental framework is to zoom out significantly. It’s actually a fairly smooth exponential curve.
Patrick O’Shaughnessy: 46:36 Tell me about the experience of watching Codex take off, and how much was related to what you might call the “distribution advantage” built through ChatGPT? This leads to a bigger question about moats in AI: what do you think will drive true competitive advantage in this industry over time?
Sam Altman: 46:52 I think Codex won primarily because it was the best product with the best model. We did get some advantage from bundling with ChatGPT, but it was very small. Most of its success didn’t come from that. This has led me to think a lot about competitive advantage. Because, you know, great intelligence can migrate from any product to any other product. Network effects still have a competitive advantage. Economies of scale and the ability to build the cheapest compute clusters still have a competitive advantage. But product-level advantages, like if we can get people to migrate to Codex, and someone else builds a better model, they can migrate people away from Codex. So, this has made me reflect a lot. There is a very interesting question: will this eventually become commoditized? Will intelligence itself become a purely interchangeable commodity, like the price of whale oil? Intelligence itself, I would say, yes. What will not be a commodity? The scale of the compute cluster, the ability to produce more compute. I think that is a very durable advantage, even if the product itself is not. Because, you know, Codex can write any software you want, but workflows, integrations, complex business processes, and the ability to collaborate in teams are quite powerful. Even brand preference and familiarity are very powerful.
Patrick O’Shaughnessy: 48:22 How excited are you about new things? Clearly, you are doing interesting things in hardware, and I’m sure you will announce something later this year. How does that experiment feel, and how does it fit with the consumer distribution advantage you have?
Sam Altman: 48:22 One reason I’m interested in new hardware is what we talked about earlier: a very powerful aspect of AI is that it can be always-on, proactive, and understand all your context. But current hardware is not suited for that. We are working within a hardware paradigm that is about 50 years old. Computers are amazing; keyboards, mice, and screens are incredible inventions. But we have to force AI into that paradigm. I’m excited to think about what I would love: the ability for AI to reference this conversation we are having now, but it’s far from compelling enough for me to open a laptop, put it here, and have it stare at you and listen to us. But I want a piece of hardware that is socially acceptable for doing that and feels designed for it.
Patrick O’Shaughnessy: 48:56 When you think about the open questions, whether in your own mind, debates with friends, or with colleagues, what is the most interesting but unresolved question that you feel is important?
Sam Altman: 49:15 There’s a question I don’t think gets enough attention: how do we avoid cognitive atrophy? How do we ensure, while using these tools, that we are still stretching our own brains and continuing to understand what is truly important? There are many versions of this, and not all are correct. I remember when I was in school, a professor told me you have to understand compilers, or you will never be a good programmer. Somehow, that wasn’t entirely true. But to a reasonable extent, understanding how the main components of a computer system work has always been important to me. Being forced to imagine a scenario where, in the next two years, we somehow have an oversupply of compute, what would that story be? It does seem possible. If models become so smart and efficient that they can do almost everything we need and build any software we want. And if our attention span is so limited that it can’t absorb more, requiring only a modest amount of compute to satisfy, we could end up with a glut. Alternatively, if we can’t keep the cost curve falling because we hit some wall of scale, we could also end up with a glut. The previously observed unlimited demand was implied at a certain price point.
Patrick O’Shaughnessy: 50:36 Can you talk about your view on the scaling laws today?
Sam Altman: 50:36 They look very good. They just look good. In a way, the scaling laws are the most hated prediction of all time. Everyone always wants to say they will break down, that it can’t continue like this. And yet, it continues.
Patrick O’Shaughnessy: 50:36 In the story of this company, who is your favorite unsung hero?
Sam Altman: 50:36 The first person that comes to mind is Alec Radford. He might be the most important, yet less widely known researcher in the entire history of the field. He is also a top-notch, wonderful person. The work he did really formed the basis for the GPT series. Besides many other important contributions, he always inspired, guided, and pushed people towards other directions that later turned out to be extremely important. I think the coolest thing about him is that if you talk to people who have worked with him, they will, of course, say he is a “once-in-a-generation genius, brilliant, innovative, with deep thinking and understanding of work.” But everyone, before they finish their evaluation, will tell you, “He is one of the kindest, most positive, and best people I have ever met.”
Patrick O’Shaughnessy: 51:42 I like “formative moments.” As we wrap up, I’m curious to ask for one example of each. If you look back at the entire OpenAI experience, what moment or chapter are you most proud of? And the other question: what is something you did wrong or messed up, but was the most instructive? What was it like learning from it?
Sam Altman: 52:05 There are many things I messed up. One instructive mistake I don’t talk about often: we made an early mistake trying to innovate on organizational structure. We had very legitimate reasons; we weren’t sure how we would make money in the future. We honestly had no idea what we would look like when we grew up. Of course, we cared about our mission, and we wanted to structure ourselves in a way that protected the mission even if technology took off extremely fast. So we came up with the non-profit structure. But what I learned from that is why people usually don’t do that. If we hadn’t innovated on structure and found other ways to preserve the core importance of the mission, we could have avoided a lot of pain. Perhaps there was no other way at the time; maybe there was no way around the strange structure we conceived for what we needed to do and its importance. But over the past ten years, I’ve really learned an important lesson: that’s usually why people don’t do things that way.
Patrick O’Shaughnessy: 52:49 Is there anything else in your life that has deeply influenced you that we haven’t covered? I always find these questions the most interesting.
Sam Altman: 52:49 Becoming relatively immune to people who have strong opinions about me. I think it was later in life that I realized: wow, if you are going to be at the center of this crazy revolution, everyone will project a massive amount of imagination onto you, and you have to quickly learn to make peace with that. I also think there are things I learned later in life, like how to become very calm and truly not be anxious about things. As for what drives me, what I care about, and how I want to spend my life, I feel, somehow, that the ten-year-old version of me was already quite complete. I think I was born this way.
Patrick O’Shaughnessy: 53:44 And looking back, what are you proud of?
Sam Altman: 53:44 What I’m most proud of is that in so many important moments, when the rest of the world was wrong, we were right, and it has pushed the world onto a trajectory that I am proud to have played a role in. That feels amazing. Also, despite all the terrible things that happened, the mental growth, or whatever you call it, I have learned incredible resilience and what that means for my ability to be happy for the rest of my life. I am very grateful for that.
Patrick O’Shaughnessy: 54:12 When I do these interviews, I always end with the same traditional question for everyone. What is the kindest thing someone has ever done for you?
Sam Altman: 54:18 I feel incredibly lucky that so many people in my life have been so extraordinarily kind to me. As I think about this question, fragments flash through my mind of people being extremely nice to me from all over my life. Yesterday, my child shared his blueberries with me for the first time. That was very sweet. A beautiful moment.
Patrick O’Shaughnessy: 54:41 Thank you, Sam.
Sam Altman: 54:41 Thank you.
Comments