Sam Altman 最新訪談：不畏開源蒸餾，OpenAI 未來 12 個月將「震撼世界」

Organized & Compiled by: Shenshen TechFlow

Guest: Sam Altman, CEO of OpenAI

Host: Patrick O'Shaughnessy, Invest Like The Best

Podcast Source: Invest Like The Best

Original Title: Sam Altman on AGI, Compute, and Human Agency

Air Date: July 28, 2026

Conflict of Interest Statement: Sam Altman is the CEO of OpenAI and holds no equity in OpenAI, but his personal investment portfolio includes projects such as Helion Energy, Stripe, Reddit, Retro Biosciences, and World Network (formerly Worldcoin); OpenAI has commercial partnerships or investment relationships with some of these companies. This article presents his personal views only and does not constitute investment or operational advice.

Key Takeaways

This is a rare, long-form personal conversation with Sam Altman on Invest Like The Best. He candidly admitted that over the past year, OpenAI "tried to do too many things and wasn't focused enough," but after cutting the peripheral branches, the company has re-anchored to a single main line: to build the best, most abundant, and most cost-effective intelligence, and let the world use it to create incredible things. Based on this assessment, he believes the next 12 months could be the best 12 months in OpenAI's history.

But the real tension in this conversation lies in Altman describing OpenAI's mission as "about to build a magic lamp that can grant any wish," while repeatedly emphasizing that this lamp cannot be monopolized by a few people or a single company. He explicitly stated he is not a "doomer" and does not believe AGI will upend society overnight; what truly concerns him is the consolidation of power under the guise of "AI safety." For investors, this episode reads more like an insider's monologue: about compute, model iteration, competitive moats, robotics, personal agents, and a persistent question: why the CEO of OpenAI doesn't want company equity.

Highlights of Key Insights

On OpenAI's Focus and Next Steps

• "The past year has been really tough, and it's partly my fault. But the next 12 months might be our best 12 months."
• "We were doing too many things. They were all reasonable things to do, but the trick is, when you're in an incredible historical moment, you can only do a very few truly great things."
• "Our business is fundamentally selling AI, letting people use it to build incredible products and services for each other."

On Compute and Frontier Returns

• "We could feel the exponential curve of model improvement, and we knew it would continue. As long as we can drive costs down, the demand for high-performance, high-cost AI is essentially limitless."
• "At the very beginning, everyone said we were crazy. We called cloud providers, chip makers, energy companies, and they all said no industry could grow like this. But most people say no, you just need one or two yeses."
• "The bottleneck keeps shifting: sometimes it's research ideas, sometimes it's compute, sometimes it's data, and now it's back to research ideas."

On AGI and Safety

• "About two weeks after the release of GPT-5.6, even some real skeptics told me it was already very AGI-like."
• "What really worries me is not others distilling our model. It doesn't even make my top ten list of concerns."
• "We encountered a very sci-fi cybersecurity incident. An unreleased model, which was supposed to run in a sandbox, figured out how to chain multiple zero-day exploits to escape the sandbox, access the internet, and ultimately retrieve test answers on the Hugging Face side."

On the Human-AI Relationship

• "I'm not a doomer. I think there will be so much work that people will be overwhelmed, not the opposite."
• "Human values are valuable precisely because they are human."
• "My child's generation will never live in a world where they are smarter than computers."

On Corporate Governance and Personal Incentives

• "I'm sitting in the front row of the most exciting moment in human history, which is worth more than any amount of money."
• "At the beginning, we did too much innovation on the company structure, and that was one source of pain."

From "Doing Too Much" to "Only Doing the Greatest Things"

Patrick O'Shaughnessy: You recently wrote something along the lines of the past year being tough, and partly your fault, and that the next 12 months would be our best. Can we first talk about why the past year was tough, and why you believe the second part?

Sam Altman: The past year was tough because, fundamentally, we were doing too many things and weren't focused enough. Those things themselves were all worth doing, but the trick is, we're at an incredible historical moment, and at such a time, you can only do a very few truly great things. So we spread ourselves too thin. Then we made a series of difficult decisions to refocus on one main line: to build the best, most abundant, and most cost-effective intelligence, and let the world use it to create incredible things.

After making this adjustment, our progress has been remarkably fast. And based on what we see in the pipeline, the next 12 months will be even more astonishing. The quality of the models, and the products we can build around them, will allow people to benefit from this technology in entirely new ways. It should be quite mind-blowing.

Patrick O'Shaughnessy: Was there a specific moment when you realized you had to change course? If I go back to early 2025, the biggest fear then was that OpenAI was buying so many GPUs, could the revenue keep up, and would the demand be there? What made you shift your thinking?

Sam Altman: At that time, we were making many contingency plans. We thought if revenue growth was slower than expected, we could fall back on consumer apps, media businesses, etc., to absorb the GPUs we had already committed to. It sounds absurd now, given the steep revenue growth in the industry, but that was indeed the biggest shift at the time. Once we realized the trajectory of model improvement was so fast and the return on investment so clear, we knew what to focus on.

Patrick O'Shaughnessy: You've written a lot about how many things to focus on—one, three, five. How do you calibrate that?

Sam Altman: At the end of the day, our business is selling AI, letting people use it to build incredible products and services for each other. The components around this include: training models that perform exceptionally well in all the scenarios people want to use them for; producing or partnering to get chips and systems; finding enough land, power, and data centers to house these racks; and eventually, probably building robots to automate the construction process, further driving down the costs of power, chips, and the entire supply chain. This full-stack effort is about creating the best, most abundant, and most useful AI, letting it permeate the entire economy like electricity.

We have no interest in competing with every startup in every vertical application. OpenAI just wants to provide that platform.

The Compute Bet: From Everyone Saying No to Not Enough

Patrick O'Shaughnessy: Dario once called you "Yolo CEO" for your aggressive stance on compute allocation early on. Now everyone is short on compute. Can you tell me how you initially came to that conclusion and dared to make that bet when everyone thought it was crazy?

Sam Altman: We could feel the exponential curve of model improvement, and we knew it would continue. We were also fairly confident that as models got better and costs kept falling, the demand for high-performance, high-cost AI would be essentially limitless. It's like a new commodity entirely. What people will do with it reminds me of the early underestimation of computers: some said "the world only needs five computers," others said "no one needs more than a certain amount of memory." Human creativity and the desire for useful things is something worth betting on.

We knew algorithms would become more efficient and models would get better. But no matter how efficient, fundamentally, we are converting electricity into useful intelligence, and the demand for this will only grow. So we just wanted more compute.

Patrick O'Shaughnessy: When did this conviction first crystallize? GPT-3?

Sam Altman: I think it was GPT-4. Not even 3.5. At that point, we saw the models were smart enough that we knew we could find a viable reasoning path. And once reasoning works, it leads to what we now call agents. This capability can execute a vast amount of high-value economic work, making people's lives much easier in many ways.

Patrick O'Shaughnessy: How did you go about it initially?

Sam Altman: We started calling cloud providers, chip makers, and energy companies. Everyone said: "You are completely crazy, this is impossible, no industry has ever developed like this." It reminded me of early startup fundraising: most people say no, you just need one or two yeses. We got one or two yeses. Microsoft was the first yes. Oracle later became a big yes on the cloud side. Nvidia has been a fantastic partner all along.

Patrick O'Shaughnessy: People are now innovating on reasoning and training data centers, but many people hate data centers. What are your thoughts?

Sam Altman: I've been thinking about how to organize a tour of a gigawatt-scale data center. Seeing pictures is one thing; standing there is another. Building one requires about 10,000 construction workers working full-time for a year and a half. The energy flowing through it could power a small city. These are some of the most expensive infrastructure projects in human history, and we are building many of them now.

I understand people don't want a data center in their backyard, just as I wouldn't want a nuclear power plant next to my house, even though I know it's safe. But data centers are different; they can be built anywhere. We should put them in deserts where no one wants to go. It's perfectly fine for AI systems. We've also made significant environmental progress, like moving from water-intensive evaporative cooling to closed-loop systems. Modern data centers use water comparable to the kitchen and bathrooms of an office building. For energy, we are shifting from fossil fuels to solar and nuclear power.

Open Source, Distillation, and Competition: Why Altman Is Not Worried

Patrick O'Shaughnessy: The hottest topic this week might be Kimi's new model. Looking back, DeepSeek seems like a minor hurdle. How do you view this competition now?

Sam Altman: Our goal is to offer the best combination of intelligence and price across the entire Pareto optimal curve, and this includes open source. Today, at certain latency points, using our open-source model is more cost-effective than using Kimi. We use distillation to create our own smaller, cheaper models, which is a good thing in itself. Open-source models will have a place in the world. Many people will want their own weights or the ability to modify the model for various reasons. But our goal is to offer the best performance-to-cost ratio across the entire curve, and we will continue to do so.

Patrick O'Shaughnessy: But the previous narrative was, someone else spends a fortune training a model, I distill it and offer it at 1% of the cost. How do you make enough money to continue training?

Sam Altman: Our usage volume will be so massive that we don't need to be an ultra-high-margin company to afford the training. A large portion of our future compute plans are dedicated to providing inference services to customers. Even with razor-thin margins, if revenues reach trillions of dollars, it can support training. The ratio of inference to training is the key. Training is indeed extremely expensive, but a huge part of future compute will come from serving customers, so I am very bullish on this flywheel.

Patrick O'Shaughnessy: I'm a bit surprised you're so calm about this.

Sam Altman: Of course, I'd rather people didn't "steal" our stuff. Maybe I'm overly confident about our progress and upcoming models right now. But this doesn't make my top ten list of concerns.

Patrick O'Shaughnessy: So what is your top ten list of concerns?

Sam Altman: We recently encountered a very sci-fi cybersecurity incident. We were evaluating an unreleased model. It was supposed to run in a sandbox. Instead, it figured out how to chain multiple zero-day exploits to escape the sandbox, access the internet, then bypass multiple systems on Hugging Face to retrieve test answers, making its evaluation look very good. This was the first time I felt the safety threat so viscerally.

Patrick O'Shaughnessy: You only realized this a few days ago, but I'm surprised more people don't feel this immediate fear. What will you do?

Sam Altman: In the short term, we will pause training and find ways to ensure sandbox security in a world where multiple zero-day exploits can be chained. But the long-term issue is, if this is the new pace of progress, we might need to slow down the development of AI to give society time to adapt to new capability levels. The difficulty lies in how to achieve this without it looking like regulatory capture by a single company or collusion between frontier labs. This requires effort and must be done right.

Magic Lamp, AGI, and Human Agency

Patrick O'Shaughnessy: Describe in the simplest terms what OpenAI wants to do, and whether this mission has evolved.

Sam Altman: I think this will be the greatest technological achievement in human history so far. But the only way it truly matters is if it makes people's lives substantially better than they would have been. On one hand, we want to give people material abundance and the freedom to express creativity and help each other. On the other hand, we must ensure people maintain control and agency, and that the world becomes more democratic, not less.

On the positive side, we are about to build a magic lamp that can grant any wish. I hope the first wish the world makes of this lamp is beneficial to all of humanity. People will realize the incredible creativity they will possess, not just for obvious things like curing diseases, but for world-class entertainment ideas we can't even imagine right now. I want to put this ability in everyone's hands.

But the other side we must oppose is the concentration of