Compiled by | AGI 2050

Editor's Note

This dialogue originates from the well-known tech business podcast "Figuring Out with Raj Shamani." In this in-depth discussion, Altman reviews many life-and-death decisions made within OpenAI: Why did he insist on launching ChatGPT despite unanimous opposition and a mere 5% internal test retention rate? At the secret point right after GPT-4 was trained, how did he gamble the entire company’s fate on computing power? Why does he believe that "the vast majority of entrepreneurial advice is noise"?

This episode is from the Figuring Out podcast, aired on November 3, 2026, with an original video duration of 68 minutes.

Below are selected dialogue excerpts, totaling approximately 8,200 words, with a reading time of about 8 minutes.

Highlights from This Episode

During ChatGPT's internal testing, the retention rate was a pitiful 5%, but I disregarded the majority and pushed it live. Because the fervent love of a few is far more important than the "it's fine" sentiment of thousands.

Almost all scholars once mocked the large models as merely predicting the next word based on existing materials, claiming they could never create new knowledge. Their biggest mistake was failing to realize that the models acquired a general "reasoning ability" amidst massive predictions.

In August 2022, right after training GPT-4, we made a decision that seemed suicidal to the entire industry: despite the world not even having seen what it looked like, we committed the entire company to massively procure computing power.

If your first version of a product doesn’t make you feel a hint of shame or embarrassment, it means you released it too late.

Many former colleagues of OpenAI have left to start their own ventures, driven by a common mindset: when the stakes of artificial general intelligence are immense and directly related to humanity’s future, they trust only themselves to accomplish this.

Listening to 800 episodes of business podcasts will absolutely not transform you into an excellent entrepreneur. The vast majority of outside advice is noise; the only underlying principle is this: create what people truly want.

What many people call 20 years of experience is merely mechanically repeating the first year's experience 20 times. The real gap in life is created by continuously pushing yourself to do uncomfortable, non-obvious new things, keeping you on the exponential compound interest curve.

On AI Bias: Why is India, not Silicon Valley, the most optimistic about AI globally?

Question: If faced with someone who knows nothing about you, how would you explain who you are in one minute?

Altman: I am a tech enthusiast, deeply curious about the world, and a firm believer—I believe artificial intelligence is touching the lives of all of us and will bring about mostly positive changes. Outside of work, I am a father, deeply love my children. That’s how I define myself.

Question: Were you like this as a child, always following your curiosity? Did you feel insecure when you were younger?

Altman: Almost every child experiences insecurity, and I am no exception. I was a typical bookworm, terrible at sports, and during my growing-up years, being good at sports was the coolest thing. I had many friends and got along with people, but I was never one of the cool kids at school.

Question: Did this sense of being on the fringes trouble you?

Altman: A little, but it didn’t completely crush me like it does some people. I’ve seen many peers who bear the deepest childhood imprint of feeling "I don’t fit in, I’m not cool, I hate this feeling." I felt a bit frustrated, too, but I had my computer and a small group of close friends; overall, I was very happy and fulfilled during my childhood.

Question: An interesting phenomenon is that when discussing AI with ordinary people in the U.S., they often carry a defensive or even pessimistic attitude; but in India, people’s attitude towards AI is extremely positive and full of enthusiasm. Where does this stark emotional contrast come from?

Altman: This difference is indeed quite remarkable. In terms of user numbers, India has become our largest single market globally, even bigger than the U.S., and its growth is incredibly rapid. You can feel a strong sense of "energy to create things" there.

I recently heard about an Indian student who needed to raise funds for his school; he had no prior relevant experience, but he used our set of tools to complete the entire process from start to finish.

Why is there such a contrast in emotions? Many people have observed a pattern: countries that are rapidly developing and have economies that are racing ahead have a strong demand for products and services, even facing severe supply shortages; these countries are often more welcoming to AI than developed economies. I think that’s one reason.

The Harsh Truth of the Scale Law: The Data Was Plainly Before Everyone; Why Did Only OpenAI Dare to Bet Back Then?

Question: Many foreign media have recently portrayed you as a controversial figure. You were the first to ignite this technological revolution, yet in your own company, OpenAI, you didn't hold any shares, and suddenly, many started to label you as some sort of “villain.” Why did this change occur?

Altman: In American culture, there’s a natural distrust of those wielding great power, especially deep skepticism about the supreme power held by the tech industry. Objectively speaking, this oversight is somewhat healthy, but I think it has gone a bit overboard now.

As for not taking equity in the company, looking back, that was a mistake. I thought I had good reasons at the time, but it completely confused the outside world. In the real world, if an artist pours their heart into crafting a piece or an athlete trains hard without purely for monetary gain, it seems natural and earns admiration; but once you bring it into a business context, if you create a company without taking a dime in equity, people think it's very strange. And the public inherently rejects things they do not understand; they endlessly speculate: “If he doesn't want money, then what does he want?” For me, this was simply because I believed it was the most worthwhile and important thing to do in the world, and participating in it itself was a tremendous fortune.

Question: You once wrote: “You should use a short-term low social status to exchange for long-term high status, but most people are unwilling to do so. As long as you are right, being misunderstood by the vast majority is actually your advantage.” The issue is, before the final outcome is clear, how do you determine whether you are a “misunderstood genius” or “just being foolish”?

Altman: The painful part is, before the dust settles, you can never be completely sure. What you can do is build a high-belief confidence based on the evidence you have and continuously adjust your probabilities as new signals emerge. This uncertainty itself is part of the risk.

In 2015, when we just started OpenAI, the social status of this endeavor was actually very low. Everyone knew that if successful, it would be extremely valuable, but very few took us seriously: General artificial intelligence? That might be something that won’t be achievable for 100 years, or even forever. The outside world saw us as a bunch of inexperienced kids without a proper academic background, met with a flood of mockery. But we calculated the odds, and recognized it as a highly asymmetric bet worth taking.

Question: What was the underlying signal that supported your firm belief that this would succeed?

Altman: I’ve always felt that in life, you can only grasp a very few foundational great insights and repeatedly apply them in different unknown fields. For me, one of the most crucial insights is: “Scale has a shockingly magical power.”

This is not empty speculation. Whether in the early Y Combinator, the expansion of mega cities, the evolution of super multinational corporations, or in personal career leaps, I have repeatedly witnessed this phenomenon: as long as a system can run successfully on a very small scale, pushing it to extreme scale often releases superlinear, exponential returns.

This is also true in the field of deep learning. Years before we fully grasped the precise "scale law," the early AlexNet had already demonstrated this momentum: simply making the neural network bigger would elevate its capabilities. I had a strong intuition: we had to push up parameters and computing power significantly.

What perplexed me the most was: during 2015, 2016, and even up until 2018 to 2021, why did the top minds in the industry, looking at the exact same data, not feel the shock like we did? Why didn’t they immediately say, “This is crazy, we must bet everything on it!” For a prolonged period, we internally felt like we were crazy, because the answer was plainly before us—why was the whole world blind to it?

Question: So, what mental shackles impeded other top scholars from seeing this?

Altman: Because it seemed “too simple” at the time, too simple to be believable.

The elites in academia scoffed at the naive mistakes made by neural networks, saying, “How could such a clumsy and fragile system possibly solve all problems just by increasing computing power and data?” At the time, the smart people insisted on sticking to the “symbolic representation” route, and their counterarguments sounded impeccable.

I remember someone told me back then why our path was doomed to fail; their logic sounded tightly knit: “What these large models do is merely to statistically predict the next word based on the training dataset. This means they can never create ‘new knowledge’ since they can't surpass the boundaries of the training set. You show them a million math problems; they might learn to solve similar problems, but they can never derive a new theorem that humanity has never proven.”

At the time, this reasoning seemed very convincing. But today, we all know how absurd that view was. They overlooked the most crucial point: the important thing these giant models learned under massive next-word predictions is not a specific math problem, but a universal, generalizable reasoning ability. And reasoning can be extended to entirely new areas.

The Gamble on GPT-4’s Computing Power: How to Differentiate “Top-Level Vision” from “Complete Delusion”?

Question: Take me back to that specific time point: everyone around you was advising you to back down, but you made a decision that was crucial for OpenAI’s survival with half certainty and half what some would call “delusion.” When was that moment?

Altman: Around August 2022, we had just finished training GPT-4.

At that time, it was about seven or eight months before we officially released GPT-4 to the world. We kept it locked on an internal network for daily testing, everyone understood it was extraordinarily powerful. But even so, it took us a while to fully comprehend what it meant—what new research directions it would open, how quickly the model would improve after GPT-4, and the thinking chains and reasoning models we later developed started from here.

At that juncture, we made a fateful decision for the company: despite the world not having seen GPT-4 yet, and the whole industry not understanding what it could accomplish, we decided to gamble the entire company’s future by rushing to buy massive computing power globally.

From the outside, this appeared to be a completely irrational and insane action. But we had enough internal confidence because we had personally used the model, and we had ingrained the mathematical logic of scale effects deeply within us. If we had hesitated even a little and chosen a more conservative option, we would have slowed down the investment in computing power. If GPT-4 ultimately turned out to be merely a minor improvement at the level of GPT-3.8, we would have never had the courage to make such a big bet.

Question: Given your current position, how do you ensure you’re not surrounded completely by echo chambers, just people saying “yes boss”?

Altman: Many people in high positions ultimately fail because they are surrounded by people who are afraid to tell the truth. You must have an extreme desire for genuine opposition.

My approach is: first, try to keep old friends around me, retaining those who have worked with me for over ten years in the core circle; second, establish a clear culture within the company—when senior decisions are foolish, those who can strongly oppose should be rewarded and respected.

But this requires extremely proactive maintenance. The more times you are proven right in reality, and the greater the resources and power you hold, the harder it becomes to hear an honest truth.

Question: But if your old partners, mentors, and extremely intelligent companions all point at your nose saying “you are completely mistaken," in that moment, what makes you think you can stick to your viewpoint?

Altman: I have actually rarely pushed through decisions that contradicted the consensus of the entire team.

I have seen many entrepreneurs have a pathological obsession where they treat “the whole world says I am wrong, so I must do it” as some kind of hero badge. This way of thinking is extremely dangerous. My principle is: unless absolutely necessary, never easily go against the collective wisdom of the core team.

If you have gathered a group of the smartest people in the world with countless efforts, you should listen to their judgments most of the time. If you decide to gamble recklessly in the face of opposition, you must be acutely aware that it's a risky bet, extremely expensive, and likely to lead to downfall. Unless you have unbreakable foundational first-principles support, never foolishly seek to be heroic.

Non-Consensus Betting and Product Minimalism: Why Release ChatGPT When Retention Rate Was Only 5%?

Question: Since the cost of opposing consensus is so high, how should ordinary entrepreneurs find their own non-consensus opportunities?

Altman: When we established OpenAI in 2015, there were very few labs globally that explicitly targeted AGI as a goal; it was perhaps just us; yet in the same year, there might have been tens of thousands of startup teams scrambling to create photo-sharing software.

At that time, most people thought creating photo-sharing apps was a much better idea. But the problem is, there are a thousand smart people fighting it out on the same track, plus the network effect of the giant Instagram is unbreakable.

As a founder, your most critical moat is to find those “seemingly unlikely to succeed, but once achieved hold incredible value” asymmetric levers. Many people's mindsets come from exam-oriented education—throughout school, we are taught how to answer others' good questions, almost never how to pose an unprecedented good question. Learning to ask good questions is my best method for finding non-consensus opportunities.

You can conduct a test: find an extremely intelligent, passionate friend with rigorous thinking and try to explain the non-consensus logic you see. If you can present solid facts and reasoning, ultimately changing their mindset to accept “this is risky but definitely worth doing,” that indicates your idea is rooted, not self-indulgent delusions.

Question: In OpenAI's history, has there been a decision that almost everyone opposed, but you pushed through based on your intuition?

Altman: The most representative example is the launch of ChatGPT.

At that time, most other executives didn’t want to release ChatGPT. It wasn't that everyone opposed creating a chatbot; rather, the majority felt it was too early and not good enough.

But I insisted we must launch it immediately. This entirely came from the product instinct I summarized during my time at YC: “If you feel the slightest bit of shame about your first version of a product, it means you released it too late.” Dreaming in the lab is meaningless; you need the real-world beating and feedback.

Even more exaggerated, the early prototype we tested internally had frighteningly poor real retention data: the retention rate was around a miserable 5%. This data itself had some inflation, as we had better models and other product interfaces in hand. But many people seeing that number would say a product with only a 5% retention rate shouldn’t even launch; it’s too terrible. Usually, that’s true.

But if you talk to that 5% of users, you’ll find some who really love it. Although the retention for that internal prototype was very low, the people who liked it used it extensively.

This is also the core principle that YC taught me: for a completely new exploratory product, having a small group of fervent fans who love you is worth far more than a large group of people who think you’re "pretty good." As long as you capture that real passion, even if the product is extremely rough, you must throw it to the market immediately.

The Paradox of Power in Superintelligence and Executive Departures: “The Stakes Are So High That Everyone Only Trusts Themselves”

Question: You once wrote: “The vast majority of people actually possess the ability to distort reality, but they have never even attempted to use it.” How do you understand this ability?

Altman: The vast majority of people spend their whole lives doing everything possible to stay within the “safety zone of group consensus.” This stems from the biogenetic instincts developed by humans over hundreds of thousands of years—in ancient times, being unique or exiled from the tribe often meant death. But like many tech people, I naturally feel a bit less of that inherent pressure to “conform to the collective trajectory.” In other contexts, this may be a flaw; but when pioneering hard-core technology, this insensitivity to others’ opinions becomes a critical trait.

Question: As OpenAI has grown larger, we’ve also seen many regrettable rifts: Sutskever started his own venture, Dario Amodei founded Anthropic, and Mira Murati also left. They all started at the same table; why did they ultimately go their separate ways?

Altman: When the potential impact and historical stakes of a technology are nearly infinite, the minds of many top geniuses inevitably produce an extremely strong psychological mechanism: “Precisely because this thing concerns humanity's future and the stakes are so high, at critical moments, the only person I can fully trust is myself.”

In my view, the rivalry over paths and trust issues are interconnected. It’s probably a mindset like: “His path is wrong, so I can’t allow power to concentrate in his hands.”

Question: After experiencing board controversies, lawsuits, government investigations, and the scrutiny of the whole internet, has your personal mental state changed? Do you often feel anxious or even out of control?

Altman: If you want to be involved in creating superintelligence, you can absolutely not expect a tranquil life; I have long accepted this reality.

I now have an extremely high threshold for triggering extreme stress; storms that feel earth-shattering to the outside world are just background noise to me. Most of the time, I find it hard to feel a rush of panic.

The last time I was really nervous enough to have a racing heart was when we were preparing to announce the Navier-Stokes equations (one of the Millennium Prize Problems). I felt that this was insane; this was the real deal: something humans have never been able to achieve, and now AI has done it. I was indeed very nervous at that time; it may have been the most recent moment that truly made my heart race.

There was a time when I meditated a lot, several hours a day. Although I don’t have as much time to sit now, that training created a physiological instinct: whenever I encounter an extremely chaotic and terrible situation, as long as I close my eyes for a few seconds, I can immediately force my brain back into a state of complete calmness.

The Minimalist Methodology of a Startup Mentor: Why Do 99% of Startup Advice Count as Garbage?

Question: Many entrepreneurs spend their first day racking their brains over "moats" and "industry monopolies." You also questioned teams about how they would form monopolies at YC. Is this the most important thing in the early stages of entrepreneurship?

Altman: Absolutely not. It doesn't even rank in the top three.

The core issues always boil down to three extremely simple points: what exactly is your product? Why do users desperately need it? How do you plan to acquire more real users?

The so-called defensive moats and long-term barriers are merely post-hoc means to ensure that after the business scales up, profits are not wiped out by competitors; it's part of commercial consideration but must not take precedence.

In those years at YC, I encountered a myriad of successful founders. If I had to summarize the core qualities evident within half an hour, I would prioritize the following:

1. Extreme clarity of expression: Can they articulate incredibly complex logic in plain language clearly and effortlessly?

2. Top-level charisma: After a half-hour chat, do I have a feeling of "I want to work with him"? Early hiring, fundraising, and client acquisition all rely on this raw infectiousness.

3. Rapid cognitive iteration speed: How quickly can they overturn old assumptions and propose new ideas? In the face of continuous dire circumstances, can they always smile and chew through the tough bits?

Question: Can a person's ambition be developed through later training?

Altman: Many people think ambition is innate, but my most surprising discovery at YC was that ambition can indeed be learned later. It may not apply to everyone, but it can definitely be achieved.

The methods to train ambition are very simple and effective: first, throw yourself into a circle of peers all making insane breakthroughs; second, continuously face a series of progressively challenging brutal tests. Once you repeatedly climb mountains you thought you could never overcome and achieve success, the cognitive boundaries within you will be permanently expanded, and your ambition and confidence will be forged at a truly healthy compound rate.

Question: What is the worst piece of advice you’ve heard for today’s young entrepreneurs?

Altman: The vast majority of outside advice for entrepreneurs is garbage.

If I had to give entrepreneurs one core meta piece of advice, it would be: “Stop listening to so-called advice.”

Listening to 800 episodes of business podcasts or taking entrepreneurial tips from various masters will never transform you into an excellent entrepreneur. The general core advice distilled from so many unicorns at YC boils down to just five to ten points, and the most unshakable first commandment is simply this: “Create what people truly want.”

Apart from these almost common-sense truths, every entrepreneur faces deadly situations that are extremely specific; what you need is to take off your suit and jump into the mud, and carve out the path for yourself in the real world.

Compound Life and End Goals: Using Superintelligence to Conquer Physics and New Sciences

Question: You once wrote an article titled “13 Rules for Success,” with “compound accumulation” being the top rule. In today’s hyper-accelerated technological era, how should ordinary people understand compounding?

Altman: There’s a harsh but insightful joke: “Many people claim to have 20 years of work experience, but in reality, they’ve just mechanically repeated the first year's experience 20 times.”

A life truly possessing compounding effects is one where every stage's cognition and ability achieve geometric leaps based on the very high base of the previous stage. But to do this, you must always maintain a mindset that seeks self-inflicted challenges, forcing yourself to continually do those extremely uncomfortable and non-obvious new explorations that could upend your existing understanding at any moment.

In the real world, the vast majority of people lay flat too early in their careers and actively slide off the exponential curve.

Even today at OpenAI, I force myself to overturn existing assumptions every day. Many people misinterpret the so-called “firm belief” as blind obstinacy, but I believe true firmness is when you try to overturn yourself, rigorously challenge your logic from every conceivable angle, and if your existing understanding still stands, then you pursue it wholeheartedly without reservations.

Question: Setting aside specific business battles, in your lifetime, what is the ultimate result you personally most wish to create?

Altman: The outcome I most urgently want to see is new scientific discoveries driven by AI.

All the improvements in living standards, elimination of poverty, and accumulation of wealth over the past few hundred years in human society are fundamentally based on breakthroughs in foundational scientific understanding.

If there’s a machine that can help humans completely understand the underlying laws of the physical world, conquer deep theoretical mathematical problems, derive entirely new physics models, find new therapies, discover new energy sources… this acceleration of fundamental scientific understanding is the ultimate lever that can truly reshape the fate of the human species. Everything we do today is paving the way for that machine.

Question: One last question: as the “Sam Altman” in the spotlight today, bearing the expectations and controversies of the entire trajectory of human technology, what has been the real cost for you personally?

Altman: Subjectively speaking, this is my entire life’s current status.

If there’s a trivial but real regret, it’s that I have forever lost the right to “live in the world as an anonymous ordinary person.” I can no longer sit alone in an unremarkable café on a sunny Saturday afternoon, ordering a large cup of coffee, opening my laptop, and distractedly browsing the web, staring out the window at passersby—if I tried to do that today, I’d have to have two security personnel sitting at the next few tables. I used to immensely enjoy that pure simplicity and freedom, but those have now become memories that I cannot return to.

But I have no sense of victimhood whatsoever. Being at the forefront of human technology, actively participating in building such a grand future, I feel incredibly fortunate. This luck is enough to wash away all the fatigue, scrutiny, and storms swirling around me. This is the life I have chosen.

Source: “Figuring Out” podcast, Raj Shamani dialogues with Sam Altman, aired on November 3, 2026, original video on YouTube

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到[email protected]，本平台相关工作人员将会进行核查。