Ultraman 68 Minutes Deep Talk: From a 5% Retention Rate Strongly Pushing ChatGPT to Betting the Entire Company on Computing Power

Compiled by | AGI 2050

Editor's Note

This issue's dialogue comes from the well-known tech business podcast Figuring Out with Raj Shamani. In this in-depth conversation, Altman reviews many life-and-death decisions made within OpenAI: Why did he push for the launch of ChatGPT despite universal opposition and a mere 5% retention rate during internal testing? How did he gamble the entire company's fate on computing power at a secret juncture right after GPT-4 was trained? Why does he believe that "the vast majority of startup advice is noise"?

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

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

Highlights of This Issue

When ChatGPT was in internal testing, the retention rate was a pitiful 5%, but I pushed for its launch against all odds. Because the fervent love of a very small number of people is far more important than the "it's okay" sentiment of thousands.

Almost all scholars mocked large models for merely predicting the next word based on existing data, claiming they could never create new knowledge. Their biggest mistake was failing to see that the model inadvertently acquired a general "reasoning ability" through massive predictions.

In August 2022, just after training GPT-4, we made a decision that seemed like suicide to the entire industry: we bet the entire company on massively purchasing computing power before the world had even seen what it looked like.

If your first version of a product doesn't make you feel a bit ashamed or embarrassed, it means you released it too late.

Many former colleagues at OpenAI left to start their own ventures, driven by a common psychology: when the stakes of general artificial intelligence are so high that they concern the future of humanity, they only trust themselves to do it.

Listening to 800 episodes of business podcasts will never turn you into an excellent entrepreneur. Most external advice is noise; the only fundamental principle is: create what people truly want.

Many people’s so-called 20 years of experience is just a mechanical repetition of the first year’s experience 20 times. What truly differentiates lives is the continuous push to do uncomfortable, non-obvious new things, staying on the exponential compounding curve.

On AI Bias: Why Is India More Optimistic About AI Than Silicon Valley?

Q: If you were to explain who you are to someone who knows nothing about you in one minute, how would you do it?

Altman: I am a technology enthusiast, curious about this world, and a firm believer—I believe artificial intelligence is touching all our lives and will mostly bring positive changes. Outside of work, I am a father who loves my children dearly. That’s how I define myself.

Q: Were you like this as a child, always following your curiosity? Did you ever feel insecure as a child?

Altman: Almost every child feels insecure at some point, and I am no exception. I was a typical bookworm as a child, terrible at sports, and during my growing-up years, the coolest thing was to be good at sports. I always had many friends and got along with people, but I was never the cool kid in school.

Q: Did this sense of marginalization trouble you?

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

Q: An interesting phenomenon is that when talking about AI with an average person in the U.S., there’s often a sense of caution or even pessimism; but in India, people have an extremely positive and energetic attitude towards AI. Where does this stark emotional contrast come from?

Altman: This difference is indeed astonishing. In terms of user numbers, India has become our largest single market globally, even larger than the U.S., and it’s growing extremely rapidly. You can feel a strong "energy to create" there.

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

Why is there this emotional contrast? Many have observed a pattern: countries experiencing rapid economic growth have an extremely strong demand for products and services across all industries, even facing severe supply shortages. These countries welcome AI more than developed economies. I think that’s one reason.

The Harsh Truth of the Scale Law: Why Did Only OpenAI Dare to Bet Back Then?

Q: Many overseas media have recently portrayed you as a controversial figure. You were the first to ignite this technological revolution, yet at OpenAI, you held no shares, and suddenly, many began to see you as a kind of "villain." How did this happen?

Altman: In American culture, there is a natural wariness towards those who hold immense power, especially regarding the supreme power held by the tech industry. Objectively speaking, this oversight is somewhat healthy, but I think it has become a bit overly paranoid now.

As for not taking company equity, looking back, that was a mistake. I had what I thought were good reasons for doing so, 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 chasing money, people find it very natural and admire it greatly; but once it’s in a business context, if you start a company without taking a penny in equity, people find that very strange. And the public instinctively rejects what they don’t understand, endlessly speculating: "If he doesn’t want money, what does he really want?" But for me, it was simply because I believed this was the most worthwhile and important thing to do, and being part of it was a tremendous privilege.

Q: You once wrote, "You should trade short-term low social status 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 question is, before the final outcome is clear, how do you judge whether you are a "misunderstood genius" or "just being foolish"?

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

When we started OpenAI in 2015, the social status of doing this was actually very low. Everyone knew that if it succeeded, it would be extremely valuable, but almost no one took us seriously: general artificial intelligence? That might be something that could never be achieved for 100 years or even forever. The outside world thought we were just a bunch of inexperienced kids without academic credentials, facing widespread ridicule. But we calculated that this was a highly valuable asymmetric bet worth taking.

Q: What underlying signals supported your firm belief that this could succeed?

Altman: I have always felt that in life, one might only grasp a few truly foundational great insights and then apply them repeatedly to different unknown fields. For me, the most important insight is—"scale has a shockingly magical power."

This is not wishful thinking. Whether in the early days of Y Combinator, the expansion of mega cities, the evolution of super multinational companies, or the leaps in personal career paths, I have repeatedly witnessed this phenomenon: as long as a system can run smoothly at a very small scale, once you push it to the limit, it often releases super-linear exponential returns.

The same goes for deep learning. Years before we figured out the precise "scale law," early AlexNet had already demonstrated this momentum: as long as you make the neural network larger, its capabilities will step up. I had a strong intuition: we must push parameters and computing power to the limit.

What puzzled me the most was: during 2015, 2016, and even until the long years from 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 long time, we even felt like we were crazy within the company because the answer was clearly on the table, yet the whole world just turned a blind eye.

Q: So what mental shackles prevented other top scholars from seeing this?

Altman: Because this method seemed "too simple" at the time, simple enough to be unbelievable.

Academics looked at the naive mistakes made by neural networks and scoffed, saying, "How could such a clumsy and fragile system possibly solve all problems just by increasing computing power and data?" At that time, those smart people insisted on taking the "symbolic representation" route, and their rebuttals sounded flawless.

I remember someone telling me why our path was doomed to fail, and their logic sounded airtight: "What these large models do is merely predict the probability of the next word based on the training set's data. This means they can never create 'new knowledge' because they can never step beyond the boundaries of the training set. You show it a million math problems, it might learn to solve similar ones, but it can never derive a brand new theorem that has never been proven by humans."

At that time, this argument sounded very convincing. But today we all know how absurd that view is. They missed the most crucial point: the most important thing these giant models learned through massive predictions of the next word is not a specific math problem, but a general reasoning ability with universal generalization capability. And reasoning ability can extend to entirely new domains.

The GPT-4 Computing Power Gamble That Determined the Company's Fate: How to Distinguish "Top Vision" from "Complete Delusion"?

Q: Take me back to that specific moment: everyone around you advised you to back off, but you made a decision that would determine OpenAI's life or death with half confidence and half what some might call "delusion." What was that moment?

Altman: Around August 2022, we had just completed the training of GPT-4.

At that time, we were about seven or eight months away from officially releasing GPT-4 to the world. We had it locked on the internal network for daily testing, and everyone knew it was extraordinarily powerful. But even so, it took us some time to fully understand what it meant—what new research directions it would open, how quickly models would advance after GPT-4, and the reasoning models we later developed started from here.

At that juncture, we made a decision that would shape the company's future: without the world even having seen GPT-4 and the entire industry having no idea what it could achieve, we decided to bet the entire company’s life on globally acquiring massive computing power.

In the eyes of the outside world, this is completely a mad act of losing reason. But we have enough internal confidence because we have personally used this model, and we have ingrained the mathematical logic of scale effects into our bones. If we had taken a step back at that time and chosen a safer option, we would have slowed down the pace of computing power investment. If GPT-4 ultimately turned out to be just a minor improvement similar to GPT-3.8, we absolutely would not have the courage to place such a big bet.

Question: Given your current position, how do you ensure that you are not completely surrounded by a group of sycophants who only say "yes, boss"?

Ultraman: Many high-ranking individuals ultimately stumble because they are surrounded by people who dare not speak the truth. You must have an extreme thirst for genuine dissent.

My approach is: First, try to keep old friends around, keeping those who have worked together for more than ten years in the core circle; second, establish a clear culture within the company—when senior management makes foolish decisions, those who can stand up and vehemently refute should be rewarded and respected.

But this requires extremely proactive maintenance. The more often you are proven right in reality, and the greater the resources and power you possess, the harder it becomes to hear a blunt truth.

Question: But if your old comrades, mentors, and extremely intelligent partners all point at your nose and say "you are completely wrong," how can you still feel justified in sticking to your opinion at that moment?

Ultraman: The number of times I have pushed decisions against the consensus of the entire team is actually very few.

I have seen many entrepreneurs with a pathological obsession, treating "the whole world says I am wrong, so I must do it" as some kind of heroic badge. This mindset 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 around you, most of the time you should listen to their judgment. If you decide to go all in despite everyone opposing you, you must be fully aware that this is an extremely expensive and high-risk bet that is likely to lead to your downfall. Unless you have unbreakable foundational first principles to support you, never easily play the hero.

Non-consensus Bets and Product Minimalism: Why did the first version of ChatGPT have a retention rate of only 5%, yet still force a release?

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

Ultraman: When we first established OpenAI in 2015, there were very few laboratories worldwide that clearly identified AGI as a goal, possibly only us; yet in the same year, there were likely tens of thousands of startup teams frantically clustering to create photo-sharing software.

In the eyes of most people at that time, creating a photo-sharing app was a much better idea. But the problem is that there were a thousand smart people fighting in the same arena, and the network effects of the giant Instagram were unassailable.

As a founder, your most core moat is to find those "seemingly very difficult to succeed, but once achieved, have incredible value" asymmetric levers. Many people's mindset comes from exam-oriented education—throughout their lives, schools only teach us how to solve good problems posed by others, and almost never teach us how to ask unprecedented good questions. Learning to ask good questions is the best way I have found to identify non-consensus opportunities.

You can do a test: find an extremely smart, passionate friend with very rigorous thinking, and try to explain the non-consensus logic you see to them. If you can present solid facts and reasoning, and ultimately change their perception, making them acknowledge "this is risky but indeed worth doing," it indicates that your idea is grounded, not a self-indulgent delusion.

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

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

At that time, most of the other management did not want to release ChatGPT. It wasn't that everyone opposed making a chatbot, but most felt that it was still too early and not good enough.

But I insisted that it must be launched immediately. This completely stemmed from the product instinct I summarized during my time at YC: "If you do not feel a bit of shame about your first version product, it means you released it too late." Sitting in the lab and daydreaming is meaningless; you need the real world's harsh feedback.

Even more exaggerated, the early prototype we tested internally had extremely poor real data: the retention rate was only a pathetic 5%. This data itself had a lot of fluff, as we already had better models and other product interfaces. But many people seeing this number would say that a product with a retention rate of only 5% cannot be launched; it is too terrible. Usually, that is indeed the case.

But if you talk to those 5% of users, you will find that some of them genuinely love it. Although the internal prototype had a low retention rate, those who liked it used it a lot.

This is also the core rule that YC taught me: for a completely new exploratory product, having a small group of passionate loyal users who love you is far more valuable than having a large group of ordinary users who think you are "pretty good." As long as you capture this genuine enthusiasm, even if the product is rough, it must be thrown into 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, "Most people actually have the ability to distort reality, but they have never even tried to use it." How do you understand this ability?

Ultraman: Most people spend their entire lives trying their best to stay within the safety zone of "group consensus." This stems from the biological instinct developed over hundreds of thousands of years of human evolution—in ancient times, being different or being ostracized by the tribe often meant death. But like many people in technology, I naturally feel less of that internal pressure to "conform to the collective trajectory." In other contexts, this might be a flaw; but in pioneering hardcore technology, this insensitivity to others' opinions has become the most critical trait.

Question: As OpenAI has grown larger, we have also seen many regrettable fractures: Sam Altman has gone independent, Dario Amodei founded Anthropic, and Mira Murati has also left. Everyone started from the same table; why did they ultimately go their separate ways?

Ultraman: When the potential impact and historical stakes of a technology are so great that they approach infinity, many top geniuses inevitably develop an extremely strong psychological mechanism: "Because this concerns the future of humanity and the stakes are so high, at critical moments, the only one I can fully trust is myself."

In my view, the conflict over direction and trust issues are interconnected. It is roughly this kind of thinking: "His direction is wrong, so I cannot allow power to concentrate in his hands."

Question: After experiencing boardroom ousting, lawsuits, government investigations, and scrutiny from the entire internet, has your personal psychological state changed? Do you often feel anxious or even out of control?

Ultraman: If you want to participate in creating superintelligence, you can never expect to live a calm life; I have long accepted this reality.

My threshold for facing extreme pressure has been raised very high; the storms that the outside world finds earth-shattering are just daily background noise to me. Most of the time, I find it very difficult to feel a panic that makes my heart race.

The last time I was truly nervous to the point of my heart racing was when we were preparing to announce the results of the Navier-Stokes equations (one of the Millennium Prize Problems). I felt that this was too crazy; this was serious: this was something that humanity had never been able to accomplish, and now AI had done it. That time I was indeed very nervous; it was probably the last time that truly made my heart race.

I once spent a lot of time meditating, several hours a day. Although I don't have that much time to sit now, that training formed a physiological instinct: whenever I encounter extremely chaotic and terrible situations, I can immediately force my brain back into a completely calm meditative state just by closing my eyes for a few seconds.

The Minimalist Methodology of an Entrepreneurial Mentor: Why do 99% of entrepreneurial advice suck?

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

Ultraman: 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 moats and long-term barriers are merely post-hoc measures to ensure that after the business scales up, profits are not wiped out by competitors; it is part of business considerations but must never take precedence.

During my years at YC, I have seen all kinds of successful founders. If I had to summarize the core traits that can be seen within thirty minutes, I value these aspects the most:

1. Extremely clear expression ability: Can they explain extremely complex logic in simple terms clearly and effortlessly?

2. Top-notch charisma: After a half-hour conversation, do I feel an impulse of "I want to work with them"? Because early hiring, fundraising, and acquiring clients all rely on this raw infectiousness.

3. Extremely fast cognitive iteration speed: How quickly can they overturn old hypotheses and propose new ideas? In the face of continuous dire situations, can they always smile and tackle tough challenges?

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

Ultraman: Many people think ambition is innate, but my most surprising discovery at YC was that ambition can indeed be learned. It doesn't work for everyone, but it can be done.

The method for training ambition is very simple and effective: First, throw yourself into a peer group where everyone is making crazy breakthroughs; second, continuously face a series of brutally escalating challenges. When you repeatedly overcome mountains you once thought were absolutely impossible to climb and achieve success, your internal cognitive boundaries will be permanently expanded, and your ambition and confidence will be forged at an extremely healthy compounding rate.

Question: So, what is the worst advice you have heard for today's young entrepreneurs?

Ultraman: The vast majority of entrepreneurial advice from the outside world is garbage.

If I had to give entrepreneurs one core meta-advice, it would be: "Listen less to those so-called suggestions."

Listening to 800 episodes of business podcasts and hearing various masters impart entrepreneurial skills will never turn you into an excellent entrepreneur. The universal core advice distilled by YC from so many unicorns boils down to just five to ten points, and the most unshakeable first commandment is just one sentence: "Create what people truly want."

Beyond these almost common-sense truths, every entrepreneur faces extremely specific deadlocks; what you need is to take off your suit and jump into the mud, carving your own path in the real world.

Compounding Life and Ultimate Goals: Using Superintelligence to Conquer Physics and New Sciences

Question: You once wrote an article called "13 Rules for Success," with "compounding accumulation" at the top. In today's fast-paced technological era, how should ordinary people understand compounding?

Ultraman: There is a harsh but piercing joke: "Many people claim to have 20 years of work experience, but they have merely mechanically repeated the experience of the first year 20 times."

A truly compounding life is one where every stage of cognition and ability achieves geometric growth based on an extremely high baseline from the previous stage. But to achieve this, you must always maintain a self-imposed mindset of seeking hardship, forcing yourself to continuously engage in those extremely uncomfortable, non-obvious, and potentially paradigm-shifting new explorations.

In the real world, the vast majority of people prematurely settle down early in their careers, actively sliding off the exponential curve.

Even today at OpenAI, I force myself to overturn existing assumptions every day. Many people misunderstand the so-called "firm belief" as blind stubbornness, but I believe true firmness is when, after every attempt to overturn yourself and rigorously challenge your logic from various extremely tricky dimensions, if your current understanding still holds, then you can wholeheartedly commit without reservation.

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

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

The improvement in living standards, the eradication of poverty, and the accumulation of wealth in human society over the past few hundred years are essentially all built on breakthroughs in fundamental scientific understanding at the grassroots level.

If there were a machine that could help humanity completely see the underlying laws of the physical world, tackle deep theoretical mathematical problems, derive entirely new physical models, find new therapies, discover new energy sources… this acceleration of fundamental scientific understanding would be the ultimate lever capable of reshaping the fate of the human species. Everything we do today is paving the way for this machine.

Question: One last question: As the "Sam Altman" who stands in the spotlight today, bearing the expectations and controversies of the entire trajectory of human technology, what is the true cost you have personally paid?

Sam Altman: Subjectively speaking, this has become the entirety of my current life.

If there is some trivial yet real regret, it is 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, order a super-sized coffee, open my laptop, and aimlessly browse the web while staring blankly at the pedestrians outside… If I were to do that today, there must be two security personnel sitting at the tables next to me. I once greatly enjoyed that pure simplicity and freedom, but those have become memories that I cannot return to.

However, I have no sense of self-pity or victimhood. Being at the forefront of human technology and personally participating in shaping such a grand future makes me feel incredibly fortunate. This luck is enough to wash away all fatigue, scrutiny, and storms. This is the life I have chosen.

Source: "Figuring Out" podcast, Raj Shamani in conversation with Sam Altman, aired on November 3, 2026, YouTube original video