OpenAI Ultraman talks to YC President: In the future, 1 person + 10,000 GPUs will have the opportunity to create a company with annual revenue of billions of dollars

Recently, YC President Garry Tan engaged in a conversation with OpenAI CEO Sam Altman about entrepreneurship, where Sam emphasized that the current development of AI technology is far from reaching its peak.

Sam suggested that entrepreneurs actively embrace AI as a new technology, leveraging it in the early stages of startups to achieve rapid breakthroughs, thereby defeating large enterprises constrained by planned cycles through speed, focus, and adaptability, thus achieving differentiation and a competitive edge.

When discussing the development path of AGI, Sam described the stages of evolution from "chatbots" to "agents" and then to "innovators." He predicted that future technologies will endow AI with stronger innovative capabilities, enabling it to explore complex phenomena over the long term and generate scientific breakthroughs, thereby driving technological advancement.

Sam encouraged startups to utilize the innovative power granted by AI to achieve a small team, high-efficiency operational model. He envisioned a future where companies with fewer than 100 employees, or even just one person with powerful computing capabilities, would emerge, providing unparalleled opportunities for startups.

Garry Tan: Do you think we will have companies with annual revenues in the billions but fewer than 100 employees? Perhaps 50 people, or even 20, or just 1 person?

Sam Altman: There is indeed that possibility. I don't know how to view this other than to say it is a very suitable time to be a founder of a startup. I really feel this is happening, like one person plus 10,000 GPUs.

Below is the dialogue between the two:

Sam Altman

We have said from the beginning that we want to pursue AGI, and at that time, you couldn't even mention it in this field because it seemed completely crazy.

Garry Tan

I remember you faced a lot of criticism for that.

Sam Altman

We really wanted to push this forward, and at that time our resources were far less than those of DeepMind and others. So we said, okay, they will try many things, and we have to choose a direction and focus. That way we can win here.

Most people still do not understand the value of holding quite extreme beliefs on a single bet. This is also why I am so excited about startups right now, because the world still maintains an astonishing indifference to these things.

Garry Tan

Today we have a truly special program. Sam Altman, thank you for coming. Thank you. This is actually a reboot of your initiated series "How to Build the Future." Welcome back to the series you started.

Sam Altman

That was many years ago. I was thinking, it was about that time. I'm glad to be here.

Garry Tan

Yes, it's been rebooted. That's right. Let's talk about your latest article on the "Intelligent Era." Do you think this is the best time to create tech companies? Sam Altman

At least it can be said that this is the best era so far. I hope there will be even better times in the future. I believe that with each major technological revolution, there are more things you can do than before.

I look forward to companies becoming more amazing and having a greater impact. So I feel this is the best era to date. When things progress slowly and are less dynamic, large companies have the advantage.

But when situations like this arise, such as the mobile internet, the internet, the semiconductor revolution, or even possibly tracing back to the industrial revolution, it becomes the advantage of emerging enterprises.

So yes, this kind of situation hasn't occurred for a while, which makes it very exciting.

Garry Tan

In the article, you actually mentioned something significant, which is that ASI superintelligence is still thousands of days away.

Sam Altman

Maybe. I guess that's our hope. In any case, that is indeed an exaggeration.

Garry Tan

Talk about that. I mean, it's big, really big.

Sam Altman

I can see a path where we continuously accumulate along our current work. If the pace of progress over the past three years can continue for three, six, or even nine years—let's say nine years is about 3,500 days—if we can maintain this rate of improvement or even accelerate it, then the system will be very capable of handling many things.

I believe that even systems like o1 are already able to demonstrate quite high cognitive intelligence in certain specific closed tasks. o1 is already very smart. I don't think we are near the limits of our progress.

Garry Tan

That's a shift in architecture that unlocks a lot. What I hear is that these things will continue to accumulate.

Sam Altman

We might encounter some unexpected bottlenecks or miss something. But for us, there seems to be a lot of accumulated potential ahead.

Garry Tan

I feel this article might be one of the most tech-optimistic pieces I've seen. There is so much we can look forward to, such as solving climate issues, establishing space colonies, exploring all of physics, achieving near-infinite wisdom, and abundant energy.

Sam Altman

I do believe that all these things, and perhaps many more that we can't even imagine, may not be far from us. And I find it exciting that we can even discuss these things half-seriously now.

What I love most about YC is that it encourages a slightly unrealistic technological optimism, believing that you can make it happen.

In a world that often tells people that something won't work, the early spirit of PG encouraging founders to think boldly is indeed a special existence.

Garry Tan

Abundant energy seems to be quite a significant unlocking point. You know, there might be Path A and Path B. If we really achieve abundant energy, not only knowledge work but also actual physical work can be unlocked. The combination of robots and language, and intelligence will bring about a truly abundant era Sam Altman

I believe these are two key inputs to achieving all other goals, of course, there are many other important factors. However, if we can achieve truly abundant intelligence and energy, the things we can accomplish in this world will be astonishing, not only being able to come up with better ideas faster but also turning them into reality in the physical world. Not to mention being able to run a large number of AIs, which also requires energy.

I think this is a huge unlock. As for whether I would be surprised by this or if it's just a natural result of the increasing rate of technological advancement, I'm not sure, but it is indeed an exciting time and a great time for entrepreneurship.

Garry Tan

Let's return to this abundant era. You know, maybe your robots can make anything, and almost all physical labor can bring substantial progress, not just for the wealthiest but for everyone. What if we can't unleash unlimited energy? What if there are physical laws preventing us from achieving that?

Sam Altman

The trajectory of solar energy and storage is good enough that even if we don't have major breakthroughs in nuclear energy, we can just barely get by. But there is no doubt that lowering energy costs and increasing energy abundance has a very direct impact on quality of life.

Ultimately, we will solve every problem in physics, so we will figure it out; it's just a matter of time. This is what we deserve. One day in the future, we might be discussing not nuclear fusion but Dyson spheres.

That would be great too. Yes, this moment makes us feel that abundant energy may seem far from enough in the eyes of our great-grandchildren. The universe is so vast, and there is so much more.

Garry Tan

I want to change the topic back to Paul Graham, whom you just mentioned. He brought us all together; in fact, he created Y Combinator. He likes to tell a story about how you got into YC, which was during your freshman year at Stanford.

He said this was the first batch of YC in 2005, and he said you were still a freshman, and YC would have another batch next time, so you should wait.

And you said, "I'm a sophomore, I'm coming." You are widely regarded as one of the most influential people in our community; where do you think that influence comes from?

Sam Altman

That story? I would be happy if it faded away.

Garry Tan

Oh, that's permanently recorded.

Sam Altman

In my memory, that day I needed to reschedule an interview or something, and then PG tried to get me to come back next year, and I said something more flattering, like "I'm a sophomore, I'm coming."

But you know, these stories often get exaggerated a bit. The funny thing is, I say this is not humility; I don't feel like I'm a person of prestige.

In fact, I feel like I'm really not in many ways. But I do find it a bit hard to understand why things have to be done in the existing way. So I just wanted to do some things that seemed right according to some basic principles I always find this a bit strange.

And I remember one thing I really admire about YC is that it brings together a group of unique individuals, all of whom are the kind of people who say, "I want to do my own thing." I believe that the part that aligns with self-awareness is that I truly think there are many times when you can do something or try something, and I believe the more, the better.

Garry Tan

That's true. I think both of us found a group of people at YC who believe you can do something.

Sam Altman

For a long time, I tried to figure out why YC is so special, and I think it might be because you have an incredible person telling you, "You can do it, I believe in you." As a young founder, that is indeed special and inspiring.

Of course, that is great. But it wasn't until much later that I realized what was truly special was a group of people who are also trying. Now, my biggest advice to young people is to find that group of peers as early as possible, which was extremely important for me, and I didn't realize it at the time.

I thought I could go it alone, but in fact, being with inspiring peers is really valuable.

Garry Tan

Interestingly, we both spent time at Stanford. I actually graduated, although I probably shouldn't have, but I did. You took the path of dropping out and giving back more.

But, you know, Stanford is said to be a community with these qualities, yet I am still incredibly surprised by the power that a room full of founders brings.

Sam Altman

I think the same way. I really liked Stanford, yes, but I didn't feel like I was surrounded by a group of people who made me want to be better and more ambitious.

Even if I sometimes felt that way, the competition was often about who could get which investment banking internship, and I regret to say I fell into that trap as well. That's the power of the peer group. After going through YC, I made the decision without hesitation to not go back to school.

Garry Tan

I particularly like a quote from Carl Jung: "The world will ask you who you are, and if you don't know, it will tell you." It seems very important to clarify early on who you want to be and who you want to associate with.

Sam Altman

That is definitely one of my insights; at least for myself, no one is immune to peer pressure, so the only thing you can do is choose excellent peers.

Garry Tan

Obviously, you later founded Looped, sold it, went to Green Dot, and finally collaborated with us at YC. Talk about the early stages of YC research; you brought this experimental project to YC.

I remember you coming back to the partner meeting room, talking about sitting in a room with Larry and Sergey, when AI was already a hot topic. Although it feels close, that was already ten years ago Sam Altman

I have always thought that the coolest job after retirement would be managing a research lab. At that time, it wasn't just limited to AI. When we started discussing YC Research, it ultimately funded many different projects.

I wish I could tell a story that I knew AI would succeed long ago, but in reality, we also tried many unsuccessful projects. At that time, I read some books about Xerox PARC and Bell Labs.

I felt that at that time, many people in Silicon Valley had a consensus that we needed to have good research labs again. I thought that would be really cool.

It's somewhat similar to the YC model, where you allocate funding to smart people, sometimes it succeeds, sometimes it doesn't, and I just wanted to try.

Indeed, at that time, AI was experiencing a small boom, roughly from late 2014 to early 2016, when discussions about superintelligence and the buzz around the book "Superintelligence" were happening.

Garry Tan

You both had that idea, right?

Sam Altman

DeepMind did achieve some impressive results, but the direction was slightly different. You know, I have always been interested in AI, so I thought, oh, trying to do something would be cool, but it's hard to say.

Garry Tan

Had ImageNet come out by then?

Sam Altman

It had come out, yes, for a while.

Garry Tan

Right, so you could determine if an image is a hot dog.

Sam Altman

Sometimes.

Garry Tan

That is indeed progressing. So, how did you identify the early people you wanted to involve in YC Research and OpenAI? For example, Greg Brockman joined very early.

Sam Altman

Looking back, it feels like a movie montage, like the scene at the beginning of a movie where you drive around looking for the right people.

Garry Tan

They would say, "I want to join."

Sam Altman

Exactly. For example, Ilia, I heard he was very smart, and then I watched some of his videos. He was not just smart, but a true genius and visionary, and he had a strong presence.

I saw one of his videos on YouTube and thought I had to meet this person. I emailed him, but he didn't reply, so I went to a conference where he was speaking, and we met.

Then we started chatting frequently. Greg, I knew a little, as he was an early member of Stripe.

Garry Tan

What was that conversation like? Was it "I really like your ideas about AI, and I want to build a lab"? Sam Altman

Yes. And from the very beginning, we said we wanted to pursue AGI. At that time, you couldn't even mention it in this field because it seemed completely crazy, even somewhat irresponsible. But it did catch his attention.

Garry Tan

It immediately caught his attention.

Sam Altman

It attracted the attention of all the talented young people, while those who were older but mediocre were critical of us. I think that was actually a good sign. We felt like a fearless group.

At that time, I was one of the oldest, around 30 years old. Everyone thought these young people were ignorant kids saying ridiculous things. And for those who thought, "I'm a sophomore, I'm coming," that kind of talk was exactly what attracted them.

So we moved around, meeting one by one or forming small groups. Over about nine months, the team gradually took shape.

Garry Tan

And then everything started.

Sam Altman

It began to operate. One of my favorite memories of OpenAI is that Ilia, due to some connections with Google, we announced in December 2015, but we had to wait until January 2016 to launch.

So, around January 3, 2016, everyone came back from the holidays, and we went to Greg's apartment, probably about ten people.

We sat in a circle, feeling like we had completed a huge task, and everyone was asking, "What should we do next?"

Garry Tan

What a wonderful moment.

Sam Altman

It reminds me of startup founders who feel they've succeeded after completing their Series A funding, only to realize they still need to figure out what to do next.

Garry Tan

It's not a time to celebrate with champagne; it's the starting gun just went off, and now it's time to run.

Sam Altman

And you have no idea how difficult this race is. It took us a long time to figure out what we wanted to do. But what I particularly admire is Ilia and all the early members; despite the twists and turns along the way, the initial grand vision was basically very correct.

We wrote down our goals on the whiteboard or flip chart in Greg's apartment, and later we did some other things, some effective, some ineffective, and eventually we ended up with the system we have today. Looking back, it's incredible how far we've come from there, going through many ups and downs.

Garry Tan

Was deep learning on that whiteboard back then?

Sam Altman

Yes, specifically a large-scale unsupervised model, and then solving RL (reinforcement learning). In an early team meeting, we had three goals: solve unsupervised learning, solve RL, and keep the team size under 120 people—we failed on the third goal. The predictions for the first two were very correct Garry Tan

So deep learning. The second big goal seems to be scaling up, and the idea of scalability was even considered heretical at the time, with many people feeling offended. I remember you received a lot of harsh criticism for this.

Sam Altman

Our core belief at the beginning was that deep learning was effective and would get better as the scale increased. I think that was a heretical belief at the time. We didn't know how systematically improvements would come with scaling until years later when we had the data to prove it.

Garry Tan

It started as intuition, and then came the data.

Sam Altman

But people already knew that if you made these neural networks larger, they would get better. We were quite certain of this before we even started.

At that time, there were always people saying, "Oh, this isn't real learning, this isn't real reasoning, it can't really do this." It felt more like a trick, and these were things said by the authoritative leaders in the field.

Not just saying "you're wrong," but saying "you're wrong, this belief is harmful." We just looked at these results, and they kept getting better.

Then we got the scaling results, which even broke my intuition. At some point, you had to look directly at the scaling loss graph and keep going; that was what we thought we should do.

It gradually felt like deep learning had some important emergent phenomena. Even though we didn't fully understand every detail in practice at the time, it was clear that there was something very fundamental here.

PG's saying was that we discovered a new square on the periodic table. So we really wanted to push in that direction, while our resources were far less than those of other teams like DeepMind.

So we said they would try a lot of things, while we would focus on one direction and concentrate our efforts, so we could win here, which completely aligned with the startup mentality.

So we said, even though we didn't know the future, we knew this thing worked, so we focused on it. Other teams might be trying to be too clever, while we just kept pushing in the direction in front of us.

The characteristic of scale has always attracted me, whether it's startups or deep learning models. I think it's an underestimated characteristic. If something seems to get better as it scales up, I think you should scale it up.

Garry Tan

I feel like people want things to be "less is more," but in reality, "more is more."

Sam Altman

We firmly believe in this, and we want to push it. I think one possibly lesser-known feature of OpenAI is that even when we were under the radar, we had a talented research team. If you have the smartest people in the world, you can push hard in a certain direction.

Garry Tan

You guys are motivated, or you created one of the few places in the world that can do these things. One story I hear is that even today, acquiring computing resources is still a crazy thing In the criticism from some "veterans" in the industry, some believe that you will waste a lot of resources, leading to an "AI winter," right? They think people will no longer provide resources, which is really interesting.

Sam Altman

People have never been sure whether we would waste resources or whether we are doing something vaguely unethical because too many resources are being invested, and resources should be spread across more bets rather than all concentrated in one direction.

Most people still do not understand the value of holding extreme beliefs on a single bet. So we say, well, we have this evidence, and we believe in this direction.

At that time, the conventional approach was to diversify investments, while we chose to focus. I think you will see this pattern repeatedly in many successful YC startups.

Garry Tan

Yes, that sounds right. When the world opposes you, and you feel that the opposition is unreasonable, you should stick to your path.

Sam Altman

In the world of entrepreneurship, I have witnessed this situation many times, and I am deeply grateful. Previously, I really believed that there was always a group of "adults" in the world who had everything under control and knew all the answers.

And if someone opposed you, they probably knew what they were talking about. Now I understand that no one can control everything, and no one will magically tell you what to do. You have to iterate quickly and find your own way.

Understanding this has been a huge life unlock for me. While it is important to hold onto beliefs, if you do not adapt to change and do not seek the truth, it will not be effective.

What we strive to do is to believe what the results tell us, to take each step steadily, even though there are many times when we hold beliefs but make mistakes.

But once we realize the mistake, we fully accept it. Belief is great, but once data points to right or wrong, you have to follow the data.

Garry Tan

So this is an iterative process, not about them being wrong and me being right; you have to show your work process.

Sam Altman

But for a long time, you have to act without data, and at that point, you can only rely on belief to move forward.

Garry Tan

It sounds like this also involves the issue of focus, such as you have to make choices, and that choice better not be wrong. Prioritization itself is an exercise that makes you more likely to succeed.

Sam Altman

I really wish I could tell you that we knew everything would turn out this way from the beginning. We had the idea of this language model from the start, and we went directly in that direction.

But clearly, the story of OpenAI is that we did a lot of things that helped us accumulate scientific understanding, but these were not all straightforward shortcuts. If we had known what we know now at that time, we could have greatly accelerated the process. But that is not the reality; you cannot guess right every time.

We had many assumptions at the beginning about the direction of technology, about what kind of company we would become, how AGI would develop, and so on. In these areas, we have been repeatedly confronted by facts But one of our advantages is that even when knocked down, we can get back up and keep moving forward. This applies not only to scientific choices but also to our thinking about how the world works and the forms of products.

To be honest, I didn't know that language models would become key; maybe Alec Radford did. At that time, we were working on robots, intelligent agents in video games, and so on. It wasn't until a few years later that GPT-3 was born, which wasn't obvious at the time.

Garry Tan

It sounds like you had a key insight about positive and negative sentiment even before GPT-1.

Sam Altman

Before GPT-1, yes. I remember that paper called "Unsupervised Sentiment Analysis," and I believe Alec did that work on his own.

Alec is an extraordinary genius; he did this amazing work and noticed that there was a neuron that switched sentiment polarity when generating Amazon reviews.

Other researchers might have exaggerated its value more, but Alec is Alec, so it took people some time to fully realize how important this discovery was. After that, he completed GPT-1, and others expanded it into GPT-2.

But this insight originated from an astonishing phenomenon: unsupervised learning basically didn't work at that time. However, he noticed a neuron switching between positive and negative sentiment, and this discovery ultimately led to the GPT series.

Garry Tan

Jake Heller, a YC alumnus from CaseText, I think he is one of the first to gain access to both 3.5 and 4.

He mentioned that gaining access to GPT-4 was a significant moment because GPT-3.5 would produce some hallucinations and couldn't be used in legal contexts.

And GPT-4 reached a level where if he broke down the prompts finely enough, he could get the results he wanted. He built a massive set of test cases around it and sold the company for $650 million.

I think he is one of the first to achieve success in the large-scale commercialization of GPT-4.

Sam Altman

I still remember that conversation with him. Indeed, the release of GPT-4 was one of the few moments that made me feel we had a truly powerful tool in our hands.

When we initially tried to pitch GPT-3 to founders, they thought it was cool, indeed doing some amazing things, and it was a fantastic demo, but there weren't many companies that could build great businesses based on GPT-3, except perhaps in copywriting.

Then GPT-3.5 came along, especially as startups began to show interest; it was no longer like pushing a boulder uphill. People actually started wanting to buy what we were offering.

Then came GPT-4, and what we encountered was the moment of "How many GPUs can you give us?" There was widespread demand shortly after its release. So we felt, well, we really have something good in our hands Garry Tan

So from the users' perspective, you knew right away after the model was released that it was indeed better.

Sam Altman

We were deeply shocked at that time as well. We designed a series of tests for it, and it performed exceptionally well, accomplishing some impressive tasks.

And when we tried it ourselves and saw the samples it returned, I thought, wow, it can do these things now. It can rhyme, tell some mildly funny jokes, and complete various other tasks; it really felt great.

However, you never truly know if the product in your hands is a hit until it reaches the customers. We were very excited and thought it was really good, but you only find out after real testing occurs.

Garry Tan

The real test is the users.

Sam Altman

So there was still some anxiety before that moment.

Garry Tan

I want to slightly change the topic. Before you created one of the craziest AI labs in history, you founded a company called Looped at the age of 19, which was basically a "find my friends' location" app, about 15 years ahead of Apple's similar feature.

Sam Altman

It was indeed too early.

Garry Tan

What attracted you to come up with that idea?

Sam Altman

I was interested in mobile phones, and I wanted to do something that could be used on mobile devices. At that time, mobile phones were just starting out, probably two or three years before the iPhone. But it was clear that having a computer you could carry around was very important.

Garry Tan

It's hard to imagine that phones back then were really just phones.

Sam Altman

Yes, they were really just phones.

Garry Tan

I even try to avoid using it as a real phone now.

Sam Altman

I remember the first time I got a phone that could access the internet; it had a terrible text browser and was extremely slow, making it painfully difficult to check emails.

I got it when I was in high school, at a time when phones could only make calls and send texts. I was fascinated; I thought this isn't a phone, it's a portable computer. We had to endure the legacy issues of the numeric keypad for a while, but the future would be great.

Garry Tan

Today, billions of people may not have computers, and for our generation, this was our first computer.

Sam Altman

Indeed, it's not the same computer experience, but my first computer was the LTTC2.

Garry Tan

Right, so that's what computers were like when we were kids. Now we walk around with this "black mirror," and technology has indeed come a long way. You've always had a keen insight into the development of technology, right? Sam Altman

I am a tech enthusiast, and I still am. I always think about these things on Friday nights.

Garry Tan

What made it even harder back then was that we didn't have the App Store, and the iPhone didn't exist. Yet you played an important role in that launch process.

Sam Altman

Although my role was small, it was indeed significant. It was a great experience for me, helping me understand how chaotic the platform transformation process can be, and how some small things you do can influence the future direction.

At that time, I was on the sidelines, observing others create this platform transformation, but experiencing and seeing how these things happened, how quickly things changed, and how to adapt was truly a valuable experience.

Garry Tan

Through that process, you eventually sold the company, which might have been your first time managing a team and doing a business sale. All of these were valuable lessons learned from that experience.

Sam Altman

I have to say that this company was clearly not successful, and it was a painful experience for me, but the speed of learning and gaining experience was incredible.

PG once said something, or quoted someone who said something, that has always impressed me: "Your twenties are always an apprenticeship, but you don't know what the apprenticeship is for until you really start working later on."

I did learn a lot, and I am very grateful. It was a tough experience; we never really found product-market fit, nor did we achieve escape velocity (self-sustaining growth), which is always difficult to do. I haven't heard of any other experience that provides as broad a learning opportunity as entrepreneurship.

Garry Tan

So in that sense, it was indeed good. When you were 19 or 20, you were right in the wave of the transition from regular feature phones to smartphones. Years later, your next phase even contributed to the birth of one of the major platforms.

Sam Altman

Yeah, time flies.

Garry Tan

But that's really the case now; young people at 18 or 20 choose to get degrees while also fearing they might miss the wave, because all the great things are happening now.

Garry Tan

It's proud to say that many of those who truly founded billion-dollar companies are not very clear about what's happening. It's like a survival-of-the-fittest situation.

Sam Altman

Yes, that's why I'm so excited about startups now, because the world still remains incredibly indifferent to all of this, while the founders at YC think, "No, I want to do this great thing, and I want to do it quickly."

Garry Tan

That reminds me of how Facebook almost missed the mobile internet. At that time, they were focused on building web software and were very good at it. They had to acquire Instagram, Snapchat, and WhatsApp. So it's interesting that platform transformations are always driven by young people who have no preconceived notions Sam Altman

Yes, I think that's great.

Garry Tan

Another interesting aspect is that people like you, Elon, and Bezos have all started their own founding journeys, whether it's Looped or Zip2, which could purely be software startups.

Then later on, they had opportunities for upgrades. Now, if someone says, "I want to start the craziest hard-tech project," would you advise them to go for it as much as possible? Or would you suggest they first address financial issues, like being able to invest more deeply in their next project?

Sam Altman

That's a very interesting question. It was indeed helpful that I could write early checks for OpenAI myself; I think it might have been difficult to find others to do that at the beginning.

Later, Elon did more, and I'm grateful for that. Others joined in as well, and there are some projects I've invested in that I'm very happy to support.

I don't think it would have been easy to find others to do these things back then. So financial independence is indeed a great advantage. And as we mentioned earlier, I learned extremely valuable lessons from early entrepreneurship.

But at the same time, I feel that spending time on Looped was somewhat a "waste." I certainly don't regret it; it's part of life, and I learned a lot.

Garry Tan

What different choices would you have made? If you could send a message to your 19-year-old self at Stanford through a time capsule, what would you tell yourself?

Sam Altman

It's hard to say because AI has always been what I wanted to do the most. I went to school to learn AI, and I was working in an AI lab at the time.

The only advice would be not to study neural networks because they had tried it and found it didn't work. That was a long time ago. I think I could have chosen a better project instead of Looped; I'm not sure what that would be, but everything has its own arrangement, and that's fine.

Sam Altman

Historically, many people have developed new technologies to improve people's lives, and I often think about this, like those who made my first computer.

I don't know them; many of them may have retired by now, but I am very grateful to them. They worked hard at the limits of technology, and I received a computer on my 8th birthday, which completely changed my life and the lives of many others.

They worked hard and never received my thanks, but I sincerely thank them. It's a wonderful thing to add a brick that belongs to us on this long road of progress.

Garry Tan

Yeah, it's been a great year for OpenAI, despite some ups and downs. Yes, we are indeed good at this. What did you learn from the "storm" last fall? What are your thoughts on some of the departures? I know the team changes over time, but are you okay now? Sam Altman

It has indeed been a very challenging good year. We are somewhat like a medium-sized or even large tech company that has completed a growth trajectory that typically takes ten years in just two years.

ChatGPT is not yet two years old, and this rapid development has brought about a lot of painful processes. Any company experiences different management teams as it scales, and those who excel at the 0 to 1 stage may not necessarily be suited for the 1 to 10 or 10 to 100 stages.

We are also constantly adjusting the company's direction, making many mistakes along the way, but we have also done some things right, all accompanied by many changes. I believe the company's goal is to continuously move towards AGI, making the best possible decisions at each stage.

However, this has indeed brought about a lot of changes. I hope we are moving towards a calmer period, but there will certainly still be dynamic moments in the future.

Garry Tan

So how is OpenAI operating now? The quality of your products and the speed at which you are bringing them to market have surpassed many mature tech companies, truly at a world-class level.

Sam Altman

This is the first time I feel that we really know what we should be doing. There is still a lot of work to be done from here to building AGI, with some known unknowns, but I feel we have a general idea of how to proceed. While it requires time and effort, it is indeed exciting.

I also believe there are some details to explore in terms of products, but overall we know the direction to focus our efforts and the optimization goals. With this clarity, you can indeed move faster.

We decided to concentrate on a few things and do our best at them. Our research path, infrastructure path, and product path are all relatively clear. Coordinating around these directions greatly helps to enhance speed.

For a long time before, we did not have this clarity; it was truly a pure research lab. Even with a direction, it was difficult to act decisively because there were always many other good things to try. Aligning everyone towards the same goal is indeed one of the key factors determining speed.

Garry Tan

It sounds like we have just completed the transition from Stage 1 to Stage 2, which is very powerful. We recently held a 0 to 1 Hackathon at YC.

That event was really exciting, and one of the teams that came in third was Camphor, a CAD/CAM startup.

They built a tool during the hackathon that could iteratively optimize a non-flying wing to a competitive lift level. This is somewhat like the fourth stage—the innovator stage.

Sam Altman

What you said is very interesting. I have been telling others that I think the leap from Stage 2 to Stage 3 will happen quickly, while the leap from Stage 3 to Stage 4 will be more difficult and will require some medium or larger new ideas. And that demonstration, along with some other examples, makes me believe that a lot of innovation can indeed be achieved by creatively using existing models Garry Tan

Yes, Camphor has basically built the underlying software for CAD/CAM, and your language model has become the interface for interacting with that software.

The use of such tools, combined with the ability to generate code, is a somewhat awe-inspiring idea, right? Not only can large language models write code, but they can also create tools for themselves and combine those tools, much like chain thinking.

Sam Altman

Yes, I believe the pace of development will be faster than people currently expect.

Garry Tan

It's an exciting time, and it's great to live in this era.

Sam Altman

You mentioned exploring all the questions in physics before. I once wanted to be a physicist, but I wasn't smart enough to become an outstanding one, so I can only contribute in other ways. But I truly believe someone will use these technologies to solve all the physics problems, and I'm very excited to be alive to see that.

Garry Tan

Can you briefly talk about stages 3, 4, and 5?

Sam Altman

We realize that AGI has become an overused term, and many people have different understandings of it. We try to outline what we think is the general order of development:

Stage 1 is chatbots;

Stage 2 is "reasoners," which we believe was marked by the release of o1 earlier this year;

Stage 3 is "agents," which can perform longer-term tasks, interact with the environment multiple times, proactively seek help when needed, collaborate with others, and so on. I think we will reach this stage faster than expected.

Stage 4 is "innovators," or scientists, who can explore phenomena that are not fully understood for extended periods, striving to understand the nature of problems.

Stage 5 is extending this capability to the scale of a complete company or organization, which will be very powerful.

Garry Tan

Indeed, this process is somewhat fractal. The effort to reach Stage 2 has some similarities to Stage 5, requiring multiple agents to self-correct and collaborate.

On a micro level, it does resemble an organization. Do you think we will have companies with annual revenues in the billions but fewer than 100 employees? Perhaps 50, or even 20, or just 1?

Sam Altman

There is indeed that possibility. I don't know how to view this other than that it is a very suitable time to be a startup founder. I really feel that this is happening, like one person plus 10,000 GPUs.

Garry Tan

It could happen, which is truly powerful. Sam, do you have any advice for those who are starting or about to start a business?

Sam Altman

Go with this technological trend. Betting on this trend, we are far from saturation. The models will get better, and the speed will be very fast As an entrepreneur, the difference between what you can achieve with this tool and without it is astonishing. Large companies, even medium-sized ones, and startups that have been established for several years are all bound by quarterly planning cycles; Google is even doing ten-year planning.

I don't understand how they can still operate, but your speed, focus, unwavering belief, and ability to adapt to rapid changes are the biggest advantages of startups, especially now.

Therefore, I would suggest everyone use AI to do something, fully leverage the dividends of this new technology, and when you see something new, build it immediately instead of dragging it into the quarterly planning.

Sam Altman

Another thing to mention is that when a new technological platform emerges, it's easy to feel that, well, because I'm doing something related to AI, the rules of business don't apply to me.

I have this magical technology, so I don't need to build a moat, I don't need to establish a competitive advantage or a better product because I'm doing AI, and you aren't, so that's enough.

Clearly, this is not true. But you can achieve explosive short-term growth by embracing new technology faster than others; just remember not to become complacent and still build something of lasting value. I think this is a point that needs to be kept in mind at all times.

Garry Tan

Yes, now everyone can build a fantastic presentation.

Sam Altman

Yes, everyone can build a fantastic presentation.

Garry Tan

But building a real business is what truly matters.

Sam Altman

That's right, the rules still apply. You can do it faster and better than ever, but you still need to build a real business.

Garry Tan

What are your expectations for 2025? What excites you about the future?

Sam Altman

AGI, of course, I'm really looking forward to it. I'm really excited about this. Also, the birth of my child is a bigger expectation for me than AGI; it might be the most thrilling thing in my life.

Garry Tan

Yes, it will definitely change your life completely. I can't wait either. Building a better world for us and our children, and hopefully for the future of the entire world.

Article author: Youxin Newin, source: Youxin Newin, original title: "OpenAI's Sam Altman in conversation with YC President: In the future, 1 person + 10,000 GPUs will have the opportunity to create a company with annual revenue of billions of dollars," with slight modifications