OpenAI CEO and COO's first joint interview: GPT-6 will become a universal tool, and the cost of AI in the future will approach zero

Recently, OpenAI CEO Sam Altman and COO Brad Lightcap were interviewed together on the VC20 podcast for investors. This is the first time Sam and Brad have jointly communicated with the public since OpenAI was founded over 8 years ago. During the interview, Sam and Brad shared their collaboration journey, starting from their initial interests and beliefs, to jointly driving the development of AI technology. Sam emphasized the potential of deep learning and the positive effects of scaling, while Brad, from an investment perspective, talked about his confidence in OpenAI and his expectations for the company's future development. Sam believes that in the future, computing costs will continue to decrease. They (OpenAI) can reduce the cost of very high-quality AI technology to near zero. However, he also mentioned that in the next 5 years, the biggest bottleneck for OpenAI may arise in the supply chain and computing resources. Sam also mentioned that current AI models are not yet smart enough, and people are still at a very primitive and limited stage when using GPT-4 for scientific research. However, he foresees that in the era of GPT-6, people will see it as a general tool that can help in various aspects. By the time GPT-8 arrives, people may be amazed at the wide range of tasks it can accomplish, perhaps even beyond our current imagination. Brad stated that AI technology is greatly needed in enterprise supply chains, and he hopes to integrate ChatGPT technology into business processes to achieve very quantifiable return on investment. Currently, by integrating ChatGPT into specific processes in supply chain management, expenditures in specific areas can be reduced by 20%. However, the adoption cycle for enterprises is slow, and the implementation of ChatGPT Enterprise Edition is still in progress. Both Sam and Brad agree that losing top researchers, research culture, or lacking sufficient computing resources will seriously hinder OpenAI's progress. They discussed the huge demand of artificial intelligence models for computing power, and emphasized that throughout the decision-making process, although there are only a few strategic decisions, successfully implementing these decisions requires making a large number of daily small decisions. It is worth mentioning that Sam mentioned that most members of the OpenAI leadership team are in their thirties and forties, unlike most leaders in other startups who are in their twenties and thirties. Similarly, their technical staff are slightly older on average. Sam believes that part of the reason is that the path to becoming an excellent researcher is very long . Below is the full interview translation compiled by "Hard AI": Host: Are you ready? I'm very excited about this, I've been looking forward to it for a long time. Is this the first time you two have been interviewed together? Host: This interview will be very unique. Sam, what made you believe in doing this seven years ago? Sam: I think there were two reasons. Firstly, I have always been interested in artificial intelligence since I was young, but there was no progress in this area during my university studies. However, when we started doing this, two factors seemed to be very important. First, deep learning seemed to be effective. Second, with the expansion of scale, the effect would be better. At that time, we were not sure how strong the predictability was, but it was obvious that the larger the scale, the better. This seemed to be something extraordinary. What puzzled us at the time was why others did not see this? Why didn't others get involved? But the reality was different. So we wanted to do this. Host: Can I ask, in the past few years, what made you persevere when others doubted? To be honest, few people have the same confidence. Sam: It seemed feasible to us, and we have been making progress. It was not blind faith, although if you want to do something difficult, you need a certain level of confidence. But for us, doing this was really important, and if we could do it, it would have a huge impact on the world in some way, and it might work. We have continuous data to prove that this approach is effective. Of course, it took a long time to figure out the specific details. You know, we didn't start out doing language models. Obviously, we knew that if we could continue to do things that were previously thought impossible, it signaled progress. For a long time, we had fundamental beliefs in this approach and attack vectors, and it took a long time to clarify the details, and colleagues made many brilliant discoveries. If we can do it, there is no doubt that artificial intelligence will be a big deal. It is very helpful. Like, it will be very valuable. We are becoming more and more confident in this approach, although for a while it felt like getting lost in the jungle or desert. And then, you know, if you believe in something and others doubt it, it can be a bit motivating. Yes, it can be a bit annoying, but it can also be a bit motivating. Host: I mean, venture capital firms may have opposing views, but that is not how we invest because we follow the crowd. But I do want to start talking about our relationship, because it is a very unique partnership. Again, we say this is the first time you two have been interviewed together. Brad, how did this partnership come about, why didn't you tell me? Brad: Of course. Well, I have been working with Sam for a long time, and we actually spent a lot of time at YC observing that batch of companies that were in the growth stage, all of which were very deep tech projects, such as nuclear fusion reactors, quantum computers, autonomous vehicles, satellites, and so on. From an investment perspective, I was more interested in these companies at the time. And OpenAI was the first company I saw that I felt was very unique, because it seemed to be getting better over time, unlike a binary risk of either winning or losing I remember telling Sam that I thought this company was different from the other companies we were looking at. Subsequently, I worked more with Greg and Ilia, witnessing the performance improvements brought about by the system scale growth described by Sam, which was initially unpredictable but later became predictable. I found this phenomenon to be very unique. I think we saw the same thing from different perspectives. From an investment perspective, if this materializes, it will be a significant investment outcome and will have a profound impact on the world. Therefore, I have been very confident about this early on and am willing to provide assistance wholeheartedly. Host: Are you planning to join full-time? When did you decide to make this a mission for the next few decades? Brad: Initially, I didn't have such plans. I was mainly helping Sam find a CFO. Sam: Brad actually started working full-time at OpenAI before me. Brad: That's true. This was the first time I beat Sam at something. But at that time, I was helping with recruitment. No one was willing to take on the role of CFO at OpenAI at that time, as the company was just a small non-profit research organization. I approached about 25 candidates, all of whom were rejected. To be honest, the reason I came here was that none of the 25 people were willing, which made me feel embarrassed. So I said, you know what? Why don't I help out in the evenings and weekends? Then I quickly went full-time. I spent about half of my time at OpenAI and half at YC. Host: Sam, when did you join full-time? Sam: I started working full-time at OpenAI. It was a gradual process. But I think around spring or summer of 2019. Host: Alright, Brad seems to have an edge over you in OpenAI. Sam: I think a sign of a great partnership is the complementary combination of skills. I am grateful to work with many key people at OpenAI. Of course, maybe Brad can do my job for a week. I definitely can't do Brad's job for a week. I believe that the ability to divide work as a team and establish very high-bandwidth communication channels with everyone and as a leadership team is crucial. Brad is good at many things. To save time, I will only mention two here. One is adaptability. Brad clearly joined the company to do finance, and now he's doing something different, which I think is like in the finance field but very, very different. We didn't have a business at all, or substantial business until recently. When we clearly knew we would have a very fast-growing business, I looked around and felt that we really needed someone. We needed someone to do this. I looked around the room and then asked Brad to do it. He said, "Okay, I'll figure it out. I may need some time to adapt, but I've dealt with similar business matters before, and I am capable of handling everything Brad can accept new challenges at all levels of the company and solve problems as he moves forward. However, I am completely clueless about finance, so all of this seems mysterious to me. Creating a new product category and marketing it requires a wide range of skills and great patience. It's like an obsession with customers, from products to business models, to how we handle customer support and all related matters. Brad can see the big picture and how it all fits together, that's what a company is. Today, we are here to participate in this corporate sales event. If you had told me a year ago that we would become a great organization, I might have been skeptical. Oh, we are not a great organization yet. If you said we would become a very good public company, I would have thought the likelihood of that happening was very low. But now, we are already a pretty decent organization. Host: We will discuss this issue later, as I think the IPO plan you have developed is incredible. However, from a different perspective, what do you think Sam's greatest advantage is, something that few people consider or know about? Brad: Alright, I guess some people might know this, but I think... I want to say two things that are interrelated. First, at any stage of a company's development, there are always one to three things that are truly important at that time. These things may change, but there are rarely 10 things that are truly important. I think Sam has an incredible ability to focus on these one to three important things, and this focus also affects the operation of our team. Because if I know he is focusing on something, even if we may have different views on these things, I usually think we are in agreement. But if we can at least agree on these things, they may not be the right global bets, but they seem right at the time, which helps me translate them into the team I am trying to build, whether we want to focus more on the enterprise, or we really want to change our bets on research, or we actually want to bet more on one thing than another, where we really need to get it right, this helps us move forward very quickly. I think this is the key to maintaining speed while scaling up, which most companies start to lose as the number of things and the perceived importance of things increase. The second thing is the long-term orientation. You have this idea that you are chasing a very distant goal, and in the process, justifying which one to three things are most important is actually trying to figure out the fastest way to get us to that point. Sam has a fanatical focus on that future world. My job is just to fill in everything in between. Host: Now, what are the one or two most important things for you? Sam: There are many AI organizations in the world that can replicate other people's work. Once you know something is possible, you know roughly what it looks like, and once you know people want it, it's not that difficult. Oh, this is a bit difficult. What's really difficult is trying something new for the first time and continuing for many years, if you're lucky, for decades, building a research institution, a product institution, and a whole company, pushing these things out into the world Because we are still innovating in our business model, this culture of continuous innovation is also very important. This way, we are not only making GPT-5 amazing, but also 6, 7, 8, etc., no matter what we will call them in the future. We won't keep numbering them like this. Ensuring that we are prepared for where researchers can take us, where products must go, and what the entire company must follow in their thinking is a big deal. Host: What is one of the biggest obstacles affecting the speed of innovation in OpenAI decision-making? Sam: I think we have the best researchers and the best research culture in the world. It would be very bad if we lost either of them. Not having enough computing resources would be very bad. I think we like to do cool research because scientific progress is the coolest and most exciting thing in the world. But in reality, we are here to do useful things for others. If we do the best research in the world, then do our best to make it efficient, but still do not have enough computing resources to serve everyone on Earth who wants to use it and will need to use these models more, that would be bad. So, the second thing I want to say is, think about how to get enough computing resources to meet the needs of people who want to use these services. Host: How did you approach answering this question? Sam: I might not answer that in front of the camera, but by considering it as a holistic system problem, I remain optimistic about the surprises we will bring to the world. Host: Can I ask, how do you both make decisions? How do you determine what can be delegated and what cannot be delegated? Brad: The most important thing is to reach consensus. We as an executive team and leadership team have spent a lot of time making decisions in this regard, sometimes it's obvious, sometimes not. All other things will be delegated down. So, I might make 10 decisions a day, and these decisions will not be handed over to Sam because they are not the most important. But if it really is the most important thing, our entire executive team will spend one meeting or several meetings discussing it. Host: Do you agree with the idea that a company only needs to make one or two decisions a year to succeed? Or do you lean more towards the view that making 10 decisions a day, all these small incremental decisions will eventually make progress for the company? Sam: I strongly believe both are important. One thing I like about being an investor is that this job actually involves making one or two decisions a year, or one or two decisions every ten years. By the way, the role of an operator is definitely not my instinct, it's not my natural place in this world, but one thing I've learned to do slightly better is that indeed, there are only a few strategic decisions. It feels more like once or twice a month, rather than once or twice a year, but not that many, not that big, just like what decision to make here. However, there are many ways to make decisions I think those who claim that CEOs make only one or two decisions a year or a month have never tried to run a complex company before. It's truly non-stop. However, there is a difference between making big decisions like whether to do ChatGPT or not. And then, to make that decision successful, we have to make 10,000 small decisions throughout the process. Host: Why do you think you are not an operator? Sam: To be honest, I'm not. I'm very happy being an investor. For me, it's not a satisfying job. But it's a very interesting job. And, I kind of like that all the things people use to make fun of investors are somewhat true. It's a good balance for quality of life work. But yes, I'm not naturally an operator. I'm very willing to do this because I really love OpenAI. I think AGI will be the most important thing I've ever been involved with. However, it's not for me. It's quite interesting to hear that OpenAI is the fastest-growing company. I believe Brad would agree. Brad: Yes, I absolutely agree. Host: I want to ask, we mentioned the element of computation. How do we consider the situation where marginal revenue exceeds marginal cost in terms of marginal cost and marginal revenue? I think this is a question that many people suggest we discuss today. Especially for our products. How do we view this issue? Sam: To be honest, I think this is the most boring of all the things we can talk about. No offense, this is the most boring question I can think of. Host: Really? Why is it boring? Sam: You just need to believe that the cost of computation will continue to decrease, and as model performance improves, the value of artificial intelligence will continue to rise. This equation is actually quite intuitive and not complicated to solve. Of course, there is a possibility of errors, such as if the cost of computation does not decrease as expected for some reason, or if computational power becomes unusually expensive due to factors like supply-demand imbalance, improper planning, etc., then the situation will be different. But I believe we can bring the cost of very high-quality AI technology close to zero, which will be a remarkable transformation for most fields in the world. Not everything will be negatively affected by this, but I believe the cost of AI is about to become very low. Host: How does the rise of open source further achieve or impact this? Sam: The world will have a place for open source models. Some people want these models, some want hosted services, and some use both. I think these details, while interesting, overlook a bigger picture: we are in a true, quite large technological revolution, where intelligence is transitioning from something very limited. In the past, only smart people had intelligence, and if you wanted to do something that required high intelligence, you needed a lot of smart people to collaborate For example, to create something like OpenAI, you need a large number of intelligent people, a lot of them. Imagine the entire system, not just OpenAI's employees, but also the people involved in manufacturing chips, building data centers, and so on. Eventually, everyone will be able to access rich and affordable intelligence to do amazing things. Host: Have we overestimated the pace of application for a year and underestimated people's future willingness? Sam: That's very possible. Because I think this actually reflects a profound insight into the general application of technology. No matter how amazing something is, social inertia is a big issue. Only really great things will be widely adopted, but this process takes time. So, for cool new things, people always have high expectations at the beginning, and then the enthusiasm wanes after a few years. Therefore, I may have indeed overestimated the pace of application. I think there will be a quick reversal between expectations and reality. Brad: Currently, people's expectations are very high, but the reality is still pretty bad. Honestly, these models are not that great. But I think once people get in touch with the current models, their expectations will quickly drop. Then, these models will soon become very outstanding, and you will see a reversal between expectations and reality, where people's expectations suddenly need to catch up with the speed of improvement in reality. Host: You mentioned that the actual quality of the models may not be as good as expected, and the gap between expectations and reality. Another interesting but perhaps boring question is the commercialization of models. We've never seen anything like this before, last week Mistral was all the rage, and then there were other things popping up, the media always reports on the rise of new players, it feels like things are changing every week. Does this mean that models are becoming a commodity? Sam: I remember there were over 100 car companies in the United States, or at least close to that number. If you look at some old media reports from that time, you will see statements like "Now there are better cars emerging", "Now there are better cars emerging". I think most emerging industries will go through the same process. I think it's okay, and maybe even a good thing, but I don't think that's where the lasting value lies. In the end, there will be a reshuffle, leaving only a few dozen providers offering models on a large scale. This will be a very complex and expensive engineering feat. I hope all of us can continue to compete with each other to make models better, cheaper, faster, and in a sense, achieve commoditization. Long-term differentiation will not exist in the basic models, just like intelligence is just a new emerging property of matter. Long-term differentiation will be in the model that best suits you, one that has your entire life background and seamlessly integrates with all the other things you want to do. But for now, the curve of progress is so steep that what we should focus on most is continuously improving the basic models. **Host: You mentioned your experience as an investor, and Brad has also interacted with many large global companies today. As an investor, I have seen many AI companies, but I have not invested in any application-oriented AI companies. Because honestly, I saw OpenAI launching some products themselves, which felt like killing the entire industry Sam: I believe there are two strategies for the current development of artificial intelligence. One strategy is to assume that the model performance will not improve, and then build various additional functions on this basis. The other strategy is to assume that artificial intelligence research institutions like OpenAI will continue to iterate rapidly, and models will continue to improve. In my opinion, 95% of startups worldwide should adopt the latter strategy, but unfortunately, many startups have adopted the former strategy. When we continuously improve models and their tools, it is possible to challenge those companies that adopt the former strategy, which is where the idea of "OpenAI killing my startup" comes from. However, many startups can benefit greatly from the significant improvement in GPT-5 performance. If you bet on the progress of artificial intelligence and use it as an investment logic, you will most likely succeed. Host: For investors, how can they identify which companies will not be eliminated and which companies may be eliminated? Brad: We can ask these companies if they are excited about improving model performance by 100 times. Typically, companies that often tell us they want the next model, inquire about the release date of new models, and want to be the first to try them out belong to the type that can benefit from it. There are also companies that have never shown interest in this. I think a very good criterion is that if a company can clearly explain how stronger artificial intelligence will accelerate the development of its products, then they belong to the type that can benefit from it. Host: Is Klarna an example of this? Sam: Klarna is a good example. For Klarna, if the next model performs as well as we expect, they will make huge profits. Take the example of medical advisors, although current artificial intelligence models still have some shortcomings, they are still very useful in some aspects. If the models make greater progress in these areas, Klarna can expand more business. They may pressure us to improve the models faster to save more lives and help people who previously could not access medical services. Host: I listed some questions earlier, but then realized that these questions were not ideal. Now I want to talk about the rate of model improvement. Is this rate linear? Or will there be a bottleneck period? The rate of model improvement is clearly accelerating now, so how will this speed develop in the future? Sam: From an external perspective, the rate of model improvement seems to be leapfrogging. This indicates that we are not doing well enough on a core belief. We believe that iteratively releasing models is very important, avoiding secretly developing Artificial General Intelligence (AGI) in the laboratory. The worst case scenario is that after decades of effort, we suddenly release a general artificial intelligence that catches the world off guard. For us, a better approach is to release models to the world, giving people time to think, react, and gradually increase societal engagement with artificial intelligence I believe one of the most important decisions we have made is to release models like ChatGPT to the world, making the world take advanced artificial intelligence seriously. We have tried to discuss this topic before, but with little success. The launch of ChatGPT has indeed been effective. However, when I look ahead to future models, I think we have underestimated their potential. Because we have become accustomed to these models, watching them improve little by little. We underestimate how powerful some of these models' capabilities can become even with an iterative release strategy. Therefore, when conceptualizing the next model, we are working to make the release process smoother, making the smoothness felt by the external world closer to what we feel internally. Host: As the company grows, do you think the iterative deployment strategy is still possible to move forward? You see, Farron Llama published some articles on medical science writing and faced strong opposition, leading to a withdrawal. Bard also did their thing, and their stock price dropped by 8%. As the company grows, releasing imperfect products may have such consequences. Is this iterative deployment still possible over time? Sam: Setting the right expectations is important, and I believe as long as expectations are set correctly, iterative deployment is still viable. Brad: Yes, I agree with that. We also learn a lot. For example, when we released "Sonata," we received a lot of feedback from the creative community, media, and industry. We are now starting to incorporate this feedback and integrate it into our research roadmap for this particular model. So, in a sense, we initially set our expectations very low, just trying to learn, listen to external voices, and then absorb these suggestions as much as possible. This way, when we really want to share a product, it will feel practical and people will naturally be familiar with it, even feeling like it was tailor-made for them. I think this will become a mode of operation for us, a truly iterative mode, and collaboration with the world will be even closer than people imagine. Host: Finally, I want to talk about the Go-To-Market (GTM) strategy. You previously mentioned medical consultants. I heard you are passionate about how artificial intelligence can address cancer, especially certain medical issues... Sam: More accurately, I am eager for artificial intelligence to provide assistance. It may not be able to completely solve the problem of cancer, but it can significantly accelerate the rate of scientific progress, and conquering cancer is a great example. I do think science is amazing, and there is certainly a part of personal excitement in it. But I genuinely believe that scientific progress is the most important part of social progress, economic growth, and improving the quality of life for everyone. If artificial intelligence can help people significantly accelerate the pace of scientific progress, I believe it will be hugely successful. Host: What do you think is the biggest obstacle to achieving this goal? Sam: I think the biggest obstacle lies in the lack of intelligence in the models. This may sound a bit frustrating, like a superficial answer with insufficient information, but I believe this is the fundamental reason. As long as the models are intelligent enough, other issues will be easily resolved. We will need to find various ways to integrate these tools into people's workflows. Of course, the capabilities of models in different fields will also play a crucial role. From a macro perspective, using GPT-2 for scientific research was once considered quite impractical. However, now, although people are still at a very primitive and limited stage when using GPT-4 for scientific research, I can foresee that in the era of GPT-6, people will see it as a universal tool that can provide assistance in various aspects. By the time of GPT-8, people may be amazed at the wide range of tasks it can accomplish, perhaps even beyond our current imagination. Host: Can we now discuss the company's development? I think this is very important. Historically, the scale of OpenAI's company development can be said to be unprecedented, especially considering the rate of revenue growth. Brad, you have always been a standout in this regard. It's not a very good question, but so far, how has OpenAI achieved such good development scale? What is the secret to success, and why does everything seem to be holding up? Brad: Things don't always go smoothly, but I appreciate your recognition that at least on the surface, OpenAI hasn't fallen apart. The launch of Chat GPT was a turning point, where people truly experienced the human side of this technology for the first time. We have been hearing a variety of stories about different use cases, which has been quite surprising. For example, sometimes you hear a research scientist at a company talk about how this tool has improved their work efficiency, and then you hear a software engineer at an XYZ startup say this tool helps them write code. There are even new parents who say they ask this tool 80 questions every day to help them understand how to take care of their baby. The fact that the same tool can support such a diverse range of use cases, and I believe it is user-friendly, will certainly have a significant impact on people's acceptance and usage. Of course, this will also have commercial implications, but we remain focused on continuously striving in this field. B2B business obviously has a different pace from consumer-oriented business, as enterprises often have longer adoption cycles for new technologies. We have achieved great success in the developer community. We have always prided ourselves on building a world-class artificial intelligence developer platform. Now, enterprise users have become our new focus. Therefore, developing products for enterprise users will be a process that requires more processes and timelines, but we are excited about it. Host: I would like to ask about talent. Is it not good if talent joins OpenAI just because it is the hottest company? Is OpenAI the fastest-growing company? Sam: Maybe. Host: So, does everyone have to join our mission? Sam: It's not a good thing if talent only wants to join because OpenAI is the hottest and fastest-growing company. Everyone should identify with our mission. We always emphasize the mission, but is it enough to just emphasize the mission? I have indeed seen some tech companies attract talent just because they are popular employers, but this approach usually has negative consequences. As you said, a sense of mission doesn't need to be 100% true in all cases. However, companies that lose their mission orientation and are dominated by a mercenary culture often regret it. Host: You have invested in some outstanding founders. Are there any of them who have been role models for you and shaped your ideas about building a company? Sam: I am very fortunate to work with many outstanding founders of my generation and witness their success. I am also glad that they are now willing to spend time helping me. Host: Can I boldly ask? Are there one or two founders who have particularly impressed you, and what have you learned from them? Sam: Chessie has been very helpful to me in the past year and a half. He is good at many areas where I am not, forcing me to quickly learn how to think about products, discuss products, and build excellent products. He is truly a special person. The Carlson brothers are also great. Every time I talk to them, I gain profound insights that I have never had before, all of which are non-linear ways of thinking. I have invested in many companies, so I know many outstanding founders, and I am very grateful that they are willing to help in different ways. Just as I try to learn from different investors, learning from different founders is also a great strategy. Host: Can we go back to the topic of user usage? You mentioned a variety of user groups such as consumers, parents, and scientists. You have also established amazing partnerships with some globally renowned large companies. In terms of enterprise applications, what do you think are the most important insights about enterprises adopting artificial intelligence, and how do large companies think, approach, and adopt AI technology? Brad: I think the biggest insight is that companies naturally have a strong desire to directly integrate this technology into their business processes in order to achieve quantifiable return on investment (ROI). Host: Sounds great. Brad: I manage my supply chain, spending X amount each year, and I want to integrate artificial intelligence into specific processes of supply chain management and reduce expenses in specific areas where I spend money by 20%. Things like that. It's good. We are here, happy to help you think about this. However, I think people seriously underestimate the importance and returns of simply enabling employees to use this technology. Although you cannot precisely quantify the benefits of this way of working, for example, a task that used to take two days to complete now only takes two minutes, freeing up more time for employees to do the other 85 daily tasks This will not be directly reflected in the way companies measure return on investment. But imagine if this were to happen to 10,000 or 100,000 employees, what impact would it have? Host: How should we explain this to companies? You're right, it's not as straightforward as a budget detail where you can say, for example, we saved X dollars. Brad: Yes, it is indeed difficult to demonstrate the time saved. On the one hand, it does take time to prove. ChatGPT as a commercial product is still very new. We only released the enterprise version around August or September last year, and the Self-Service Product (SRF) was released earlier this year. So the time on the market is almost zero, and companies often take longer to adopt new technologies. So I think part of it is the need for time, and another part is that employees expect to use these tools. Additionally, in the future, you will start hiring talent accustomed to using these tools, and they expect to use these tools in the workplace as well. Therefore, I believe that over time, we will begin to see this shift. However, I think there is a strange misconception among people about where artificial intelligence should be deployed, which will have a significant impact on where I think they should deploy artificial intelligence. Host: Do you think the biggest companies are not asking the right questions they should be asking? Brad: Yes, about how to use artificial intelligence, how to integrate artificial intelligence, the questions they should consider. Many companies see this as a static technology. Many companies think GPT-4 is the best model they can get. This can be understood. Because every technology they have adopted in the past has been relatively static. For example, think about the iPhone in 2009 and today's smartphones, they are essentially the same, just with slight changes in appearance, faster speed, higher resolution, but the technology itself has not changed much. The application development of cloud computing is also similar. So now they are given this new technology, and they think that's all there is to it. I think they have not fully considered the speed of technological updates, how to view the next wave of technological trends and subsequent trends, and how to think about implementing these technologies to adapt to this changing speed. Host: Your company has clearly adapted to this speed of change. European companies often cannot keep up with your rapid changes. Because they are used to existing workflows and procedures, and then you make updates, they will be caught off guard, right? Brad: Yes, it is indeed very difficult. This is the challenge of our work, isn't it? I think companies want to grow quickly, but it is very difficult to do so when your scale reaches 10,000 or 20,000 times. Therefore, this will be the biggest challenge we face in the coming years. Host: Sam mentioned the balance between research and culture, as well as the challenge of balancing the two when building a sales team, because the functions and culture of product and sales are difficult to effectively integrate. What challenges do you think this balance faces? Sam: I think this is where Brad and I have established a good working relationship because we have different views on how to balance any specific decision. I believe that we are very good at listening to each other's opinions, and the situation or feeling can have a more significant impact. But I think we have a very deep consensus on ensuring that research drives products and products drive sales, which Brad and many others may not think so. Of course, this does not mean it is entirely so. There must be feedback in another direction. One of the reasons we now like to have users is that this is the most important reward signal you can get to judge the quality of the model. How useful is it to people in the end? This is the most important thing. But we also know that to sell more products, the best thing we can do is to make the products better. To improve the quality of the products, the most important thing we should do is to conduct more in-depth research. At this point, we have never had any disagreements between us, and this is crucial. Host: Interestingly, you mentioned Alex Schultz from Meta, who was previously very good at growth, but after OpenAI, how has his growth mindset changed? Sam: Alex Schultz is indeed a genius in growth. He can talk endlessly about user retention curves and various metrics, and he knows the tricks of these aspects very well. I think in general, you don't learn much from failure, success is a better teacher. However, extreme successful cases that break all the rules actually don't teach you much either. The success of ChatGPT is a rare technological revolution, not something you can operate on advice. So, if I want to learn the ways of growth, I might not be able to consult Alex Schultz. Host: Why do you say you can't learn anything from failure? Sam: I have always believed that lessons can be learned from failure, such as which practices to avoid. At least from my own experience, I have failed many times and succeeded a few times, and I have learned much more from successful cases. Host: What is the biggest lesson you have learned from success? Sam: There are many, such as what aspects to focus on when recruiting employees. Now I tend to promote internal employees to senior positions, and of course, I am very cautious when recruiting external talent. In addition, you also need to learn how to judge whether a founder is excellent, how to judge the success of an investment, and so on. Krishna Ross has an excellent investment performance, you can consult him on how to judge whether a founder is excellent. I will definitely consider the obvious factors, and some other factors will also be taken into account. For example, I value whether the goals pursued by the founder are grand, and if successful, the impact will be significant. This is more important than people imagine because real big winners are often rare. It is better to have nine failures out of ten investments and achieve huge success on the tenth attempt, rather than pursue mediocrity with seven small wins. Excellent founders should be able to continuously generate new ideas and have the ability to iterate quickly. Of course, qualities like intelligence and resilience are also important Excellent communication skills are also something I value very much. Host: Okay. I messed up a lot, I missed out on a lot of great companies, I messed up, forgive me for being blunt, especially in the seed stage or where I tend to invest, they are not that skilled. So they don't have that kind of communication ability. Sam: As a leader of a company, you need to explain the company's development direction and goals to the team, recruit talents and convince them to work with you, promote products to customers and make them willing to try. At certain times, you also need to give speeches to a wider audience. I'm not saying you have to be eloquent, maybe I'll never be able to do that in my lifetime, but it is very important in daily work to be able to clearly explain what you are doing, why it is important, and how to get others to help you. Host: One last question, let's quickly discuss recruitment. One characteristic you mentioned earlier is that the people you hire seem to be a bit older. Do you prefer to hire experienced talents, or do you value the enthusiasm and fighting spirit of applicants more? Based on the recruitment situation you mentioned earlier, I guess you lean towards the former. What do you think? Brad: At least in the company I work for, in formulating recruitment strategies, I differentiate the composition of team members and their responsibilities. I advocate teamwork and encourage everyone to come up with good ideas. In fact, many excellent ideas often come unexpectedly and not always from the most experienced people. Therefore, my suggestion is to establish a flat team structure, treat everyone's ideas equally, encourage everyone to brainstorm, make decisions together, judge, and unleash creativity. Of course, recruiting some experienced talents is also important because they can bring more perspectives and experiences. However, I believe that truly innovative ideas that can change a company often do not come from these experienced talents. Host: Do you agree? Sam: I think some positions do require experience, while others do not, and sometimes experience can even be a disadvantage. Most of the members leading our team are in their thirties and forties, unlike most leaders in other startups who are mostly in their twenties and thirties. Our technical team's average age is also slightly higher. I don't have specific data, but I estimate that the average age of our technical team should be in their early thirties, while some other tech companies may only have people in their twenties. I think part of the reason is that the path to becoming an excellent researcher is very long. Of course, there are some exceptions in this regard. In conclusion, I think whether it's experienced talents or those with almost no experience, as long as they are outstanding talents, they are worth considering. So far, our recruitment strategy seems to be effective. This is not an either-or choice, but rather a consideration of who is the most suitable candidate. Brad: I would like to add that in certain specific fields, recruiting experienced talents is indeed very important. Because what we are doing is completely different from the past, this is a brand new field. The way people use, consume, discuss, and evaluate this technology is completely different from before. Therefore, traditional business model manuals are not applicable to this new technology, and there are no ready-made solutions to refer to In my opinion, those with 20 years of work experience may not necessarily provide better solutions to issues in these cutting-edge fields. Host: One of the great joys of emerging industries is that they provide a fair competitive environment. I think you especially see this in the field of encryption. Sam: Talents at the age of 19 and 45 can both make significant contributions because this is a brand new field without established experience standards. In general, if you want to evaluate employees at OpenAI, you can look at their positions, responsibilities, and impact, and then judge whether they meet your expectations in terms of having more or less experience. Host: Are you ready for quick answers? Each question in 60 seconds or less, let's get started. Sam, what is the biggest challenge that OpenAI faces in the next 12 months and the next five years? Sam: The most important challenge in the next 12 months is how to conduct the best research and translate the best innovations into products. Five years later, the biggest challenge may be the supply chain and computing power. Host: Brad, if you could change your mind, what was the biggest challenge in the past 12 months? Brad: I think the speed at which companies adopt artificial intelligence is actually much faster than people realize. I think we are breaking conventions. People think that companies are slow in adopting new technologies, but I don't think that's the case. Host: Do we have a lot of experimental budget? Brad: Yes, we do have a lot of experimental budget, which is very helpful for our work. Host: Sam, what is the global issue that concerns you the most today? Sam: I feel that global issues are becoming more serious, including geopolitical issues, socio-economic issues, and political issues. I feel that the world is more unstable now than at any time since I started paying attention. That's the key issue, that's the root cause. I feel the overall macro instability is very high. Host: Brad, what has been the most unexpected development in the process of OpenAI's growth? Brad: What surprised me is the consistent scalability of the models. Despite this trend having continued for six years, I still find it unbelievable that performance predictably improves as the models grow larger. It's really amazing. Host: Brad, what is something you know now that you wish you knew when you started working at OpenAI? Brad: I wish I understood the sequence in which this technology actually has an impact. For example, the importance of this technology in the creative industry exceeded our expectations, while we previously focused more on applications in knowledge-intensive industries or industrial sectors. We did a lot of research on robots early on, thinking we would collaborate with robot companies to manufacture robots and work with game companies to develop intelligent agents, but the actual development turned out to be completely different Host: Sam, if time allows, what else would you like to do? Sam: I hardly read books now, I used to read a lot. This is a change that I regret. Host: Do you want to free up more time to read? Sam: Maybe, but it may not happen in the short term. The current way of sacrificing personal life for career development is acceptable to me. I know it's not a long-term solution, but I still feel a bit sad. Host: Sorry, this question is a bit deep. Are you happy to meet Elon? Sam: I am very happy. Although I can't say I enjoy the process, I am indeed very proud of the progress made. Host: Both of you got married in the past year, which is really exciting. Can you share some tips on how to maintain a romantic relationship and happiness in such a busy situation? Brad: Communication is key. I am still learning how to communicate better, to have empathy, and to understand that this job may be one of the hardest jobs in the world. However, the one who truly pays the price is not you, but your partner. Sam: I am really lucky to have married Christie, but he has indeed sacrificed a lot for my work. Our life used to be very calm, but now he supports me a lot, understands my work. He says, go ahead, I'll be here waiting for you, we will have plenty of time together. We still make time for dates, but having a supportive partner, not just supportive, but a passionate partner, is really important. He says, go do it, I'll figure out how to make things go smoothly, I'll try to be more flexible. I really appreciate this. Host: Brad, what are your expectations for OpenAI in the next 10 years? Brad: I hate making 10-year predictions. Host: It's okay, you can say 5 years or 20 years. Brad: It doesn't make a difference. I don't know, moving on. I know Sam hates it more than I do. Host: For example, when you look ahead 10 years, how do you view the world at that time? Are you excited about the future state? Brad: Yes, otherwise we wouldn't be in this job, at least I wouldn't. Sam: Very excited. I hope that by then people will look back and say, "Wow, people in 2024 lived so primitively and barbarically!" just like we look back at people's lives hundreds of years ago. It's not that we don't appreciate the good life now, but people get sick, they die prematurely from diseases. Not everyone can receive a good education. Everyone should be able to do things and manage their time according to their own wishes, not to mention there will be new things in the future that we can't imagine now. Of course, the future is not all good, there will be things we lose. But overall, I am extremely excited about a truly prosperous world Host: Thank you very much for taking the time to do this interview. Honestly, it's great to be able to have a face-to-face conversation with both of you! Once again, thank you for joining me in today's interview. Sam and Brad: Thank you for the invitation, this is awesome!