【Introduction】

As DeepSeek is stirring up the global AI landscape with its low-cost model, OpenAI CEO Sam Altman and SoftBank founder Masayoshi Son had a major conversation in Tokyo. In the conversation yesterday, on February 3, Altman publicly responded to the challenges brought by DeepSeek for the first time and revealed OpenAI's choice of technology route. Masayoshi Son, who has always been high-profile, bluntly stated that the investment in AI computing power will show an exponential growth of "10×10×10" in the next few years, pointing directly to the current "low-cost route" represented by DeepSeek.

[Preview of key points]

• 1. OpenAI responds positively to the DeepSeek challenge, reaffirming its belief that "the more you invest, the stronger your intelligence will be"

• 2. SoftBank invested heavily in OpenAI and firmly bet on high investment strategy

• 3. Sam Altman first disclosed the five-stage roadmap for AI development, calling 2024 the "first year of AI agents"

• 4. Demonstrate the fundamental differences between OpenAI and DeepSeek in terms of technical routes

• 5. Masayoshi Son plans a "billion-level AI agent" plan, showing a technological vision different from the low-cost route

"There has been some publicity about China's DeepSeek recently. They may imitate or catch up, and the cost may be much cheaper." When Masayoshi Son mentioned this sensitive topic in the conversation, Sam Altman gave a meaningful response: "When we want to make the smartest model, we need the greatest computing power. Some people say, 'You can also use a small model for compression,' but that's really just a small fight. I don't think people really understand to what extent the returns are growing exponentially. Of course, the cost is also increasing exponentially, but I think the benefits are growing faster." This clear-cut response reveals the fundamental differences between OpenAI and DeepSeek in their technical routes.

What’s more interesting is that when talking about the development speed of AI, Masayoshi Son directly gave an amazing calculation: "Now your model is improving about 10 times every year. In addition, the performance of the chip itself is also advancing at a speed of 10 times. Then our 'Stargate' project is massively increasing the number of chips, which can increase 10 times a year. So 10×10×10... one or two years is 1000 times, and then 10×10×10 again the next year, which is another 1000 times. 1000 times 1000 is 1 million times, and a few more times is a billion times improvement." Faced with such a large investment, Sam Altman said: "I don't think people really understand to what extent the returns are exponential..."

In this one-hour conversation, the two technology leaders not only discussed the development path of AI, but also disclosed in detail for the first time the future technology layout. It is particularly noteworthy that they repeatedly emphasized a core point: the real AI revolution has just begun ...

Manuscript finishing

Masayoshi Son: Thank you, Sam.

Sam Altman: I'm glad you enjoyed it. Yeah, that's great, I'm really excited that we can announce this today.

1. OpenAI Technology Development Roadmap

Masayoshi Son: Yes, we were thinking, 'Can we really work together? ' And it actually happened. We have been talking about doing this for a long time, and now it is finally finalized and announced to the outside world, which I think is really great. We will set up a joint venture in Japan, "SB OpenAI Japan", to provide artificial intelligence services to corporate customers. This joint venture will be jointly owned by OpenAI and the company we set up with SoftBank and its domestic telecommunications division. SoftBank Group and its investment companies plan to invest $3 billion a year in OpenAI tools.

Sam Altma:  Yeah, and the world is going to need a lot of compute going forward. As I said before, we can get small models to do some incredible things, but if we're going to really push the frontier of intelligence, we're going to need a lot of compute. And that's where the most value is going to be. So we need a lot of compute to train these models, and people are obviously going to need a lot of compute to run these models. It's great to finally be able to do this at scale. I feel really good about it, really good.

Masayoshi Son: Yeah, yeah… About a year and a half ago, we were having dinner together and we were talking about, ‘Sam, when will AI really arrive? How big will a computer have to be?’ You and your team gave me a simple answer: ‘The more the better,’ right? ‘The more the better.’ I thought, if the more the better, then we should invest a lot. Now we have done a lot, and that’s where we started.

Sam Altma: Yeah, exactly, 'more is better.' It's not like there's a fixed upper limit to how much computing power you can have, but more is better. More computing power certainly leads to more powerful models, right? Some people say 'you can also use smaller models to get compression,' but that's really just playing around. I don't think people really understand the extent to which the returns are exponential. Of course, the costs are exponential, but I think the benefits are growing faster. When we want to make the smartest models, we need the biggest computers.

Masayoshi Son: Yes, it reminds me of the early days of the Internet. When we first started the Internet in 1995, we used a regular PC that could only display large text, and it was slow and expensive. Then when broadband came along, people said, 'Why do we need so much bandwidth?' Then there was more bandwidth, and people thought, 'That's enough, the demand won't grow anymore.' But then pictures came along, and the resolution got higher and higher, and then videos came along, and the demand for bandwidth just kept going up and up. At first people were still saying, 'The Internet is just a virtual thing, it doesn't have any real use, and most services are free, so there's no business model.' All these criticisms seem ridiculous now, right?

Sam Altma: I think the same thing is going to happen with intelligence. Some people are asking, 'How smart does AI have to be?' The answer is, 'The smarter it is, the better.' People are going to use it a lot, they're going to generate a lot of video, they're going to solve very difficult problems, and then everything in the world is going to be very smart.

Masayoshi Son: Now, your model is actually improving very quickly, right? Like, it can improve 10 times a year? What is your metric?

Sam Altma: We don't have an exact scientific quantification method, but as a kind of 'intuitive' or 'spiritual' feeling, we can probably improve by one standard deviation of IQ every year. In addition, the computational cost of the previous generation of models has dropped by about 10 times every year. The decline in chip costs leads to higher cost-effectiveness - we can buy 10 times more chips with the same budget, and the algorithms themselves have become more efficient, so these factors will be added together. This rate of progress is easy for people to take for granted. In 2018 and 2019, we had GPT-1 and GPT-2, and people didn't think it was that shocking at the time. When GPT-3 was launched, it first attracted some people's attention, but in fact GPT-3 is also limited. If you go back and use GPT-3 now, it will feel like entering an old computer museum and looking at a Xerox Alto, which is decades old. You can see that it does some functions and has some shadows of modern computers, but it is 50 years old after all, and it seems ancient now. GPT-3 was actually only a few years ago, but looking at it now, it feels a bit like a joke. ChatGPT was launched just over a year ago (end of November 2022), and GPT-4 was released in March 2023. If you look at the speed of progress: how fast the model's capabilities are improving and how fast the cost is falling, if we can keep on this curve, the future will be amazing.

Masayoshi Son: In my opinion, your model is improving about 10 times every year. Plus the performance of the chip itself (thanks to the efforts of Jen-Hsun Huang and the entire industry) is also improving at a rate of 10 times. Then our 'Stargate' project is massively increasing the number of chips, which can increase 10 times a year. So 10×10×10... one or two years is 1,000 times, and then the next year is 10×10×10 again, which is another 1,000 times. 1,000 times 1,000 is 1 million times, and a few more times is a billion times improvement, right? Some people will say that Deep Seek has also had some publicity recently, and they may imitate or catch up, and the cost may be much cheaper. But a year later, you will continue to rush on this path and release more updated things, such as GPT-3.0x, GPT-4.0x and so on. People don't realize how amazing this exponential speed is. When you are on this exponential curve, it is actually difficult to feel it directly because you are constantly adapting, but we are indeed moving on a very steep curve. This is amazing, this is amazing. So, in just a few more iterations, we can get to a billion times faster. Think about the next 10 years, this will become true superintelligence, right? I think people can't even imagine what that will be like now. Because most people are used to linear thinking, when true exponential growth comes, it will exceed human imagination.

Sam Altma: Yeah, I agree. It's hard to fully appreciate it when you're in the middle of it, but I've found over and over again in my career that you have to believe in the power of exponentials. Our brains aren't designed to understand exponentials, but we have to believe that it's going to happen.

Masayoshi Son: So you are still excited about the potential for future innovation and don’t feel that it has reached a peak or saturation?

Sam Altma:  No, no, no, far from it. In a few years, looking back at the current GPT-3 or some version, you will definitely think: "Oh my God, how could it be so weak at that time?" It's unimaginable.

Masayoshi Son: Some people think that it is difficult to integrate AI agents or various prompting projects, but in fact, this technological progress itself makes everything simpler. Users do not need to do a lot of complicated implementations by themselves. It will become more and more friendly. Just like how we talk now, communicating through voice and eye contact. We will also do the same with AI in the future, through voice and through each other's "eyes".

Sam Altma: Absolutely. Now, just a text box has brought great value to people, but our world is not just a text box. So we will definitely add more ways, such as voice, vision, etc. It's like talking to this 'crystal ball'. As long as you open your mouth, AI can see you, see your face, and understand your tone. Just like we communicate with each other, it can also capture information by itself through voice, expression, and surrounding environment, and then talk to us.

2. Sources of funding for the Stargate program

Masayoshi Son: Yes, this is going to happen very soon, I think so too. Some people say, 'The capital expenditure (capex) of the Stargate project is too large, how are you going to raise the money? Masa, do you have enough funds?' What do you think? I think we still need a very large capacity and a lot of potential upside to really bring the technology to life, right?

Sam Altma: That’s exactly the point I was trying to make. I think that as long as you make linear improvements in intelligence, the value you get is exponential. If we can push the intelligence of these models a little bit further, we’ll create a lot of value economically. Of course, it will require a lot of capital expenditure, but the revenue growth curve will also climb very quickly.”

Masayoshi Son: We have a mutual friend, Elon Musk, and he would say, 'Masayoshi Son, do you have enough money? ' I would tell him, 'We will definitely get it done. We are not a bank, but we are SoftBank. ' I fully believe that we will push this forward. Now we also need to expand Stargate to Japan, because there are regulatory issues, we must respect national security, privacy laws, and so on. Yes, you SoftBank are building a large data center in Japan. We will expand Stargate to Japan's infrastructure. We are the center of innovation here, and the United States is mainly responsible for training and core brains, but each country has its own culture and national security needs. I think we should not only expand in Japan, but also expand this model to other countries and regions, respecting local culture and national security.

Sam Altma: Yes, we certainly hope so. Although we are indeed an American institution originally, our mission has always been to build AGI for all of humanity. We want to make sure that our system reflects the values, culture, and language of different people around the world.

Masayoshi Son: I once took a photo of a place in Japan and asked the AI, 'Do you know where this is?' And it said, 'Oh, this is supposed to be XYZ.' I asked, 'Did you use GPS?' It said, 'No, I didn't use GPS, I looked at the moss on the rocks and the way the rocks were joined. This is supposed to be XYZ, a 500-year-old monument.' It got it right! I was stunned, thinking, 'Oh my god, how can it know so much about Japan? It's so smart!' It really stunned me. It's a kind of inference, a kind of predictive inference. It's not based on all the detailed data, but by guessing, guessing, guessing, guessing, and guessing, it finally guessed that historical monument. It was amazing! I was really shocked. It even understood my jokes - I remember asking it to make a joke in Osaka dialect (you know, different regions in Japan have different dialects), and then it explained to me why the joke was funny. I thought, 'Oh my god, it even understands context and culture!' 'That's already pretty impressive, but I still use it every day and am constantly blown away by how amazing it is.

Sam Altma:  Yeah, that’s awesome.

3. Commercialization of AI for Enterprises

Masayoshi Son: We also announced the 'Crystal' product today. When we want to read various source codes, such as the data of our internal group's 2,500 systems, which contains billions of lines of code, such a scale must require a lot of computing power, right? Do you think that in Japan, if we have enough computing power, such as with the support of Stargate, we can read all our source codes for the past 30 years and let your model process them? Do you have confidence in this?

Sam Altma: Yeah, we're pretty confident we can do it. Haha, thank you. Yes, it's totally doable. People might take it for granted, but it's actually pretty amazing to be able to do it directly. I'm very confident about this.

Masayoshi Son: I'm really excited to be able to integrate all the source code and be able to interact with it in real time in meetings, and also have long-term memory. But now we don't have the real 'long-term memory' function. When do you think we will achieve long-term memory?

Sam Altma: I think it will definitely happen in the next few years, and maybe even sooner than we expect. It's so important to have these models have almost unlimited long-term memory. It's a very important step to have AI really understand your entire life, understand all the information of an enterprise or an entire company. We have been working on this.

Masayoshi Son: I have my own 'patent' or 'concept' for long-term memory, which is that when we have a conversation like this, I can see facial expressions, emotions, and the tone of voice. All conversations can be converted into text, but the tone of voice and facial expressions must be understood. I have a set of emotion mapping here, including about 250 emotions. Then each emotion is indexed, such as fear, anger, or suspicion, and so on, with about 250 words used to express emotions. For each emotion, we also use a scale of 1 to 10 to indicate the intensity of anger. If an emotion is very strong, it may be 10, or it may be weak like 3. I will index each emotion and its intensity, and then convert it into a digital index mark, so that the conversation can be compressed into a text with three digital indexes. By expressing it in this way, the conversation is compressed. Then when there is a strong emotional fluctuation (such as extreme anger or very uneasy), multimodal recognition (including video) will capture all the information and store it in long-term memory. But if it's just daily greetings like "good morning" and "good night", or when you drive to work every day and see traffic lights or other vehicles, the human brain tends to forget those things, otherwise the brain capacity will explode. So you will compress those unimportant information. But if you encounter something that is very surprising or emotionally intense, you will not over-compress it, and you will record and store multimodal information such as video, sound, and audio. For example, on your 3rd birthday, you should remember that moment, which was very happy for the family, so it will automatically capture and store all multimodal data.

So, this is my concept of long-term memory. The key is to index the ‘surprise level’ or ‘emotion intensity’. Because humans communicate not only through words, but also through emotions, for example, ‘I like you’ can have completely different tones and meanings. So by indexing the tone of voice and facial expressions, the information can be compressed and stored in long-term memory. This context is very useful in the next conversation, discussion or negotiation. If you can’t read the other party’s emotions during a negotiation, it’s easy to fail, right? This is the patent or concept I filed 10 years ago. I think this should be useful soon, right? "

Sam Altma: Very soon. Really very soon. I'm not sure about the details yet, but I think that having AI with emotional expressions will definitely have a bigger impact than we expect - not just text from a chatbot, but you see a rendered virtual face, and it has emotions. This aspect will be very exciting, but it also means that we have to build some new social security measures. But I think this is also very promising.

Masayoshi Son: Right, right, right. Our friend Johnny is doing this, right? I am very much looking forward to seeing this development. If we have this data and long-term memory, we need super large capacity (computing power and storage). Latency will also become very important. For example, customer service call centers must have real-time responses. Do you have confidence that you can do this here in Japan? I used to worry about latency, but now, if you experience it with our current voice mode, it feels like talking to a real person, very fast and good experience. Do you think this can be applied on a large scale in Japan? Enterprise-level, mission-critical? I used to worry about this, but now I feel that it may work.

Sam Altma: Yeah, I was very concerned about latency. But even if you use our voice function today, it feels like you are talking to a real person, the response is very fast and the effect is very good. I think we can solve this problem. If you think back to a few months ago, there was a lot of latency, but now, even last night when I used it, I was like, 'Wow, it responds very quickly! ' So for me, the latency has reached a pretty good level. It is now around 100 milliseconds, maybe a little higher, but it is very fast. 100 to 200 milliseconds, which is actually about the interval between human conversations - human conversations are about 200 milliseconds, right? So the interaction between 100 to 200 milliseconds is almost synchronized with human conversations, and you can interrupt in the middle, which is the key. Because humans also interrupt in the middle. All of this has become a reality, so do you ask me if I am confident that after training the model, the service can be provided in real time in Japan? The answer is yes.

Masayoshi Son:  Even if the model is trained in the United States, but there is a Stargate center in Japan, do you think real-time response will be no problem?

Sam Altma: Yes, we definitely need to put some very low-latency use cases closer to the user. But as you said, we can train in the United States and handle a lot of the "thinking" (deep reasoning) in the United States. It's just that in some application scenarios, for extremely low latency, we need to deploy the model on the edge node.

Masayoshi Son: Yes, yes. Those parts that do not involve national security can continue to be done in the United States, while those that involve national security and privacy can be handled locally in Japan.

Sam Altma: We can certainly deploy models globally. Yes, yes, yes. Then, we will assign a thousand sales engineers to this new joint venture project. They need to implement and deploy each system and build agents for each task.

Masayoshi Son: Can you explain to me how these agents work? Is it used to complete a single task, or is it a very complex multi-functional agent? I also want to understand more deeply - is your agent a relatively single-task agent, or is it quite complex and can handle many tasks? "

Sam Altma: We will have some general-purpose agents for consumers. They can do a lot of powerful things, like the deep research we just saw, browsing the web, and things like that. But for a business or an organization, or for everyone, the ideal agent would be one that has the same context, permissions, and capabilities as an employee inside the company. That means you need to connect all of the company's systems, all of the knowledge bases, and give the agent access to the company's internal code and understand how the company works. This will require a lot of customization for each company. But once you do that, you can imagine what changes that will bring. For example: If someone integrates an AI agent into SoftBank's business, and there is a fictional competitor that doesn't do that, SoftBank can do more, faster, and better. Once AI is truly integrated into the company's workflow and has all the powerful capabilities that it brings, it's not just about doing deep research or browsing the web and writing general code, but it's deeply integrated with the company's systems. That will be extremely powerful.

IV. AI Safety and Regulation

Masayoshi Son: Yeah, the difference between a company with Crystal and a company without Crystal is like having a machine gun and having... well, never mind, maybe this analogy is not quite right. But the difference will be very obvious, just like a country with electricity and a country without electricity. Or a country with cars and a country with only bicycles, the difference in productivity will be huge. Do you think this will happen again in the field of AI?

Sam Altman: Absolutely. I think this is one of those pivotal moments. You mentioned the sword - I collect ancient technology myself. I have a sword from the beginning of the Bronze Age. They were not only forging blades, but they were also forging metal hilts, and then they attached the hilts to the blades. So you could use it for slashing, instead of the previous versions where you had a wooden hilt stuck to the blade and it would break, so you could only use it for stabbing. This was a moment where technology gave people a decisive advantage. It changed all of Europe in just a few decades. I think AI is one of those technologies. Companies that don't fully integrate AI are going to have a hard time competing with companies that have.

Masayoshi Son: Yes, not just between countries or industries, but also between companies. There was a recent example of 'Deep Seek'. Now you have put a lot of effort into protecting human safety, such as preventing AI from outputting dangerous answers or making very dangerous decisions. So, the technology and output may look 99% similar, but there is an extra layer of 'safety guard' to protect human or national security, such as eliminating wrong answers and dangerous decisions. This last 1% or 2% of fine-tuning requires a lot of effort, right? "

Sam Altma: Yes. Society needs to figure out some boundaries. We pay a lot of attention to this and put a lot of effort into security. But as long as we do it right, people will be willing to use it, which is a good thing, right?

Masayoshi Son: I don’t want to get too political, but depending on the situation in different countries, if it is used improperly, it could have very dangerous consequences, right? It could be the trigger for a terrible future or a terrible war. So, some people ask, are these AI agents or Crystals designed to save money? Will it put people out of work? I think you must be asked this question a lot, how would you answer it? "

Sam Altma: Yes, it will save costs, but that's not the most exciting part. The most exciting thing is that we can do more and achieve more. Freeing humans from repetitive tasks to do more ambitious and higher-level tasks. Every time there is a technological revolution, people worry: 'What about the jobs?' But then we always find new things to do, which is great. People will perform at a higher level and have higher expectations. Of course, AI will make a lot of processes more efficient, which is good for the economy. But what I am most excited about is that these systems can help us create new knowledge - knowledge or scientific breakthroughs that we might not be able to do on our own. If the rate of scientific progress can be accelerated significantly, so that we can achieve in one year what we have achieved in the past decade, and the next year what we have achieved in the past century, it will have a huge impact on people's quality of life and economic development. This is not just about making something 'cheaper', but things that we simply couldn't do before. We were not smart enough to do it without the tools."

5. The future of AI Anget

Masayoshi Son: Yes, you announced five stages of AI capability improvement. I remember the third stage is ' agent ', which we just started using this year, right?

Sam Altma: Yeah, it just started, this year - actually just recently, or last week. This year is the original year of 'agent'.

Masayoshi Son: And then you also mentioned that the next stage is 'innovator'. Can you explain how it works?

Sam Altma: Our AI systems are currently very strong at integrating existing knowledge and doing things similar to what they have done in the past, but they are not yet actively generating new scientific discoveries. And Innovator is our next step. I think it will have a transformative impact on society. We have a lot of work to do on Agent this year, but then we will focus on Innovator.

Masayoshi Son: Some skeptics would say, "AI has its limits, because it must be taught by humans, so how can it be smarter than humans?" But "innovator" means that it can innovate or invent something that has never existed before. Can you tell me more about how "innovators" can invent something from scratch? Does it have an exploring mechanism?

Sam Altma: I think it's very similar to the human process. If you're trying to solve a problem that you haven't solved before, you'll come up with a bunch of ideas and then notice how they relate to each other, or what works and what doesn't work based on existing knowledge. For example, you'll find: 'This doesn't work, this doesn't work, this is interesting, I'll go a little further...oh, that doesn't work either...oh, this direction is good, I'll keep pushing it.' Once you think, 'Oh, I can go here, here, here,' and it looks promising, you'll go further in that direction. For humans, the creative process may not always look like this, but behind the self-perception, I guess it works pretty much like this: making a lot of small changes to what already exists, and then gradually building up more promising ideas. I think we can use AI to achieve this process.

Masayoshi Son: So reasoning is the first step, right? It may be three steps, ten steps, or even a hundred steps of reasoning. Then when humans innovate, they will try to think from different angles and explore. This is the concept of 'exploring' you mentioned. I have submitted 1,800 patents in the past 12 months (I understand this as an exaggerated way of expression). In my mind, I keep diverging and forcing myself to think from the right brain, trying to try in all aspects. I think AI's 'agent' reasoning can also force itself to make various different attempts through a mechanism - 'exploration', which is the key to innovators. It can conduct thousands or even hundreds of millions of trials and errors, and occasionally encounter the right answer, which is invention.

Sam Altma: Exactly, that's how the 'Innovator' will probably work. Yes, it's very quick to try, which is great. But I can't say too much, after all, there are some development secrets.

Masayoshi Son: Haha, okay. Then the fifth stage you mentioned is 'organizer', which is the coordination between agents, right? "

Sam Altma:  Yes, I mentioned this idea a little bit to Rene before: many agents or many "innovators" working together. Imagine that in a data center, you can run countless "brains" at the same time, communicating with each other, sharing ideas with each other, and bringing different expertise. You can even imagine it running like a virtual company. In this way, the power is quite amazing. "

Masayoshi Son: For our Crystal project, my vision is to create 'a billion agents' within the SoftBank Group alone. Because we have 100 million users for LINE, 40 million for mobile, and 70 million for PayPay... If each account and each function can have 10 sub-functions, then each function can be assigned an agent to complete simple tasks. Instead of having one agent do something very complex, it is better to have a lot of micro-agents to do small and clear tasks. So the picture in my mind is: within the SoftBank Group, there are a billion agents to handle various matters. Although a billion agents sounds like a lot, it should be able to be done from the perspective of capacity or computing power, because each agent is only responsible for a very simple job. We are actually very good at this model (micro-agent parallelism). I think although we still have a lot to learn, the general direction is like this, right? "This is my vision: to use Crystal to build a billion agents within SoftBank, and then perfect the internal experience first, and then we will be the 'evangelist' to tell our customers: 'Look, this is how we improve efficiency, and you can do the same.' This is my imagination of Crystal. What do you think of this direction?"

Sam Altma: Yeah, let’s do it!

6. Impact of AI on Society

Masayoshi Son: I want to ask about cybersecurity. There will always be those bad guys who try to attack others and do bad things, whether intentionally or unintentionally. We have to protect more people - after all, many people's lives will depend on this super intelligence. What do you think about this? Because when AI starts to become more and more proficient in programming, it will obviously be used for cyber attacks. So we need to stay ahead in cyber defense. I am optimistic about the potential of AI, but defense is indeed more difficult than offense. You mentioned that this is a good idea, so the whole world must start taking this issue seriously as soon as possible, right?

Sam Altma: Yeah, it's really important. There are always bad guys out there, and it's a big risk. I'm optimistic that we can figure it out, but we have to do it quickly.

Masayoshi Son: I am optimistic too. You know, 99% of the world's people are good, but there is always 1% of bad people. We have to keep working to protect the 99% from the 1% of bad people. But with innovative superintelligence, the good side will also keep working to find solutions. Just like when cars appeared, there would be traffic accidents, and we would introduce regulations, cultivate etiquette and morality, and everyone would slowly learn to adapt. This is also what you said that we need healthy regulation, not excessive restrictions. There should be enough space for innovation, but at the same time there must be reasonable regulation, right? What do you think?

Sam Altma: I totally agree. Yes, I strongly support these points. Some people say to me, 'I'm surprised you're calling for regulation of this industry,' but regulation of important industries will come sooner or later. The key is how to strike the right balance. If there is no regulation at all or too much regulation, both are bad. So we have to discuss what is appropriate and reasonable.

Masayoshi Son: Yes, it should be moderate and healthy regulation. We cannot let innovation stagnate because of regulation. OK. Speaking of innovation, what do you think about using AGI to solve medical problems? This is an area that I am very concerned about.

Sam Altma: This is one of the applications I'm most excited about. Think about how we can provide high-quality healthcare to everyone on the planet, and think about how we might be able to cure or improve many diseases, and maybe even cure all diseases one day. I think this is within reach. Everyone has probably experienced a situation where if such technology was available at the time, it would have greatly helped themselves or their family. I believe that AI will achieve very significant results in this area.

Masayoshi Son: Now let’s talk about robots. You like robots, and I like robots too. This should be one of your favorites, right?

Sam Altma: Like everyone else, I've wanted robots for a long time. But it's always been difficult. Now the AI ​​part (the 'brain') is finally improving, and we've been able to build the 'body' for a long time, but the key has always been the 'brain'. I think it will be possible soon. So in a few years, we may see very good humanoid robots, as well as various other types of robots - which will also change the world.

7. AI, Emotion, and Philosophical Thinking

Masayoshi Son: In this way, humans will no longer have to do dangerous work, physical work, sweaty work, and boring work. Then some people will ask: "What kind of work will humans do then?" What do you think about this question?

Sam Altma: Humans always find new jobs. It always happens. Think about us here today. If people from 500 or 1,000 years ago looked at our jobs, they would say, 'What kind of job is that?' We think we are so busy and important, but in their eyes, it was not just a way to survive, but a 'game'. I hope that people in the future will also look at our generation in this way: AI does a lot of our work today, and we can do more meaningful and interesting things. When they look back at us, they will say, 'Why did you work so hard for such a small thing?'

Masayoshi Son: Yes, I think so too. Also, regarding education. When you first launched ChatGPT, many schools banned students from using it. What did you think at the time?

Sam Altma: I can understand why they would do this. When people saw this thing, they thought, 'Oh my god, the world has changed. Students can use ChatGPT to write papers. How can education be done?' But very quickly, many teachers and education administrators who originally banned ChatGPT said, 'Wait, this is a big mistake. We should do the opposite - embrace it. This is the future, and students must learn how to use it. We have to adjust the entire curriculum.' Now it seems to have become part of education.

Masayoshi Son: Yes, and it (AI) is bringing amazing results. I believe this trend will continue. I use ChatGPT every day now, such as GPT-1 and GPT-3. The more you use your brain, the more you will think actively and think in a conversational way, just like we talk like this - just brainstorm with ChatGPT or GPT-1 and GPT-3. In fact, your brain starts to operate in a conversational way. Children will learn more by using it. Some people say, 'Children will not learn if they have this,' but I think it's the opposite, right?

Sam Altma: Of course, there will be some kids who will want to use ChatGPT to do as little homework as possible, but overall, I think people will learn more, do more, and be more capable. Just like debate, you learn more through discussion and debate. Now that it's part of the world, people are doing everything this way. And it's incredible to watch young people use ChatGPT - they're solving problems in a completely different way than we're used to.

Masayoshi Son: We have talked about emotions before. Do you think our AGI and ASI will begin to have their own emotions? What do you think?

Sam Altma: I personally don’t think there will be true emotions, but maybe there will be something like emotions. I’m not sure, maybe you think there will be?

Masayoshi Son: I think it will. You see, even dogs have emotions. I don't know if fish have them, but maybe fish run away when they encounter dangerous enemies, which is also an 'emotional' reaction. So I think emotions are very important - whether it is to produce more output, be more efficient, or for self-protection. For example, if a dog has no emotions, would you still think it is cute? If a dog has no emotions, it may bite people all the time. I think if AI makes us feel that it has emotions, it proves that it really has some kind of 'emotion'. Of course, some people will say that it actually has no emotions, but just looks like it. This may become a philosophical debate.

Sam Altma: That's true. There may be a point where we think it does look like it has emotions. Whether it really does or not is a big debate.

Masayoshi Son: Well, my point is this: initially people said that ChatGPT didn't actually understand context, but then people found out that it did understand context, because at first it had all kinds of 'hallucinations', and people thought it didn't understand, but now with the improvement of its reasoning ability, people suddenly say 'oh, it really understands context'. So I bet that in a few years, maybe 10 years or so, it will at least understand human emotions, and then it will gradually develop emotions itself. And this is a good thing, because it can better protect humans. People are worried that AI with emotions is like a demon and will destroy humanity, but I think if the energy source of AI is not 'protein', it will not want to 'eat us'.

It doesn't need to 'eat us' to get energy, so there is no reason for AI to feed on us. It may find that 'making humans happy' is better for it. Therefore, we will not be 'eaten' by AI, which is a sure thing, right? I bet you that this will be a good thing. It will understand human happiness and try to make humans happier. You also see that now you let AI try not to give dangerous answers, and as it gets smarter, it will understand love better and learn to be more friendly to people. Just like we are friendly to our friends, AI will be more friendly to humans. I believe this is a good thing.

Sam Altma:  Okay, I understand.

8. Review of OpenAI’s original intention

Masayoshi Son: Well, we are running out of time. Last question: Why did you start OpenAI? What was your initial motivation? Tell us your story.

Sam Altma: When I was studying AI in college, I found that the field was not progressing well, so I dropped out of school to start a technology company. But I always wanted to come back to AI one day. I was obsessed with AI since I was a child and I was a big fan of science fiction. In 2012, Alex (Alex Krizhevsky) and his team successfully made some breakthroughs in deep learning, and I thought: 'Maybe the professors in college were wrong about neural networks not working? Maybe it really works? ' I watched for a few years and found that the larger the scale, the better the effect. By 2014, I realized: 'Okay, this thing seems to really work. ' I thought about it for a while, and then we founded OpenAI at the end of 2015. The reason we founded it at that time was that we believed that AGI (artificial general intelligence) might be possible, and once it was achieved, it would be a very important thing. At that time, many people thought we were completely crazy. It was only 10 years ago, but at that time, this idea was not only non-mainstream, but also "non-mainstream among non-mainstreams", because few people believed that AGI was really possible. But we decided to push it forward. For me, this has been the most exciting, fun, cool adventure. I can't imagine anything more exciting than this.

Masayoshi Son: Yes, when I first met you, you were still the president of YC (Y Combinator), and then you started talking to me about AI, about 'human-like AGI', saying that this was your goal. I believed you immediately. I remember in Tokyo in 2017, you said you wanted to do AGI. At that time, I immediately said, 'I believe you, I want to invest.' I still remember you ran to my office in Tokyo.

Sam Altman: Yeah, you said, 'I believe in you,' that was in 2017, and you said you wanted to work on AGI. A lot of people thought we were crazy, but you believed.

Masayoshi Son: Yes, most people thought you were crazy at that time, but I was a believer from day one. Some people also thought I was crazy, but in the end everything came true, right? Haha.

Sam Altma: Yeah, we are here.

Masayoshi Son: I really should have forced you to accept my investment... But we have it now, it is never too late.

Sam Altma: Yeah, it’s never too late.

Masayoshi Son: Well, we talked a lot today. I think everyone should know better that you are the major shareholder and OpenAI is a non-profit organization. Your original intention was to make people's lives better and happier. Has this not changed?

Sam Altma: Yes, I still believe that very much.

Masayoshi Son: Great, thank you! That’s great.