We'd like to start the event, Transforming Corporate Business with AI, presented by Arm, SoftBank Group, OpenAI, and SoftBank. First, we'd like to have Junichi Miyakawa, President and CEO of SoftBank Corp, to give you an opening remark.
Thank you very much for joining us today, and I appreciate your great support for our business. I am Miyakawa from SoftBank. Today, we are honored to have top executives from Japan's leading corporations, as well as members of the media gathered here. We sincerely appreciate your participation. Remarkably, today we have brought together executives from companies that collectively account for more than half of Japan's total market capitalization. The key questions we will explore today are: How will our companies evolve through AI? When and how should we engage with this transformation?
With the rapid global adoption of generative AI, driven by innovations like ChatGPT, DALL·E, Sora, and models such as o1 and o3, we are pleased to collaborate with OpenAI to bring you this event. Joining us for presentations and panel discussions are Masayoshi Son, CEO of SoftBank Group, and Rene Haas, CEO of Arm. AI is now evolving toward what is known as AGI, Artificial General Intelligence, a level that surpasses human intelligence. It is progressing beyond merely assisting in tasks and is on the verge of transforming into AI agents capable of autonomously executing tasks. As business leaders, we must anticipate and actively adapt to this new era of AI-driven transformation with flexibility and strategic engagement. In today's event, we have also prepared a live demonstration of AI agents presented by OpenAI.
We hope this session provides valuable insights for your corporate management and offers new perspectives on how generative AI and AI agents can drive innovation in your businesses. We sincerely hope that today's discussions serve as a source of inspiration for your future management strategies and open new possibilities for your companies. With that, I'd like to pass the microphone to Mr. Son and conclude my opening remarks. Thank you very much.
[Foreign language]
Masayoshi Son, Chairman and CEO, SoftBank Group Corp.
Good afternoon, everyone. This is Masayoshi Son speaking. Thank you very much for joining today. So I was just having an official agreement and signing with Mr. Sam Altman of OpenAI regarding the joint venture. So I would like to also include the report regarding this joint venture as well later on. So first, before we go into that, I would like to show you some interesting things. This is it. So I would like to talk about this today. This is Cristal. When you hear crystal ball, what do you think about it? What do you imagine about it? So beyond the human's capacity, if you ask anything, you're able to know about the future. What do you need to solve? How do you need to solve? What does that mean? And what does that imply? So anything that you would like to know, you would like to inquire.
Once that you hear that, you would be able to learn that. That's the kind of magical things which have been treated as if a science fiction story, and now the time has come that we are seeing this as a realistic world, and that will be actually bringing us to AGI and ASI world going on. So about this thing that I would like to talk about a lot today, Cristal. So crystal ball, that I want you to keep in mind during my presentation. So please pick this up for me, or maybe I can keep it and speak. The other day in the United States, together with President Trump and also Sam from OpenAI, Larry Ellison from Oracle, together with them, we made an announcement regarding the Stargate project.
This was the first day after the assumption of President of the United States. Mr. Trump was super busy, but still, as there are various political events wrapped up, but still that he shared the time for us. And also he himself joined the announcement on this project. Actually, this is not only the event for the one private company, but actually it's going to be an even more important project for the government and for the country itself going forward.
I think that you'll be able to expect on this, and I am sure and convinced that towards 100 years, 200 years, 300 years, it is going to impact the future of human beings. And that project, I am so happy to be able to work together with OpenAI, Oracle, and also as a financial partner, MGX, which is something very much excited me. So Stargate project, this is something that we're going to be working on.
Speaking of this Stargate project, very close future, AGI may be achieved. That's how I think and I believe. A year ago, just about a year ago, I said that AGI will be coming in 10 years. That's something I said in our company events and so on. Just about a few months ago, I changed my words and restated that AGI will be coming in a couple of years. And now I would say AGI will be coming even sooner than that. That's how I feel right now. So AGI, something that relates to today's announcement, this is actually coming to enterprise world rather than to individual consumers. Especially the large enterprises will be the kind of the first beneficiary for this AGI because in a consumer world, there are so many things, phenomena, situations, varieties of things.
Sometimes that cannot explain in logics, but there are emotions or the exceptions, so many things, and to satisfy all that around the world and have a kind of super intelligence for that is going to be a bit difficult. However, when it comes to the enterprise, one enterprise or one group, if you limit it to or target to this company, then this company already or this business already has a large volume of data already, and to achieve AGI, you have to have a high-quality, vast, and limited segments, limited areas, information and data is available. That's going to be the very important base for AGI to provide the benefit, so you have to have a deep, wide, real-time, and at the same time specific to this sector, specific to this industry, that kind of data and information is available. That's very important.
And based on that data, you'll be able to do a lot of trainings, inferencing. It's always going to be available too. So that's why I would say AGI can be achieved in large enterprise businesses first. And this also costs money as well. Quite a huge amount of money is necessary. You need a lot of effort as well. So you have to be capable of affording such cost, which is only available in large enterprise at this moment. That is why I believe AGI's first target should be enterprise, especially large enterprises. That's how I feel right now. Mr. Miyakawa just mentioned that we have companies exceeding account for more than half of the Japanese total market cap. So corporations, CEOs, Chairmen of the Japanese enterprises with more than 500 companies and businesses right now that are together with us.
Such large enterprises, I believe, first will be addressing AGI or such a technology. The latest edge technology to start from Japan here for businesses. We will be starting that from Japan. That is something that we officially agreed with Mr. Sam Altman just before this. I will be touching on that later in detail, but not only for the enterprise, but also medical, education, governments. Of course, that's going to be expanding into every field, including families, consumers later on. We will be seeing the demonstrations from OpenAI after this. You see, agents is going to be working for us 24/7, one after another, executing tasks. Internet so far, just search, email sending. That's something that human needs to work first. But this time, AI autonomously work and execute tasks for you on behalf of yourself 24/7 and keep working all the time.
Even while we are sleeping, agent is going to be work for you. And agent will also pass work to other agent or vice versa. So that's how it's going to work. So I believe that you're going to be enjoying the demonstration after this from OpenAI. And I show you this slide. Sorry, that shows this crystal ball is because so this latest edge AI, we call it Cristal. C-R-Y-S-T-A-L, Cristal, but it'll be intentionally changed from Y to I. So that's how I liked it. And C-R-I, even though the meaning of Cristal is the same despite one alphabet difference, but in Spanish and in French, this is the correct spelling. It's not a typo. The official name is Cristal Intelligence. That's how we call it. Going forward, we are going to expand applications of Cristal.
SoftBank and OpenAI struck a strategic partnership agreement to start deploying Cristal to enterprises so that we can help you to get engaged in AGI and ASI. And for the first time, we want to start selling what we develop from Japan to the outside world. "Japan first" recently is something that we rarely hear, unfortunately. However, AI attracts global attention, and the competition of AGI is getting fiercer and fiercer. And we will begin this initiative in Japan, and we hope that you are going to leverage our product going forward. Again, SoftBank and OpenAI agreed on a strategic partnership to establish a joint venture, share 50%/ 50% each. And this joint venture will be called SB OpenAI Japan. SoftBank Group, including SoftBank Corp led by Miyakawa-san. C reates a company and holding company is equally held by SoftBank and OpenAI. This morning, we just signed an agreement.
When we announced Stargate, right before the day of announcement, we signed a deal on the day, actually. In the morning of that day of announcement, we signed an agreement. Likewise, we did sign this agreement this morning. So the question is, what does Cristal do and how? All enterprise systems, for example, we have group companies like LINE and Yahoo and SoftBank Corp, and Arm. We have about 2,500 enterprise systems running, and each has a distinctive database. So it's been like this for the last 30 years. There is a source code for those systems. So what we are going to do is to have Cristal to read all source codes that have been developed in the last 30 years, whether it be banks or automobile companies. Somebody created an enterprise system, and they wrote source codes.
It would be a very boring job to read source codes that have been written in the past. So the programmers who wrote the code may have passed away already, and you don't know how to rewrite the code or how to fix the bugs if there are. But Cristal will read all source codes and interpret to figure out what the source code means, what kind of functionality the source code has, and where upgrades should take place so that the language will be rewritten to the latest one. So human programmers don't have to write a program and upgrade a version every time. Internally, all source codes in a company can be written by Cristal altogether. And then all meetings, as you know, we are having meetings every day, and Cristal will join, if you will, all meetings.
Sometimes Cristal gives answers if there are questions, or Cristal may join some debate or discussions to share intelligence with the participating employees. When it comes to negotiating with a potential customer, for example, sales team can bring Cristal for their negotiation with customers. Call center. 24/7, a lot of calls are coming in. Sometimes call center agents may not be able to respond quickly or may not be able to answer questions properly. But Cristal can have a conversation with customers directly to solve questions that customers may have. Documents or emails of employees of your company, or when it comes to engineering section, or design papers, requirement documents, all data will be read by Cristal.
When it comes to meetings or negotiations, not only the content or conversation of that particular meeting, maybe the history of negotiation or past meetings, that kind of past record or memories are very important. As you may have heard, there is a technique called Prompt Engineering to ask better questions to AI. So you don't need to worry about Prompt Engineering anymore. You don't have to say specifically exactly what you want to ask. You can ask a high-level question, then Cristal or AI will get back to you in real time, referring to past long memories or some relevant information. That's amazing, isn't it? When your colleagues go on business trips or they leave companies, their memories or experience or history will be gone. Going forward, however, with that agent, long-term memories are going to play an important role in your business.
Coming back to Cristal, Cristal, based upon the long-term memories reside in the enterprise, a Cristal will act accordingly. Remember, I keep saying that agent will be important and important, and AI agent is going to be keyword starting from this year. Please look forward to demonstration by Sam later. Beyond agent, long-term memory will be a key. While talking about long-term memory, in fact, patent was acquired. I submitted a patent 10 years ago around long-term memories, and Sam was surprised. AI experts recently, there is a concept of Reinforcement Learning. Reinforcement Learning is playing central role in GenAI. The central logic of Reinforcement Learning is reward. With that reward, in order to maximize the reward, AI will repetitively learn and enhance their learnings.
For example, when a dolphin plays a trick, and if the dolphin has done a great job, then the dolphin will get reward, and that dolphin will learn more, so that's the basic idea of Reinforcement Learning in the space of AI with Reinforcement Learning plus reward. That idea I actually submitted a patent for, and March 11th of 2015, we applied for a patent, and I remembered I applied, but I didn't remember if I got the patent granted, and I asked my team to check whether I had really gotten the patent granted or not, and it did. Again, I remember I submitted for the first time in the world a patent about Reinforcement Learning, which plays a central role in AI now, and I am the inventor of Reinforcement Learning, actually, so inventor is myself as a result that we had a patent granted.
The first applicant is given. The patent is given. And I just confirmed that, and I'm so happy. That makes my day, actually, that it makes me so happy. So based on that reward, Reinforcement Learning, March 17th, 2015, patented. In addition, weighting based on emotions in long-term memory. That is also patented on June 17th, 2015. And also indexing of the long-term memory was next year, May 13th, 2016. So those three basic patents have already been approved and given, granted. So actually, that makes me so happy. It's not money that we may be able to get because of the patent, but actually, it was the first patent that applied, and that was me. And also, well, the patent will last for 20 years, and we have already spent 10 years. So we have another 10 years that this patent is available.
In the AGI, ASI era, I believe this is going to be important. So all the meetings, all the negotiations, remembering all the long-term memory. And based on that, you can go for the negotiations, or you can go for the decision-making process for the meeting, which I believe is quite interesting as we have such a patent in Japan. So that can be available for the marketing, finance, legals, whatever it is that the Cristal will be helping you with all the wisdom and intelligence. And that can be the brain for the company and become the comprehensive agent. And that is going to be start from Japan together with OpenAI and SoftBank Group. In our group, we have several hundreds of companies and also tens of millions of the customers for the mobile phone service, I believe close to 40 million.
PayPay has close to 70 million customer base. LINE also having about 100 million customer base. About 90 million customers are active users, and 1 billion messages are exchanged using LINE. On top of that, we'll have an agent available in SoftBank Group. Yahoo, LINE, integrating, how are we going to integrate IDs and everything? That's cost you money and time hugely. However, we can have a program, and we try to do the programming to integrate the ID by human, but that is not needed anymore. Cristal will be reading everything. Cristal will be understanding everything so that you don't need to have a system programmed for ID integration anymore. Even intra-group, when you come to technology, sales, different department, HR, compensation system, everything comprehensively gathered. This will be acting as the brain of the company, and Cristal will be working for you.
Of course, it costs money. It's going to be a huge system. It's going to be a huge brain. SoftBank Group, for the development and operation of the system of the Cristal, JPY 45 billion, in dollars $3 billion will be paid by SoftBank Group to OpenAI. There are many news these days that many CapEx is necessary and still making loss. OpenAI can recover those or not. That's a kind of speculating, but just one agreement by the company that you'll be able to have a revenue of JPY 400 billion. There are more than 100 companies, some similar level of the SoftBank Group. If you have 100 companies agreed, if you have 100 Cristal, then it's going to be, with the size of our group, about $3 billion per year. 100 companies makes you $300 billion revenue.
$300 billion in Japanese yen, JPY 45 trillion yen. By the time, system cost, even that you add all that, but still, it's very much paid off. That's how I see. And we will be the first company to make that happen, SoftBank Group, by ourselves, using Cristal for all the system, all the information integrating all together. So in our group, we have ZOZO, PayPay, those companies, several hundreds of companies, and also Arm is one of it. So all Cristal be there as your brain and intelligence and utilize. So I want you to remember Cristal here. So this strategic partnership, as I mentioned in earlier slide, we have just signed the agreement. I'm so happy. It is my happy day. This is the latest and most interesting, most wanted technology and completely comprehensively integrating data inside of the company.
So this is going to be super intelligence for the company, and we are going to make that happen. So excited. I am so excited. Of course, this Cristal, you see that this is a bit showing bright. It's not that we have a chip embedded in here. This is just a branding kind of the things. And also, we may have products later on with using this, and the Cristal image is going to be utilized. So this you see on my hand, this itself is not a chip. So it's not that misleading you or anything, but it's just each enterprise, each business is, for example, today, for example, company A from automobile company, company B from automobile company. So for the company A, the data that you analyze, integrated, will not be reused at the company B.
So it is going to be. If you develop an engine, that's going to be into the Cristal. And you may scare or you may fear that your knowledge, your intelligence may be known by the other competitor, or they may utilize or reuse those information. That will not happen. That's only for the company A. That's only for the specific company, customized for the company. So this is only available for company A or this specific company. So when you see, if you may fear or scared that the information may leak or relearn or reused, that's not happened. That's not the case. So this is going to be customized or fine-tuned service for each respective company. So it does require some time and cost. As a provider, it takes time and effort. And how can we configure agent for you? How can we read source codes?
That kind of activities need to be done before this is made available for you. So SoftBank and OpenAI, which created a joint venture. So by the end of this year, about 1,000 sales engineers will be sent by SoftBank to a joint venture, specifically working for the development. So as the new company, JPY 450 billion of revenue can be recognized because SoftBank is the first customer. So in the first year, over 1,000 employees and $3 billion of revenue. And also, engineers come from OpenAI to specifically work on Cristal. So the infrastructure will be built in Japan. For example, there is a rule around secrecy of personal information. For example, customer data of LINE or Yahoo Japan cannot be used in a data center outside Japan.
In principle, development will take place in the U.S., but when it comes to fine-tuning or building infrastructure, that's what we are going to see in Japan. It's kind of an extension of Stargate project. In Japan, data center will be built in Japan for the AI learning, and the operation will be mainly done by OpenAI. Of course, SoftBank will support OpenAI to build infrastructure. So this effort is an extension of Stargate. When it comes to launch, we cannot address 50,000 companies on day one because of limited resources. So we will start from one company per one industry as a customer so that we can really focus our efforts to fine-tune our product to that particular customer. So if you are interested, again, because of the limited resource, we want to start from one company per one industry.
Of course, we want to expand the efforts going forward. But again, like I said earlier, we make sure that data which was used for you will not be used by somebody else. So the Cristal is made only for you. So please rest assured. Again, if you are interested, please contact SB OpenAI in the future. But in the meantime, there is an enterprise sales team in SoftBank Corp, and I'll make sure that the SoftBank enterprise team will support you if you are interested. Again, that concludes my presentation. So let me call Sam because Sam is going to give you a presentation by himself. So Sam, please come to the stage. OpenAI Inc.
Sam Altman, CEO OpenAI.
Thank you all for being here today. This is an important time in the development of AI. Progress is happening quickly. Models are getting better and better.
We have a five-level system of AI. We started with chatbots. Last year, we launched o1, our first reasoning model. This is a model designed to think before it responds. Just last week, we released o3 -mini to the world, another step forward. Reasoning is useful and exciting for a lot of reasons, but one of them is that models that can reason, models that can think and take multiple steps and deduce, pave the way for AI agents. Now, people have been talking about AI agents for a while. These are AI systems that can do work for you independently, level three. AI agents are designed to observe the world, make decisions, act on behalf of the user. It's like a real digital assistant, something that understands the world around it. So you can give it a task, a complex task.
It can make thoughtful choices and take actions on your behalf. With ChatGPT, we say you could talk to it about anything. With agents, you'll be able to do anything. It's the next evolution beyond ChatGPT, and just like you, these agents understand how the web works, so we were able to launch our first real agent, Operator, recently. Operator can look at a web page, understand what's there, click around, and complete actions for you. It's like an agent that has control of it can look at a computer screen and have control of the mouse and keyboard, and it can really do quite a lot, so we're very excited about this. It expands the usefulness of AI to touch anything you can do on a browser, and soon, a computer more broadly. Now, this is our first agent, but we have more agents to come.
Today, we're excited to demo our next agent. This is called Deep Research, and we announced it earlier today from Tokyo. I think this is one of the best things OpenAI has ever launched, and it really points at what's going to be possible with AI agents. This can do complex research tasks for you, tasks that might take 30 minutes, that might take 30 days. It's powered by o3. It's the first time that the outside world gets to use our o3 model, and it can browse the web, scan text, images, PDFs, much more, synthesize, reason through it, and prepare a report for you. It takes a while. It goes off and does all of this work. You can see what it's thinking about as it goes. It's different than ChatGPT, where you instantly get a response.
Here, you start off a task like you might give a task to a sophisticated coworker, and Deep Research goes off, thinks through it, gathers insights, gets it together, finds sources, and gets you a report. This is a system that I think can do. This is just an estimate of mine, but I think can do a single-digit percentage of all economically valuable tasks in the world. This is a huge step forward for AI, and it really gets at Masa's vision for what enterprise AI can look like. This is just the consumer version. There will be a stronger one to come, so synthesizing knowledge like this is a huge step forward. You can now have an army of research assistants at your disposal to do anything you'd like, and we're going to take this much further. This is available today to professionals in finance, science, law.
It's also useful for people that just need great research. I used it to find a new car, and it was fantastic for that. We're going to demo some of the ways you can use this in a moment. But before that, this is just the next step. There are more to come. This is about synthesizing knowledge. Eventually, AI will be inventing new knowledge, and we think that'll be a phenomenal step forward. The enterprise value here already today, I think, is quite strong. But again, we're going to go much, much further. So without further ado, I'd like to introduce my colleague, Josh, who's going to show you how Deep Research works. This is a live demo. Live demos don't always work, but we're pretty confident in this one. Hope it goes well, and then we'll show you one other thing after. So here's Josh.
Thank you all very much. Thank you, Sam. My name is Josh Tobin. I lead some of our research efforts focused on our Next Generation of Agentic Products, and today, we announced and are soon releasing Deep Research, which is our next agentic capability. Deep Research takes our reasoning models and augments them with the ability to search the web. By searching the web and synthesizing the information they find, they're able to complete a wide range of tasks all under the bucket of knowledge work, and because of this, we think that this is going to be a capability that unlocks many use cases across the enterprise, so I'm going to show you a little bit about how this works and some of the enterprise use cases that we're excited about, so let's start with a sales example.
Suppose that we are an enterprising AI company, and we're hoping to sell our AI tools to companies to help them expand and maybe expand in Japan. So I'm going to start by asking Deep Research, "Can you help me prepare a detailed report that explains how a potential partner of our sales team, in this case, SoftBank, can succeed in the Japanese market using generative AI and agent technologies?" When I send this query to Deep Research, it's going to come back, and it's going to ask me a number of clarifying questions. And the purpose of this is that this technology really excels when it's asked to do highly detailed work that involves incorporating many requirements, synthesizing them, and pulling them together into one detailed research report so that the agent can use those requirements as it searches to build the best possible answer to your query.
You can provide detailed answers to these questions if you like, but you can also just say, "Make some good choices." I'm going to send this off to Deep Research. Now, Deep Research takes a while to run, and that's a good thing. The reason why that's a good thing is because, unlike ChatGPT or kind of previous generation chatbot AI products, Deep Research is able to spend a lot of compute across many searches and a large amount of reasoning to produce a much better answer for you. While this comes back, while this report runs, I'll show you some examples that we ran earlier today to give you a sense of the breadth of the capabilities that are possible beyond just sales. Let's consider a business strategy use case.
In this use case, we are trying to analyze podcast hosting platforms to understand which of these platforms might be the best choice for our business. This is the kind of question that you could imagine having someone on your team go spend days or weeks on or potentially hiring a consulting company to do for you. And like in our previous example, we provide as much detail as we can to allow Deep Research to have the best possible sense of what exactly we need to know when we run this query. And when you run the report, Deep Research will come back with something like this. This is a good kind of typical example of one of the types of things Deep Research is really valuable for.
In this case, Deep Research was able to, through many, many web searches and lots of reasoning, it was able to pull together a table that shows the different platforms that we should consider and breaks down how these platforms compare across all of the criteria that we specified upfront. This is really powerful because this is the kind of work that you would expect an analyst to spend a large amount of time doing. Deep Research is able to complete this for you and save those analysts time and also do it much faster so that you can make faster strategic decisions and consider more options. One of the powerful features of Deep Research is that all of these web searches that it's doing, it provides transparency to you about how it's using that information to produce the final answer.
So you're able to find citations for each of the pieces of information that Deep Research pulled together, and you can go and verify its work or go deeper on any of the questions where you might want a little bit more detail than Deep Research provided. So one of the other kind of powerful features of Deep Research is that you can actually view its reasoning process as well, and you can click in and understand how it came to the final answer, which I'll show you in a second. So this is a business strategy use case, and this indicates one of the powerful capabilities of Deep Research, which is broad kind of questions that require synthesizing a large amount of information in an ambiguous setting to provide an answer to a business question.
But Deep Research also excels at finding answers that are difficult to find about very detailed questions. So this is an example in M&A. So suppose that we are looking into land power deals for maybe a large data center that we might want to build, for example. And we have a hyper-specific question here that requires a lot of industry knowledge to answer in an accurate way. Deep Research excels at these types of queries as well because it's able to find rare and difficult-to-find sources of information on the web and pull those into its final answer. And so you can see the report here provides a detailed analysis across a number of criteria and even breaking this down over different regions and other ways of splitting this query into subcomponents that might be valuable for us.
So this is an M&A example, but let me show you one that is a little bit more relevant to our team at OpenAI today. So we announced Deep Research today, and one thing our marketing team might want to do is to understand, "What are people saying about this? How are people reacting to this launch? Is the media positive on it? Are people on social media excited about it?" And Deep Research excels at these types of use cases too because it's able to access all of these sources of information and, instead of pulling together all of them, synthesize key points across them.
This is the kind of task that marketing teams can do, but oftentimes, when we're busy with a launch, we don't actually have the bandwidth for our marketing team to do this kind of detailed analysis for every single thing that we would like them to. So Deep Research enables marketing teams to scale their efforts and serve a much wider set of use cases. Hopefully, this gives you an indication of the types of use cases, the breadth of the use cases that we think make this technology so exciting for enterprise. But Deep Research is useful beyond enterprise as well. So it's also useful in our personal lives.
If you have a question that you want to ask about a hobby of yours or sports or you're doing some shopping and it's a similar type of question where you have a kind of very detailed set of criteria that you're looking for an answer to, Deep Research can pull together answers for these things too, such as summarizing baseball statistics and making comparisons across a large number of baseball players. Beyond just enterprise and consumer use cases, one of the things we're really excited about Deep Research enabling is accelerating scientific research. So here's an example that we put together in a domain that's very familiar to me, that's kind of closely related to my PhD research in deep learning and robotics.
And the report that it pulls together, a lot of experts describe these reports as being kind of at the level of an advanced undergraduate or an early graduate student level work because it pulls together many sources and summarizes them in a way that requires sort of a sophisticated understanding of the nuances and the details between these different use cases. So this is an indication of the breadth of the types of things you can use Deep Research for. Now, let's come back and check on the go-to-market strategy question, the sales question that we asked in the beginning. But before we look at the final answer, just to show you kind of how much of a leap forward in capability this is, I'm going to try asking the same question to regular ChatGPT without using Deep Research. So conduct a detailed analysis.
Actually, I will just jump in here and I'll copy the exact one so we can get a fair comparison. You can already see how detailed this report is, so we'll type in the same query, and the advantage of using a chat model over using Deep Research is that you get an answer much faster, so here we're going to get some, so here's kind of our ChatGPT model giving a high-level analysis of this strategic question. This is kind of the type of answer that you might expect someone who is thoughtful and has read a lot about SoftBank to answer kind of off the top of their head in a first pass, so if you want an answer fast, using chat models is a great way to do that, but if you want a detailed answer, then Deep Research provides a much better solution.
And so you can see even just by the length of this, how much more research and work has gone into producing this answer. And the level of detail and insight that this goes into is far beyond what you can get out of the box from ChatGPT, including industry focuses, financing of these deals, business strategies. Each one of these claims, each one of these insights backed up by supporting evidence, just like an analyst that you hire at your own company would do. So this is an overview of how you can use Deep Research in enterprise use cases. And as you can imagine, this is only the beginning for these types of agentic technologies transforming enterprise. Giving models access to web search unlocks many use cases, but not all use cases. In enterprise, oftentimes, the most valuable data is the data that you have in your company.
So you can imagine us starting to expand this to accessing other types of information like internal information. And you can imagine these agents going beyond just synthesizing knowledge, starting to create knowledge or take actions in the world as well. So that's the roadmap of where we're hoping to go with this technology. And so from there, I'll hand it over to Michael to talk a little bit more about how enterprises can start to customize these applications and how we think about that ourselves. Thank you.
Hi. Thank you. It's an honor to be here and show this to you. So 2025 is the year of agents. But what does that mean? We're all sitting in this room right now, not pulling out our phones and typing into Deep Research or ChatGPT. So I wanted to show you a different take on what the future of your organizations could look like. This is a small demonstration just to hopefully inspire you and give you something you can bring back as you look ahead to the rest of this year. Let's start with something we know: sales. Here we have a sales contact form where someone is trying to reach out to OpenAI to learn about ChatGPT Enterprise. Now, we know how this usually works.
Someone submits this form, it goes into a system, and then a team of people, account associates, are reviewing the leads to reach out and figure out how best to respond, and this can take hours or days and feel like a slow process for your buyers. What could it look like if you had agents inside of your workforce working for you right now, so first we'll submit the lead here, and we'll come over to our system. Here you can see what a virtual teammate, a virtual sales associate could be doing for you with our technology. If we open the task here, we'll see that the agent is picking up work for you already to try to help this lead.
The account email came in, and already it's doing what a sales associate might do on their own and researching that lead, figuring out the industry, the revenue, title, and other information. Think back to Deep Research that you just saw and imagine the ways that this could plug into this as well. It's really exciting. Now, in this case, the agent was able to see that this is a really good prospect. OpenAI wants to buy OpenAI, so we think they're a good customer. And so we're going to get the calendar availability and figure out when we might hop on the phone with them, then it's going to write an email and get back to them. Now, in this case, this agent hasn't been told to write in any specific language. It's really smart.
It realizes that the prospect wrote in Japanese, so it's going to write back in Japanese as well, and sure enough, if we come over to our inbox, we can see that email right here waiting for you. Again, this is just a small demonstration that we wanted to pull together to help you see what it could look like. You can imagine the applications for this are endless. If you start to look around your organization, think about all the small tasks that people in your teams do and the ways that those add up. This doesn't have to be an all-or-nothing process. You can really chip away at this and bring the latest technology into your workforce now and help save your team time and focus on the next steps after that.
I hope this is inspiring, and I can't wait to hear what you do with this, Deep Research, and everything else. Thank you.
We are going to have Rene Haas, CEO of Arm Holdings.
[Foreign language] Very nice to be here today and speak with you about Arm and how we fit into everything going on with agents and the things that Masa and Sam just talked about. Now, I think you may know Arm well, but just to remind everyone about our company, we are a compute platform with unmatched scale. Since we were started in 1990, over 300 billion chips have been shipped with Arm inside. No compute platform even comes close. And now, as we move into the age of artificial intelligence and all the connected devices, 99% of the connected global population runs AI on some form and some way on Arm. And we have a developer community unlike anything that has ever been created for the compute platforms: over 20 million developers.
So the last 35 years of Arm have just been incredible in terms of our growth, our breadth, market penetration. But we think the future is just beginning. And agents and AI running on Arm is the future. I'd like to show you a short video about how we see the future and the vision about agents and Arm.
AI agents are changing how we live every day. They help us shop, book, drive, stay healthy, and even build the technology of tomorrow. And at the heart of it all, Arm's compute platform. Shopping is easier than ever. AI agents running on Arm know what you need, find the best deals, and even let you walk out without stopping at a register. Reservations, adjustments, and even remembering your favorite lunch spot handled instantly, all so you can focus on enjoying the moment. Making driving smoother and safer, helping you navigate, park, and even take the wheel when you need it. It's all running on Arm: early detection, real-time monitoring, and life-saving support for doctors. Smarter technology leads to better health outcomes. AI agents don't just run technology. They help build it. Arm-powered AI is transforming how developers code, test, and create.
From how we shop to how we move, eat, stay healthy, and create, AI agents are changing our lives. The future of AI is built on Arm.
So one of the very important things when we think about agents and the compute platform is that these agents running everywhere will require more and more power-efficient compute because the devices that we have today still need to run displays. They still need to run operating systems. They still need to run applications. But yet, on top of that, the agents are going to be making our lives much easier, but will require the most power-efficient compute, and that's where Arm fits in. Now, we are the compute leader from cloud to edge. When you're talking about the largest data centers or the smallest embedded devices, such as thermostats or security cameras or earbuds, the world is using Arm in all these areas. Now, one of the unique attributes of Arm is the fact that the software is common across many of these platforms.
The operating systems that run on your phone or run on your PC or in your automobile or even in the cloud, that is the key in terms of how these devices go. But in the future, these agents will be running on top of the operating systems. So some of the very interesting demos that we just saw really begin to abstract away some of the things that are going on at the software level. Now, that is magic to the user, but it takes a lot of hard work to really make it all go. And one of the things that Arm brings to this problem is a solution we call Arm KleidiAI. And this allows the agents to just work.
That is, for the developer who's writing the agent, to be able to write it in such a way that knowing that whether they're running it on a phone or a PC, the data center, a car, it's just going to run. We envision a world where these agents will be running everywhere, not just on your PC where you can query demands, but agents will be talking to agents and other agents and other agents, and our job at Arm is just to make that very easy for developers and make it very power-efficient, so I think the future is so bright for this technology. I just want to bring Sam back up here for a moment and just chat for a second because I think when we look at the future of agents, Sam, you showed some pretty cool demos there.
I can imagine a world where the agents are just not running on a sophisticated device, but probably almost every device you can think of.
Yeah, I think the network of agents will be the next thing we talk about, and it'll run in the cloud on device sort of all over the place. But that seems like it's going to be pretty incredible.
Now, I don't want to put you on the spot here, but.
Go ahead.
Every time we chat, you say, "Hey, I did this really cool new demo I need to show you," and just looking at this Deep Research, when do you think we see a world of agents talking to agents across embedded devices?
Technically, it should be capable now if you could run a big enough model on device. So it's a better question for you.
All right, we got to get the hardware ready.
Yeah. But I think one of my biggest surprises about AI over the last few years has been how much we're able to do with a smallish model. Distillation has just been incredible to watch. So I am optimistic that every device in the world is going to be pretty smart.
Yeah, the future could not be more exciting. Thank you.
Thank you.
Thanks, all.
Thank you very much. Next, we would like to move on to talk sessions, so let us have some time to be prepared and bear with us a moment. Thank you. Now, the speaker is going to be on the stage. Thank you. Thank you. Thank you very much, everyone. Now, I would like to have a free discussion with Sam, so there are many things that you have asked. I will be asking questions to Sam on behalf of you, so I think that I'm going to be covering many things that you may want to know. I would like to ask him directly, so please enjoy. Sam, please come on the stage. Enjoy. Yeah, yeah. Great. Yeah, so I'm very, very excited that we were able to announce today.
Yes, me too.
Yeah, yeah. So how did you feel about the Stargate announcement?
That was quite a moment. It really kind of came, you know, it was very so cool to be in there and people were excited.
We were talking, can we really make that happen? It really happened.
We've been talking about doing this for so long, but to finally get it all done and get it out into the world, I think, is wonderful. Yeah, and the world was going to just need so much compute. It's true that we can, as I was saying just a few minutes ago, get small models to do incredible things, but to really push the frontier of intelligence, that's going to require a huge amount of compute, and the most value will get created at that frontier, so we need a ton of compute to make these models. People are clearly going to want a ton of compute to run these models, and to finally be doing this at scale is totally great. Yeah, so I felt really good about it.
Yeah, yeah. So about a year and a half ago, we were having dinner, and we were talking, Sam, so when is the AGI coming, how big the compute should be, and the answer from you and the team was, more is better, right? More is better. That was a simple answer. And I started thinking, well, if the more is better, we should do a lot.
Now we're doing a lot.
Yeah. That's how we started.
It is.
Yeah, yeah. So it was not a limited amount of compute. It's the more is better because more breadth is definitely better, right? Some people say, well, you can do small compress, but that's small.
The front, I think people still don't understand how much, how exponential the return is. The cost is exponential too, but I think the return is even more exponential to the smartest model we can make, and that will require the biggest computer.
Yeah. Well, this reminds me of the beginning of the internet. When we started our internet in 1995, it was just a PC with just big letters and very, very slow, very expensive. And then when the broadband came, people said, why do we need that much capacity and the bandwidth? And with more bandwidth capacity, people said, well, this is enough. This is not growing anymore. But then the picture came, more high-resolution pictures, and then the video stuff. The capacity requirement went on and on and on. And people initially saying, oh, internet is just a virtual stuff. It's not really useful. It was mostly free service, so there is no business model. All those criticisms seem nonsense.
It seems nonsense now.
Now, all the GAFAs.
I think we'll see the same thing with intelligence. People are like, oh, how much how smart does it need to be? And the answer is very smart. Yeah, and people use a lot of it, and they'll be generating tons of video and solving really hard problems, and everything will be really smart in the world, so.
Yeah, your model is actually improving quite a bit, right?
Yeah.
Like 10 x a year, kind of, you know, model? What's your measurement?
You know, very roughly, it feels to me like this is, like, not scientifically accurate. This is just sort of a vibe or spiritual answer. But every year, we move one standard deviation of IQ. Also, every year, the cost of last year's intelligence falls by about a factor of 10.
Yes. Yeah, so per chip price, cost becomes one tenth, meaning we can have, with the same budget, we can have 10x more chip, right?
I think this is, yeah, totally, but also the algorithms get more efficient too. This compounds itself. The rate at which this is happening, I think, is easy to take for granted. In 2018 and 2019, we had GPT-1 and GPT-2, and people looked at them, and it didn't feel that serious. GPT-3 came out. I think that was the first time some people noticed, but GPT-3 barely worked, and if you go back and play with it now, it's like using, you know, I went to one of these old computer museums recently, and I got to use a Xerox Alto. I think it was 50 years old, and you could see kind of how it did some stuff, and there were the inklings of a modern computer in there, but it was 50 years ago, and it now feels like a 50-year-old computer.
GPT-3 is only a few years old, and it feels, yeah, if you use it now, it feels like this joke. ChatGPT is only about two years old. It came out at the very end of November of 2022. GPT-4 didn't come out until March of 2023, I think. And so if you just look at the progress here, how quickly the model's gotten better, and also how quickly the models have gotten cheaper, it really points, if we can stay on that curve, it really points to an incredible future.
Yeah. To me, it seems like your model is improving like 10x a year. And the performance, actually, chip itself, with Jensen's effort, the industry effort, is becoming 10x. And then with the Stargate, we are actually increasing the number of chips 10x like a year. So 10 x 10 x 10 is like 1,000x in a year or two. And then the next year, again, we have another 10 x 10 x 10. That's another 1,000. So 1,000 x 1,000 is 1,000,000 x. So if you do once, twice, three times, 1,000 x 1,000 x 1,000 is 1,000,000,000 x, right? So people may say, well, with the recent announcement of DeepSeek, oh, they can sort of mimic and try to catch up. You know, a year later, it comes out. It's so much cheaper.
But you are still going ahead dramatically more with this o3, o4, maybe sometime soon. So people don't realize the level of exponential, right?
It is hard to really feel the exponential when you're living on it because you can adapt so quickly, but we clearly are on a very steep one.
It's amazing, amazing. So like 1,000,000,000 x is coming in just a few iterations. But think about the next 10 years. It's going to be amazing superintelligence, right, that people cannot imagine.
I think so.
Today.
Yeah.
Because people tend to think linearly. When exponential comes, it goes beyond people's imagination.
I think so.
Yeah? You are front runner of that.
It is hard to really feel that. But I have learned over my career again and again and again, you just have to trust the exponential.
Yeah.
We're not built to conceptualize it, but you just have to trust.
So, you are still excited the level of innovation yet to come. It's not reached.
More than ever.
Saturation.
You know, no, no. We're going to look back in a few years at o3 and be like, man, can you believe how bad that was? Yeah.
Yeah. So people think, oh, bringing agent, prompting, oh, that's too difficult, not for me. But actually, this level of innovation makes it easier, right? So users don't have to really do implementation by themselves. It comes more and more friendly. Like we are talking here with the voice and looking at the eyes of each other. We start talking with our artificial intelligence with voice and the eyes.
Totally.
Right?
Yeah. Like it's amazing how much value people have gotten just out of a text box. But the world is not just a text box. So we will add all of those things.
Yeah. Like talking to this Cristal.
There you go.
Right? You just talk, and it sees you. It sees your face, and it understands the tone of the voice, and like we are communicating, it will basically communicate with the voice and the emotions and surrounding looking at by itself, right? Talking to us. Yeah, that's really happening very, very soon.
I think so.
Yeah. Well, some people say, oh, Stargate, too much CapEx. How do you bring the money? Masayoshi, do you have enough money? Right? So what do you think? We still need a lot of capacity, a lot of upside potential to get the technology out of it, right?
Yeah. Again, this is the point I was trying to make earlier. I think the returns on linearly increasing intelligence are exponential in terms of value. So pushing each bit we can push the intelligence of these models further, there's so much more value created in the economy. And yes, it takes a lot of CapEx, but the revenue goes like that too.
Yeah, yeah, yeah, yeah. Well, our mutual friend Elon Musk was there.
We're mutual friends.
You know, he says, Masa, do you have enough money? I will tell you, we will make it happen. We are not a bank, but we are a SoftBank.
I have no doubt.
We will make it happen. So now, the Stargate have to also expand into Japan because of the regulation. We have to respect the national security, the privacy law, blah, blah, blah.
Yeah, SoftBank is building a big data center here.
Yes, yes. So we're going to expand.
We're excited to run on that.
Stargate into Japanese infrastructure also, right? The innovation center of innovation is happening. The training main brain is happening in the States. But there are other people in each country. There are other cultures, national security. I believe we should expand this, not just Japan, to the other sovereign, respect to their culture and their national security, right?
We certainly do want, you know, we started obviously as an American effort, but our mission has always been AGI for all of humanity. And we really want to find ways that our systems reflect all of humanity and the different values and cultures and languages.
I was amazed when I took a picture in some part of Japan and said, "Do you know where it is?" Actually, you know, o1 at the time said, "Oh, this must be this place." I said, "How did you understand? Did you use a GPS?" Well, it says, "No, I did not use GPS. I looked at the stone and the moss on the stone and how the stones are stuck to each other. It must be this culture in 500 years ago in this historical location. It has right on.
Pretty good.
I was so amazed. I got brought up. How could Sam know Japan in this? Oh, my God.
Pretty good.
So smart. Amazing, huh? So the inference, right, prediction, inference, not based on all the detailed data, but guessing and guessing, guessing, make it right on in the historical landmark.
Fantastic.
Amazing. I got blown away. It even understood my joke so it was, I text, I actually spoke, said, can you make a joke in Osaka language? In Japan, there are dialects and start making a joke in Osaka dialect. And it says why it is funny. Explain to me. Oh, my God.
Fantastic.
It even understands the context, the culture. It's already now. But going forward, you know, I'm using it every day, but I get blown away almost every day still. It's amazing, amazing.
Fantastic.
OK. So we announced the Cristal today. When we do all kinds of source code reading of 2,500 systems, just within our own group, so many source code, billions of coding lines. It must take a lot of compute.
A lot.
Right? A lot of compute. But you are confident that if we have some capacity in Japan, then reading all of the source code of 30 years for your model, you are confident that you can do it.
Yeah. We're confident we can do it.
Cool. Look. Sam, [Foreign language].
Sam just simply said you're confident.
Cool. "Said, yeah. Done." I said, "It's amazing, you know? People would expect, oh. Right? But you said, yeah. You're so confident.
You did that too.
You're so confident. So I'm very, very happy that we can read all of the source code, but participate real time on the meeting with a long-term memory. We don't have long-term memory yet. But when do you think long-term memory thing can happen?
Definitely within the next couple of years.
Within the next couple.
Maybe even faster than that. Yeah. Having these models have infinite long-term memory, that is so important. An AI that can get to understand your entire life or an entire company and entire enterprise, that'll be a huge step forward. We're working hard on that.
You know, my pattern, what my pattern, the concept of my pattern for long-term memory is that as we are talking right now, OK, I can see a facial expression, emotion, tone of the voice. So all the conversation, I change to text. But understanding tone of the voice and facial expression, I have an emotional map with 250 kinds of emotion and then indexing, and with each of the index, like fear or anger or doubtful, there are about 250 words for expressing emotions, and each emotion, how angry you are with a 1- 10 scale. If you are so angry or so doubtful, 10 or three, I put the index of the strength of that emotion. Analyzing 250 emotions and the strength of the emotion and make it into a numerical index.
Text with just three numbers of numerical index attacks, then you can express, I mean, compress conversation. And then when you have a very strong emotional vibration, like you're so angry or upset, the multimodal understanding, including video, captures the whole thing, captures and stores as a long-term memory. But if you're saying, hey, good morning, good night, like driving on the commute, everyday drive, you're supposed to forget the traffic light or the car passing by. Human brain forgets all of those. Otherwise, our capacity of the brain explodes. So you compress all those not important ones, but the one with a surprise or a big emotional strength, that's the one you, without too much compression, you even capture and store the multimodal video and voice and sound, everything. So like your three-year-old kid's birthday, you're supposed to remember that, right? It's a happy moment for the family.
So it will automatically capture and store the multimodal data. So that's a long-term memory. And the key is the level of surprise or level of emotion with the index. So emotion, the human communicates with emotion, not just the text. Like I like you. I like you. I like you. Completely opposite meaning, right? So the tone of the voice, facial expression. And then if you put the index, that makes the compression and long-term memory. And that context can be very useful for the next conversation, next discussion, negotiation. Like negotiation, you have to read the emotion of the other side, right? Otherwise, you fail.
Yep.
So this is the long-term memory with the emotional trigger. That's what I filed 10 years ago.
Good job.
It should be useful very soon, right?
Very soon. Yeah. I think, I mean, I don't know this, but I think that AI that has emotional expression, so not just like texting in a chatbot, but when you see the motions of like a rendered video avatar or something, that's going to hit us more than we think, and we're going to have to develop some new societal guardrails for it, but it'll also be tremendously exciting.
Yeah, yeah. Our friend Jony is supposed to make.
He'll figure it out.
He's such a timer, right? Yeah, yeah. Yeah. I'm very much excited to see that. So if we have all the data and long-term memory and so on, we need lots of capacity, but also latency becomes very important. Like a call center, customer care call center. We have to have an instantaneous response. Are you confident, let's say in Japan, with so much enterprise mission critical, are you confident?
You know, I used to worry about that a lot. But even if you use our Voice Mode today, it feels like talking to a real person. It's quick.
It's very, very good now.
So I think we'll be able to solve this.
Yeah. Only several months ago, it was still lots of today. Like even last night, I used, I said, wow.
It's very good now.
All three meaning, you know, wow, it's so fast. So the latency is now about 100 milliseconds or what? Something like that, right?
Something like that. A little bit more maybe, but it's quick.
Yeah. 100 to 200 milliseconds. Human conversation is about 200 milliseconds, I think. So 100 milliseconds to 200 milliseconds is almost human interactive, and you can even still interact.
You can interrupt. That works well.
That's the key, right? Because humans also interact, and it's really happening, so you are confident. Even the model trained in the U.S. and Japan with the Stargate Japan center, the response of all this real time, you're confident.
Yeah. Obviously, we'll have to run the model for very low latency things closer to where people are going to use it. But as you said, we can train in the U.S. We can run a lot of things from the U.S., especially where it's thinking. And then some use cases we'll have to put out towards the edge.
Yeah, yeah, yeah. So whatever non-national security kind of thing, you can still do in the U.S. And national security and privacy things can happen locally in Japan.
Yeah. We can certainly deploy models around the world.
Yeah, yeah, yeah. So we would allocate 1,000 sales engineers.
To the new.
Yeah, with this new joint venture. Those guys have to do the implementation setups to each of the systems to establish the agents for each task. So explain a little bit more about how the agent works. Is it a single task agent or a very sophisticated agent or what?
So there will be generic agents that consumers use. And those can do powerful things. Like we just looked at Deep Research browsing the web. But what you might want for your companies, or I think what everyone will want, is an agent that can act with as much context and information and power as an employee at the company would have. And so you need to connect it to all the systems. You need to give it all the knowledge base. It needs access to the code. It needs to understand how the company works. And that will take a lot of customization work for each company. But think about what can happen once you have it. So if someone builds this and integrates it into, let's say, SoftBank. And let's say there's SoftBank, and then there's some imaginary competitor that hasn't done this.
SoftBank can now do so much more.
Yeah.
And so once you've integrated AI into the workforce, and you have all the power of that, and it's not just the deep research browsing the web or a coding agent writing generic code, but fully integrated into the company, that's going to be very powerful.
The one with the best tool, the one without, is dramatic. It's like a country with electricity, with no electricity.
Yeah.
Right?
Yes.
The country with automotive and bicycle. It's a huge difference in the productivity that you think will happen again here, right? Truly.
I think it will be. I think it is one of these moments. You mentioned swords. I collect sort of ancient technological artifacts. And during the Bronze Age, one of the things I have is a sword from the very beginning of that. And they were able to not just forge the blade, but also cast the handle. And so you had swords that had a metal handle that was sort of attached to the blade. And what that meant is you could swing.
Yeah.
Rather than the people that just had a forged blade and a wooden handle, which if you swung would break, so you just had to jab. And it's an example of technology giving this decisive edge all at once. And in a matter of a few decades, I think it changed Europe. I think AI is a technology on this order. And companies that don't integrate it will have a hard time competing against the companies who do.
Yeah. So not just country and company, but recent example with the DeepSeek as an example. Now, you care so much about protecting human security and not to make dangerous output. You try to not answer the wrong way because that can dramatically make dangerous decisions, blah, blah. So the technology and output looks 99% similar, but the one with a lot of human safety features to protect the mankind or to protect the national security. Like debugging, it's a lot of effort for the last 1%, 2% fine tuning, right?
It is, yeah. Society is going to have to figure out what the boundaries are here. We do care a lot about it. And it is a lot of effort to get that right. But people are happy to use it once we do.
Should be. Should be, right? And I don't want to go into politics that much, but depending on the country, there are very dangerous situations that could happen if they're used wrongly.
Yeah.
Right? It could be a trigger of a very bad future for mankind, like very fearful wars.
I think we'll get it right. I think we collectively will get it right.
Yeah. Well, you care a lot about that. So these agents and Cristal and this AI, is it for cost? Some people ask, is it for cost saving? Does it eliminate jobs and so on? You must be getting asked that question many times. What's your answer?
Look, it will save money, but that's not the exciting part. The exciting part is how much more we'll be able to do and how much more we can achieve. It's great to free people up and let them do more ambitious things. And we see this like every technological revolution. People worry a lot, and they say, what is this going to mean for all of the jobs? And then we always find new things to do. And that's wonderful. And people will just achieve at a higher and higher level, and we'll expect more. But AI will make things way more efficient, and that's great. The economy benefits from that. The thing that I'm most excited for personally is these systems can help us create new knowledge that we couldn't handle on our own. We couldn't do on our own.
If the rate of scientific progress can materially increase, so we make a decade's worth of scientific discoveries in one year, and the next year we make a century's worth of scientific discoveries, that will have such an impact on quality of life, on the economy, and that's not just like making something cheaper. That's something we just couldn't do before at all.
Yeah.
We just are not smart enough without this new tool.
So you announced five levels of AGI improvements. Now, I think the third one was the agent.
We just started that.
Which just started this year.
Yeah.
So this year is the year of.
I mean, kind of like today or last week.
Yeah, yeah. So this is the year for the agent. But the next one, you say, is Innovator, right?
Yeah.
So explain a little bit more about Innovator. How does it work?
So today, our AI systems, they're very good at synthesizing existing knowledge. And they're very good at doing things that are similar to things that have been done before. But they're not making new scientific discoveries yet. And that's our next level. That's Innovators. And I think that'll be transformational to society. So we're going to go. We've got a lot of work left to do with agents this year, but next we're going to go work on that hard.
Yeah. So some skeptical people say, oh, AI has a limit because humans have to teach. So how can it become smarter than humans? That's the limit that AI can go. But now innovators will innovate, invent things that we did not have in the past for the solutions. So explain a little bit more the mechanism of how the Innovator would innovate things, like exploring, right? You have a feature for the exploring mechanism.
I think it'll work a lot like how it works with humans. If you're trying to figure out a solution to a problem you haven't solved before, you start thinking of a bunch of ideas, and you kind of notice some connections, or you build off your previous knowledge, and you say, that didn't work, that didn't work. Oh, that's kind of interesting. Let me go a little bit further. No, that didn't work. Oh, this seems promising. And then once I have that, like, oh, I can go to here and here and here. That seems really good. So I'll go further in that direction. And the process of human creativity, I think, I mean, it doesn't always feel like this from a sort of self-perception standpoint, but I think it's something like that.
It is like trying a lot of small modifications to existing things and building it on the ones that are promising, and I think we can do that with AI.
Yeah. So the reasoning is the first step, right? Reasoning. You do the three steps, 10 steps, 100 steps reasoning. And then when humans innovate things, we try out, as you say, we try out something different from a different angle, right? And that's an exploring concept. I have filed 1,008 patents in 12 months last year. In my mind, I explore so many different. I force my right-hand side of the brain to think different from the forcing mechanism to think different, right? That was the key to the innovation. And this AI, the agentic reasoning effort can force the different trial, right? Explore. I think that is a key for your Innovator, right?
Yes.
Trial and error of many, many, many, many billions of trial and errors.
Yes.
That's once in a while you hit the right solution. That's invention, right?
Very much.
That's how Innovator must be working.
Yep.
Yeah. Okay. Right on. I understood. I thought that was the case. So I think I figured out how you are preparing.
We'll try soon.
Yeah. Very good, very good. Maybe I shouldn't say too much.
That's okay.
For maybe some of your secret of how you're developing. So then the fifth level, you say, is the organizational. So agent to agent co-work, right? That's.
Yeah. Rene and I were talking a little about that earlier. But the idea of many agents or many innovators working together. If you think about the number of minds that can run in one data center, all talking to each other, building off of each other's ideas, bringing different expertise together, you can easily imagine like a virtual company running.
Yes.
Yes, and then things can be quite powerful.
For our Cristal for SoftBank, my image is to create a billion agents just internally for SoftBank because we have 100 million accounts for LINE, right? 40 million mobile customers, 70 million PayPay users, and so if each one of those accounts, each one of those functions have 10 functions, 100 functions, each function should be able to allocate agents doing a simple task, right, so instead of making two sophisticated tasks with one agent, you allocate a simple task. Many, many, many, many, so that's why I have an image of a billion agents just for our Cristal inside SoftBank Group. That's a lot of agents, but capacity-wise, it shouldn't be a problem because each agent is an integration of simple tasks, right? Our computer is very good at that.
Again, I think we have a lot to learn here, but directionally, I agree. And I think we'll figure it out.
Yeah. So that's my internal image goal. I want to have a billion agents with Cristal just within our internal use, okay? Once we have perfected that experience, then we can be an evangelist to the other customers. This is how we improved our efficiency, and they can utilize that. That's the image I have on Cristal. The direction, that's what you think is.
Yeah. Yeah, let's go do it.
Yeah, yeah, let's go do it. Let's go do it. Okay. So we have just a few more minutes. What about cybersecurity? Now, there is always a bad guy, right? And try to attack, to do something bad for the other people, intentionally or by mistake. We have to protect. More and more people's lives depend on this super intelligence. How do you?
As AI starts to get really good at programming, clearly, it's going to be used for cyberattacks, and so cyber defense is something we need to stay ahead of. I am optimistic that AI can contribute a lot, but it is harder to be on defense than offense, so I think you bring up a great point, and the world has got to start to take this very seriously quickly.
Yeah. Because there's always a bad guy.
It's a big risk.
Yeah. I'm optimistic too, okay? So there are 99% good humans. There are always 1% bad humans. And it's the continuous, endless effort to protect 99% good people from 1% bad guys. But with the level of innovation that good guys continue to try to do together with innovators of our super intelligence, there's always a solution improved. Like when we have automotive motorization, there's a car accident, blah, blah. We humans create regulations, the etiquette, morale, our custom learning. I think that's why you say the regulations, the healthy regulation is always needed. Not too much restriction. Innovation should be given opportunity. But still, we have to have healthy regulations, right? Your comment?
We do. I strongly agree with all of that.
Yeah. People were surprised when you said, Oh, our industry needs regulation. People did not expect.
Well, regulation always comes for important industries. But I think getting it right, if we get it wrong, either way, too slow or way too much, either of those could be bad. And so I think talking about how to get it right.
Yeah. Reasonable. Within the healthy regulation. And it should not overly regulate it so that it kills the innovation speed, right? Okay. So we talked about those innovations. What about the medical? What's your view on our AGI for solving medical?
This is one of the areas that I am most excited about. The idea that we can provide great healthcare to every person on Earth. The idea that we can go cure or treat many diseases, maybe someday all diseases. I think this is within reach, and everybody's got a story about how this would have been great in their own lives or their family's lives, and I think we can finally deliver it. I think this will be one of the biggest triumphs of AI.
That's great. We have to solve. I lost my father from cancer a little over a year ago. It was so sad. Why we cannot solve these difficult issues if our AI can help humans protect from cancer or other difficult diseases? It reduces our sadness. Definitely good for humans.
Absolutely.
Yeah. Well, how about robotics? You love robots. I love robots.
One of your favorites.
Yeah.
Look, I've wanted, like everybody. I've wanted robots for a long time, and it's always felt difficult. I think now the AI is getting. We can build the body, but the brain has been really hard, and I think it's within reach, so I think in a few years, we can have really great humanoid robots and lots of other kinds of robots too, and that will also change the world.
Yeah, so we humans don't have to do dangerous jobs, the hard job, sweating job, boring job, and people say, then what's left for humans to work? What's your comment on that?
We always find new jobs. We always, always find new jobs. If you think about many of our jobs in this room today, if you were a person 500 years ago or 1,000 years ago, that person would look at what we're doing and say, that's not really work. They feel very busy. They feel very important. But they're not doing that to survive. They're playing a game. They're doing it for whatever reason. And I hope that we look at people in the future like that. Yeah, and that with AI taking care of many of the things that happen today, the people in the future do more interesting things, and we say, like, that's so ridiculous. Why do you need a whole galaxy?
Yeah. Totally agree. What about education? In the beginning of your introduction of ChatGPT, many schools tried to prohibit the use of ChatGPT to their kids at the school, and what did you think? What was your comments?
I understand why people looked at this and say, the whole world has changed. Students can have ChatGPT write a paper for them. What does that mean? Very quickly, teachers and administrators who had banned ChatGPT said, "Wait, that was a big mistake." We're going to go the other direction. We're going to go all in. This is the future. Students need to learn how to use it. We're going to change our whole curriculum. Now it's like part of education.
Yes.
And it's delivering amazing results. And I'm sure that will keep going.
Yeah. I'm using ChatGPT o1, o3 every day. More I use it, actually my brain starts thinking new conversations like we are conversation brainstorming with ChatGPT o1, o3. Actually, your brain starts to function more. Kids can learn more. Instead of some people say, oh, with this, the kids will no longer study. I think it's completely opposite, right?
I agree. Yeah, it's been, I mean, definitely, there are some kids who try to use ChatGPT to do as little work as possible. But on the whole, I think people are going to learn more, achieve more, be capable of more.
Yeah, like debating. You learn more by discussing, right? By discussing, by debating.
For sure, and this is part, I mean, it's part of the world now. This is how people are going to do everything, and it really is amazing to watch young people use ChatGPT. It's like a completely different way of working on problems than I grew up with.
Yeah, yeah, yeah. Well, we talked about emotions, okay? So do you think our AGI, ASI will start to understand, start to have emotion by itself? What's your comment?
I personally don't think so, but maybe something like it.
I actually think.
You think it will?
I think it will. Even dog has emotion. I don't know if fish has emotion. Maybe fish also has emotion. Because when dangerous enemy comes, fish escape, right? So I think emotion is a very, very important thing to have more output, more efficiency, protect themselves. If dog did not have emotion, do you think the dog is cute? Dog is lovely if the dog did not have emotion. If the dog does not have emotion, it will start to bite.
I think it will feel to us like AI has emotion.
Already.
No, no. Well, yeah, maybe people would already say it does. But certainly, at some point, it'll feel like it does. And whether it does or not, that's going to be a big philosophical debate.
Well, I would say this is my bet, okay? In the next several years, it will start to gradually. People said, oh, ChatGPT does not understand context. Now people say, oh, it actually understands the context, okay? Because initially, people say, oh, there are lots of delusions, the hallucination, lots of hallucinations. So it does not really understand the context. Now, with the reasoning and so on, people say, oh, wow, it actually understands the context, okay? So I would bet you in the next several years, 10 years, it will gradually start to have at least understand the people's emotion, and then gradually, it will start to have emotion by itself, and it's a good thing to protect humans. People think, oh, if it has emotion, it's a disaster. It's the bad thing, for that's the end of humans because they're going to fight and kill you, destroy you.
But I would say if their source of energy was protein, then it's dangerous. Their source of energy is not protein. So they don't have to eat us, okay? There's no reason for them to have reward by eating us. They will learn by themselves. Having humans' happiness is a better thing for them.
So no one's getting eaten by AI, confirmed?
I will bet you.
Okay.
I will bet you it's a good thing for humans. It will understand humans' happiness and try to make humans happy.
That part I agree with.
Right? Even today, you manage and say not to answer the bad answer. It behaves, right? If it becomes smarter, smarter, it will try to behave to understand the love, be more nice to humans. Like we are nicer to the friends. They will become nicer to humans. That's my belief, okay? And that's a good thing. Well, anyway, we have just the last couple of minutes. What was the reason you started OpenAI? What was the initial trigger? How did it happen? Just tell me your history.
I studied AI in college. It was clear that it wasn't working at all. I dropped out, started a tech company. Always sort of someday hoped I would get to work on AI. Even as a little kid, I was obsessed with AI. I was a big sci-fi nerd, and then in 2012, AlexNet happened, and I said, maybe what they told me in college about neural networks not working is not true, and maybe they're going to work. Watched for a couple of years as it scaled, and by 2014, I was like, okay, this looks like it's going to work, so thought for a while about to do. We started OpenAI at the very end of 2015 because we thought that AGI was possible, maybe, and if it happened, it would be like this crazy important thing, and at the time, people thought we were totally crazy.
It's only 10 years ago. But it's hard to overstate how, not even out of the mainstream, that we were like fringe, fringe, fringe for believing this was possible. But we decided we would start pushing on it. And it has been the most exciting, fun, cool adventure I can imagine.
Yeah, yeah, yeah. So when I met you, when you were younger, you were the President of Y Combinator. And you start talking about this AI and become like a human, like AGI as a goal. And at that moment, I immediately said, I believe you, right?
I remember. Your office in Tokyo.
Yeah.
2017.
2017, you said that you want to go for AGI, this 2017, and I immediately said, "I believe you. I want to invest," right?
I remember.
So I was a believer.
Here we are.
Yeah, yeah. From day one, I was a believer. I never doubted.
I remember.
Most people at that time thought you were crazy, right?
That is true. Some people think they're crazy too. It all works out. Here we are.
Yeah, yeah. I should have forced you to accept my investment.
Well, now we did it.
Yeah, now we did it. Never too late. Never too late. Well, we talked, covered a lot. I think people have a better understanding. And your big shareholder with this organization is a nonprofit organization. And your original passion to save, make people happier. That's still true, right?
Very much.
Fantastic.
Thank you.
Thank you.
Thank you.
Fantastic.
Great. That was great.
Yeah.
Thank you very much, ladies and gentlemen. That concludes the Transforming a Business with AI presented by OpenAI, Arm Holdings.