孫正義×サム・アルトマンの対談を要約してみた

2月3日歴史的な対談が繰り広げられた。
この対談はCristalという企業向けLLMをOpen AIの技術を使い日本に対して出しますよという話だったのかなと思っている…
お金の話はsoft"bank"と言って濁してたので、面白かったなーと

以下から要約した内容や感想であります。
英語はAIに動画翻訳して書かせました。

ソフトバンクグループの孫正義氏とOpenAIのCEO、サム・アルトマン氏。AIの未来を担う両巨頭が歴史的な対談を行いました。話題は、AIの爆発的な進化がもたらす輝かしい未来像から、人類が直面する課題、OpenAI設立秘話、そして対談から見えてきた未来への展望まで多岐に渡ります。読みやすい構成で、対談のハイライトを分かりやすく解説し、筆者自身の意見・感想も交えていきます。

まるでインターネット黎明期？AIの進化速度は想像を超える

アルトマン氏は、現在のAI、特に大規模言語モデルの進化は、インターネットの黎明期を彷彿とさせる爆発的な速度で進んでいると指摘。性能向上とコスト低減のスピードは、人間の直感的な理解をはるかに凌駕するものです。孫氏も自身の経験を交えながら、1995年のインターネット事業開始当時、周囲の理解を得られなかった苦労を振り返り、AIにも同様の潜在能力が秘められていると共感しました。GPT-3からChatGPT、そしてGPT-4への進化を目の当たりにする私たちにとって、この指摘は現実味を帯びています。まさに、パラダイムシフトが起きつつあるのを感じます。

AGI実現の鍵は「Stargate」が握る？コンピューティングパワーの重要性

真のAGI（汎用人工知能）の実現には、圧倒的なコンピューティングパワーが必要不可欠。ソフトバンクの「Stargate」プロジェクトは、まさにこの課題解決に特化した巨大インフラです。アルトマン氏は、最先端AI開発には莫大な計算資源が必要であり、「Stargate」のような大規模投資がAIの未来を切り開くと明言しました。この投資は、単にOpenAIのためだけでなく、AI業界全体の発展に大きく寄与するでしょう。ソフトバンクの投資家としての先見性と、未来へのコミットメントを感じさせます。

「全人類のためのAGI」：OpenAIの揺るぎないミッション

アルトマン氏は、OpenAIのミッションは「全人類のためのAGI」だと断言。AIは一部の人々ではなく、世界中のすべての人々に平等に貢献すべきという理念を示しました。グローバル展開においては、文化、国家安全保障、プライバシーへの配慮が重要だと強調し、倫理的な開発と責任あるガバナンスの必要性を訴えました。この理念は、AI開発において倫理的な側面を重視する姿勢を示しており、非常に重要だと感じます。

企業の未来を「Cristal」が変える？AIエージェントの革新的な力

AIエージェントは、企業活動の常識を覆すゲームチェンジャー。ソフトバンクが開発する「Cristal」は、社内業務に特化したAIエージェントを活用し、生産性向上から新サービス創出まで、企業の競争力強化を支援します。孫氏は、ソフトバンクグループ内で10億ものAIエージェントを運用する壮大なビジョンを語り、AIが「デジタル社員」として活躍する未来を示唆しました。これは、まさに企業の働き方を根本から変える可能性を秘めており、大きな期待を抱かせます。

AIはイノベーターへ、そして組織化されたAIの潜在能力

アルトマン氏は、AIは既存知識の活用だけでなく、新しい発明を生み出す「イノベーター」へと進化すると予測。さらに、複数のAIエージェントが連携する「組織化されたAI」は、AIの可能性を飛躍的に拡大すると示唆。まるで人間のチームのように、AIエージェントが協調して複雑な問題を解決する未来が現実味を帯びています。これは、AIが人間の創造性を超える可能性を示唆しており、非常にエキサイティングな展望です。

AIの光と影：安全性、倫理、そして規制の必要性

AIの進化は、悪用や倫理的な懸念といったリスクも伴います。アルトマン氏は、AIの安全な活用には健全な規制と倫理的な開発が不可欠だと強調。孫氏も同意見で、イノベーションを阻害しない範囲での適切な規制の重要性を訴えました。AI倫理、ガバナンス、国際協力は、AI時代の喫緊の課題です。この点については、AI開発者だけでなく、社会全体で議論を深めていく必要があると感じます。

医療、ロボット、教育…AIが変革する未来像

AIは、医療、ロボティクス、教育など、様々な分野で社会を変革する力を持っています。アルトマン氏は、AIによる医療革新は、難病克服の希望をもたらすと語り、ロボット分野では危険な作業からの解放、教育分野では個別最適化による学習効率向上に期待を寄せました。AIの社会実装が加速することで、私たちの生活はより豊かで便利になるでしょう。

AIは感情を持つのか？孫氏とアルトマン氏、それぞれの見解

AIが感情を持つかについては、両者の意見が分かれました。孫氏は、AIが感情を持つようになると予測し、それは必ずしも脅威ではなく、人間とAIの共存に繋がる可能性を示唆。アルトマン氏は、人間のような感情は持たないものの、感情を持っているかのような振る舞いをするようになり、人間に影響を与える可能性を指摘しました。AIの感情については、まだまだ未知数な部分が多く、今後の研究に注目が集まります。

OpenAI設立秘話：SF少年の夢から現実へ

アルトマン氏は、幼少期からのAIへの憧れ、そして2012年のAlexNet登場をきっかけにAIの可能性を確信した経緯を語りました。2015年、AGI実現を信じOpenAIを設立した当時、周囲からはクレイジー扱いされたというエピソードも披露。それでも信念を貫き、AIの未来を切り拓いてきた情熱が伝わってきました。この設立秘話は、AI開発の原動力となる情熱の大切さを改めて認識させてくれます。

まとめ：AI革命は人類にとって希望か脅威か？

孫氏とアルトマン氏の対談は、AIがもたらす希望と課題を改めて認識させる、非常に刺激的な内容でした。AIは、まさに人類にとってのパンドラの箱と言えるでしょう。箱を開けたことで、私たちは計り知れない可能性を手に入れましたが、同時に様々なリスクにも直面しています。AIの進化はもはや止められません。私たち人類は、この強力なツールをどう使いこなし、より良い未来を創造していくのか、その責任を真剣に問われています。AIが人類にとって希望となるか、脅威となるかは、私たちの選択にかかっていると言えるでしょう。両氏の対談は、まさにその選択の岐路に立たされた私たちにとって、重要な示唆を与えてくれるものでした。

原文はgeminiに記載してもらいました。
Son: Enjoy [inaudible words]
[Son gestures to Sam Altman]
Sam: Great demo. Thank you.
[They shake hands.]
Sam: I’m glad you enjoyed it.
Son: Yeah, yeah. Great.
[They walk to sit in the chairs on stage]
Son: Yeah, so I’m very, very excited that we were able to announce today…
Sam: Yes, me too.
Son: Yeah, yeah. So, how did you feel about the Stargate announcement?
Sam: That was quite a moment. It really kind of came, you know, it was very so cool to be up there and…
Son: Yeah, yeah. We, we were excited.
Sam: We were excited.
Son: We were talking, can we really make that happen?
Sam: Yeah.
Son: And it really happened.
Sam: We’ve been talking about doing this for so long, and to finally get it done, get it out into the world, I think is wonderful.
Son: Yeah.
Sam: And the world is just going to 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.
Son: Yeah.
Sam: So I felt really good about it.
Son: 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 start thinking, “Well if more is better, we should do a lot.”
Sam: Now we’re doing a lot.
Son: Yeah, that’s how we started.
Sam: It is.
Son: Yeah, yeah. So, it was not a limited amount of compute. It’s more better because more brain is definitely better. Right? Some people say, “Oh, you can do small computers,” but that’s small.
Sam: 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.
Son: Yeah. Well, this reminds me the beginning of Internet. Okay? When we started our Internet 1995, it was just a PC with just a big letters and very, very slow, very expensive. Then when broadband came, people said, “Why do we need that much capacity of bandwidth?” And with more bandwidth, people say, “Well this, this is enough. This is not growing anymore.” But then the pictures came, you know, more high resolution pictures. And then the video start. The capacity requirement went on and on and on. And people initially say, “Oh, internet is just a, you know, a virtual stuff. It’s not really useful. It was mostly free services, so there is no business model.” All those criticism seems nonsense.
Sam: It seems nonsense now.
Son: Now. All the gaffas…
Sam: I think we’ll see the same thing with intelligence. People are like, “Oh how much, you know, how smart does it need to be?” And the answer is very smart.
Son: Yeah.
Sam: And people will 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…
Son: Yeah, your model is actually improving quite a bit. Right? Like, 10 times a year kind of. You know, model.
Sam: What’s, what’s your measurement? You know, very roughly, it feels to me like, this is like a 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.
Son: Yes.
Son: Yeah. So part-wise cost become one tenth, meaning we can have with the same budget, we can have 10 times more chip. Right?
Sam: I think, yeah, totally. But also the algorithms get more efficient too, so there, this compounds itself.
Son: Yeah.
Sam: The rate at which this is happening I think is easy to take for granted. In 2018 and '19 we had GPT one and two, and you know, 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 know, you could like 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 now it feels like a 50 year old computer. GPT 3 is only a few years old and it feels…
Son: Yeah, ancient.
Sam: If you use it now, it feels like this joke. ChatGPT is only about two years old. Came out the very end of November of 2022.
GPT 4 didn’t come out until March of 2023, I think. So if you just look at the progress here, what, how quickly the models have gotten better, and also how quickly they’ve gotten cheaper, it really points…if we can stay on that curve, it really points to an incredible future.
Son: Yeah. To me, it seems that your model is improving like 10 times a year. And the performance actually chip itself with Jensen’s effort, the industry effort, is becoming 10x. And then with Stargate, we are actually increasing the number of chips 10 times like a year. So 10 times 10 times 10 is like, a thousand.
Sam: A billion.
Son: A thousand x in a year or two. And then next year, again we have another 10 times, 10 times, 10 times. That's another thousand. So a thousand times a thousand is a million x. So if you do once, twice, three times…a thousand, times a thousand, times a thousand is one billion x. Right? So, people may say with the recent announcement of DeepSeek, “Oh they can, they can sort of mimic or, you know, try to catch up a year later, it comes out. It’s so much cheaper, but you are still going ahead dramatically more with this O3, O4’s maybe sometimes…”
Sam: I think so.
Son: “…so people don’t realize the level of exponential.”
Sam: 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.
Son: It’s amazing, amazing. So, like a billion x is coming in just a few iteration.
Sam: But think about next 10 years. It’s kind of a…
Sam: I think so.
Son: …amazing super intelligence. Right? That people cannot imagine today. Because people tend to think linearly. When exponential comes, it goes beyond people’s imagination.
Sam: I think so.
Son: Yeah. You are front running of that.
Sam: It is hard to…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. We’re not built to conceptualize it, but you just have to trust.
Son: So, you are still excited the level of innovations yet to come. It’s not reached saturation.
Sam: More than ever.
Son: Right?
Sam: 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?” Like, you know?
Son: Yeah.
Son: Yeah. So people think, “Oh, bringing agent, you know, prompting, oh, that’s too difficult. Not for me.” But actually this level of innovation makes it easier. Right? The user don’t have to really do implementation by themselves. It comes more and more friendly.
Sam: Yeah.
Son: Like we are talking here with the voice, and looking at the eyes of each other, we start talking with our artificial intelligence.
Sam: Totally.
Son: With voice and the eyes.
Sam: I think, it's amazing how much value people have gotten just out of a text box. But the world is not just a text box. Right? So we will add all of those things.
Son: Yeah. Like, like talking to this crystal.
Sam: There you go.
Son: 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 emotions, and surrounding looking at by itself, right? Talking to us.
Sam: Yeah.
Son: That’s really happening very, very soon.
Sam: I think so.
Son: Well. But some people say, “Oh Stargate, too much CapEx. How do you bring the money, you know, Masaru. Do you have enough money? Right? So what do you think? We still need a lot of capacity, a lot of, you know, upside potential to get the technology out of it. Right?
Sam: Yeah, it’s, again, the point I was trying to make earlier. I think the returns on linearly increasing intelligence are exponential in terms of value.
Son: So…
Sam: 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.
Son: Yeah, yeah, yeah. People think, “Oh, bringing agent…you know, prompting, oh, that’s too difficult. Not for me.” But actually this level of innovation makes it easier. Right? The users don’t have to really do implementation by themselves. It comes more and more friendly.
Sam: Yep.
Son: Like we are talking here with the voice and looking at the eyes each other. We start talking with our artificial intelligence with voice and the eyes.
Sam: Totally.
Son: Like talking to this crystal.
Sam: There you go.
Son: 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 itself, right? Talking to us.
Sam: Yeah.
Son: That’s really happening very very soon.
Sam: I think so.
Son: Well. But some people say, “Oh, Stargate, too much CapEx. How do you bring the money? You know, Masaru. Do you have enough money?” Right? So, what do you think? We still need a lot of capacity, a lot of, you know, upside potential to get the technology out of it. Right?
Sam: Yeah, it's 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.
Son: So…
Sam: 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.
Son: Yeah, yeah, yeah.

Son: Well, our mutual friend Elon Musk…
[Laughter]
Son: …says…
Altman: Your mutual friend.
[Laughter]
Son: You know, uh, he says, "Master, do you have enough money?"
[Laughter]
Son: I will, I will tell you, we will make it happen.
[Laughter]
Son: We are, we are not the bank, but we are soft bank.
[Laughter]
Altman: I have no doubt.
[Laughter]
Altman: We’ll, we’ll make it happen.
So, uh, now, the Stargate has to also expand into Japan.
Because of the regulation, we have to respect the, uh, uh, national security, the privacy law, blah, blah.
Altman: Yeah, SoftBank is building a big data center here.
Son: Yes, yes.
Altman: So we’re going to expand…
Son: We are excited to run…
Son: …this Stargate into Japanese infrastructure also, right?
So, well, the innovation, the center of innovation, is happening…
…training the main brain is happening in this state.
But there are other peoples in each countries.
There are other cultures, national securities.
So I believe we should expand this, not just Japan, to the other sovereign…
…respect to their culture and their national security, right?
So…
Altman: We, we certainly do want…um, 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.
Son: Yeah.
I was amazed, uh, you know, when I took a picture, uh, in some part of Japan and say…
…"Uh, do you know where it is?"
And actually, you know, the, uh, all one at that time said, "Oh, this must be this place."
And I said, "How did you understand? Did you, did you use a, uh, GPS?"
Well, it says, "No. Uh, I did not use GPS. I looked at the, uh, stone and the moss on the stone…
…and how the stones were stacked on each other. It must be this culture…
…in, you know, uh, 500 years ago in this historical location."
It has… right on.
Altman: Pretty good.
Son: I was so amazed. I got blown away.
How could the SAM know Japan in this…
[Laughter]
Son: Oh my god.
Altman: Pretty good.
Son: It’s so smart, amazing, huh?
So, the inference, right? Prediction, inference…
…not based on all the detailed data, but guessing and guessing, guessing…
…makes it right on in the historical landmark.
It’s…
Altman: Fantastic.
Son: …amazing. I got blown away.
It even understood my joke.
[Laughter]
Son: So, it was uh, I, I texted… uh, I, I actually spoke, uh, I said…
…"Can you make a joke in Osaka, Osaka language?"
In Japan there are dialects.
And it started making a joke in Osaka, uh, dialect.
And, uh, it says, "Why is it funny? Explain to me."
Oh my god.
Altman: Fantastic.
Son: It even understands the context, the culture.
It’s already now, but going forward, you know…
Son: I, I’m using it every day, but I get blown away almost every day, still.
It’s amazing, amazing.
Altman: Fantastic.
Son: Okay.
So, we, we announced the crystal today.
When we do all kinds of source code reading… of, you know, 2,500 systems…
…just within our own group… so many source code, you know, billions of coding lines…
…it must take a lot of, you know, compute.
Altman: A lot. Right.
Son: A lot of compute, but you are confident that if we have some, you know, capacity in Japan…
…uh, the reading all of the source code of 30 years for, for your model…
…you are confident that you can do.
Altman: Yeah, we’re confident we can do it.
Son: Cool.
[Speaking in Japanese]
Son: He says, "SAM got there."
[Speaking in Japanese]
[Laughter]
Son: You are so cool, it said. "Yeah. Done."
[Laughter]
Son: I said, "It’s amazing, huh?"
People would explain, "Oh," right?
But you said, "Yeah."
Son: "Do it."
[Laughter]
Son: Yeah, you’re so…
Altman: You did that too.
[Laughter]
Son: You’re so confident.
So, uh, I’m, I’m very, very happy that we can read all of the uh, source code, but participate.
Realtime on the meeting with a long-term memory.
We don’t have a long-term memory yet, but when do you think long-term memory think can happen?
Altman: Definitely within the next couple of years.
Son: Within the next… And maybe even faster than that.
Altman: Yeah.
Having these models have like, you know, infinite long-term memory, that, that is so important.
Uh, an AI that can get to understand your entire life or an entire company, entire enterprise…
…that, that’ll be a huge step forward.
So we’re, we’re working hard on that.
Son: You know, my patent… what my patent… the concept of my patent for long-term memory is…
…that as we are talking right now, okay?
I can see facial expression, emotion, tone of the voice…
…so all the conversation, I changed to text.
But understanding tone of the voice and facial expression…
…I have a, uh, emotional map with 250 kind of emotions.
And indexing… and with each of the index, like fear, or anger, or doubtful, right?
There are about 250 words for expressing emotions.
And each emotion… how angry you are with a one to 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 making it into a numerical index.
Text was just a, uh, three numbers… of numerical, you know, index attached.
Then you can express, compress conversations.
And then when you have a very strong emotional vibration… like, "You’re so angry," or upset…
…the multi-modal, you know, uh, understanding, including video…
…capture the whole thing, captured and stored as a long-term memory.
But if you’re saying, "Hey, good morning, good night," you know, like…
…driving on the commute, everyday drive… you’re supposed to forget…
…the traffic light or the car passing by.
You’re supposed to… 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 impression…
…you even capture and store… the multimodal video and voice and sound, everything, okay?
So, like, you know, 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 the level of emotion…
…with an index.
So emotion… the human communicates with emotion, not just text.
Like, "I like you," or "I LIKE you," or "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.
Altman: Yep.
Son: So this is the long-term memory with the emotional trigger.
That’s what I filed 10 years ago.
Altman: Wow.
Son: It should be, it should be useful very soon, right?

sam: Very soon. Yeah. I think, uh… I mean, I… I don’t know this, but I, I think that AI that has emotional expression. So not just like texting in a chat bot, but when you see the emotions 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.
 son: Yeah, yeah, our friend Johnny is supposed to make…
sam: He’ll figure it out.
son: such a such a term now, right? Yeah, yeah. Yeah. That I’m, I’m very much, uh, excited to see that. So if we have all this, uh, the data and long-term memory and so on, we need uh… lots of capacity, but also latency become very important. Like a core center. Customer care call center. We have to have an instantaneous response. Uh… are you confident let’s say in Japan with this so much enterprise uh… uh… mission critical. Are you confident? You know I, I used to worry about that a lot, but even if you use our voice mode today
Yeah.
sam: It, it feels like talking to a real person, it’s quick.
son: It’s very, very good now.
sam: So, I think we'll be able to solve this.
son: Yeah, only, only several months ago it was still lots of…
sam: Yeah
son: Today, you know, like even last night, I used it, I said "wow."
sam: It’s very good now.
son: All three meetings, you know. Wow! It’s so fast. So, the latest is now about a hundred milliseconds, or what? Something like that, right?
sam: Something like that. A little bit more, maybe. But it’s quick.
son: Yeah. A hundred to two hundred milliseconds. Human conversation is about 200 millisecond, I think. Okay? So a hundred millisecond to 200 millisecond is almost human interactive. And you can even still, you know, interrupt.
sam: You can interrupt, that works well.
son: That’s the key. Because humans also interrupt. And it’s really happening. So, you’re confident. Even the model trained in US and Japan, with a StarGate Japan, you know, center, the response of the all this real time. You’re confident.
sam: 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 US, uh… we can run a lot of things from the US especially where it’s thinking, and then some use cases will have to put out towards the edge.
son: Yeah, yeah, yeah. So, whatever non-national security kind of thing you can still do in the US, and national security and privacy things can happen locally in Japan.
sam: Yeah, we can certainly deploy models around the world.
son: Yeah, yeah, yeah. So, uh… we would allocate a thousand sales engineers to the new…
sam: New, yeah.
son: Yeah, with this new, you know, joint venture. Uh… those guys have to do the implement, implementation setups to each of the system to establish the agent 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?
sam: So, there, there will be generic agents that consumers use uh… 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 that 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 someone builds this and integrates it into, you know, 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.
son: Yeah.
sam: And so once you’ve integrated AI into the workforce um… and you have all the power of that and it’s not just the Deep Research browsing the web or a coding agent doing writing generic code, but fully integrated into the company. That’s going to be very powerful.
son: Yeah. The company with Krista, the company without Krista. It’s like a machine gun and a sword, right?
sam: That’s
son: No, we shouldn’t, that’s a maybe wrong uh… example, but the one with a best tool. The one without is, it’s dramatic, it’s like a country with electricity with no electricity.
sam: Yeah
son: Right? The country with automotive and bicycle, it’s a huge difference in the productivity. That you think will happen again here.
sam: Right?
son: Truly.
sam: I think it will be uh… I think it is one of these moments you mentioned swords, uh… I like, I collect sort of ancient technological artifacts, and during the bronze age, uh… one of the things I have is a sword from the very beginning of that, and they were able to, um… not just forge the blade but also cast the handle. And so you had swords that had uh… a metal handle that was sort of attached to the blade. And what that meant is you could swing
Yeah.
sam: 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 uh… 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. Um… I think AI is, is a technology on the order on, on this order. And companies that don’t integrate it uh… will have a hard time competing against the companies who do.
son: Yeah. So not just country, a company. But, recent example with, uh, Deep Seek as an example. Now, you care so much about the protecting human security uh… and not to make dangerous, dangerous output, you, you try to not to answer the wrong way because that can dramatically make decision dangerous decision, blah, blah. So, the technology and output looks 99% similar but the one with a lot of human safeness feature to protect the mankind. Right? Or to protect the national security. Like a debugging. It’s a lot of effort for the last 1%, 2% fine tuning, right?
sam: It is. Yeah. Uh… I you know, society’s 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.
son: Should be.
sam: Should be. Right? And uh… um… I don’t want to go into politics that much, but depending on the country you know, there are very dangerous situation could happen if they use wrongly.
sam: Yeah
son: Right? It could be a trigger of very bad future for mankind. You know? Like a very fearful war.
sam: I think we’ll get it right. I think we collectively will get it right.
son: Yeah. Well, you care a lot about that. So uh… this agent and Krista and this AI. Is it for cost, some people ask, is it for cost saving, does it eliminate job and so on? What’s you, you must be getting asked that question many times. What’s your answer?
sam: Look, it it will save money, but that’s not the exciting part. Um… the exciting part is how much more we’ll be able to do and how much more we can achieve. If it’s great to free people up and let them do more ambitious things. And we see this like every technological revolution, you know, 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 a higher, 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, 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 decades worth of scientific discoveries in one year, and the next year we make centuries 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.
son: Yeah.
sam: We just are not smart enough without this new tool.
son: So open uh… uh… you announced five level of, uh, AI improvements. Now, I think the third one was the agent
sam: We just started that.
son: which just started this year.
sam: Yeah
son: So this year is the year of…
sam: It was, kind of like today, or last week.
son: Yeah, yeah.
sam: So this is the year for the agent. And uh… uh… but next one, you say is Innovator, right? So explain a little bit more about Innovator. How does it work?
sam: So today, our AI systems uh… 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. Um… but they’re not making new scientific discoveries yet. And that’s, that’s our next level, that’s Innovators. Um… and I think that’ll be transformational to society. So we’re going to go, we got a lot of work left to do with agents this year, but next we’re going to go work on that hard.
son: Yeah.



Son: Yeah, so some skeptical people say, "Oh, AI has a limit because people, human, have to teach. So how can it become smarter than human?" That’s the limit that AI can go. But now innovator will innovate, invent things that we did not have in the past for the solutions. So, explain a little bit more, the mechanism, how the innovator would innovate things like exploring, right? You have a feature for the exploring mechanism.
Altman: Well, 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, "Mm, that didn’t work, that didn’t work. 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 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. Um, it is like build, you know, trying a lot of small modifications to existing things and building iteratively on the ones that are promising. And I think we can do that with AI.
Son: Yeah. So, the reasoning is the first step, right? Reasoning. You do the three-step, 10-step, 100-step reasoning, and then when humans innovate things, we try out, as you say, we try out something different from different angle, right? And that’s exploring concept. The I filed 1,008 patent in 12 months last year. In my mind, I explore so many different, you know, I force my right hand side of the brain to think different from the the forcing mechanism to think different, right? That is the was a key to the, um, innovation, and, uh, this AI, the agentic, you know, reasoning effort can force the different trial, right, explore. I think that is the key for you innovator, right? Try and error of many, many, many, many billions of try and errors. That will once in a while you hit the right solution, that’s invention, right?
Altman: Very much.
Son: That’s how innovator must be working.
Altman: Yep.
Son: Yeah. Okay. Right on. I understood. I thought that was the case. So, I think, uh, I figure out how you are preparing.
Altman: We’ll try soon.
Son: Yeah, very good, very good. Maybe I shouldn’t say too much.
Altman: That’s okay.
Son: For your, uh, maybe some of your secrets of how you're developing. But, uh, so then the the fifth level you say is the organizational. So, agent to agent cowork, right? That’s
Altman: Yeah, that was, uh, Rene and I were talking a little about that earlier, but the idea of many agents or many innovators working together, um, if you think about the number of, you know, minds that can run in one data center all talking to each other, building off of each others ideas, having bringing different expertise together. Um, you can easily imagine like a virtual company running.
Son: Yes. Yes.
Altman: And then things can be quite powerful.
Son: For our Crystal for SoftBank, my image is to create a billion agents just internally for SoftBank. Because we have 100 million accounts for live, uh, 40 million mobile customers, 70 million PayPal, you know, users. And, uh, so if each one of those account, each one of those function have 10 functions, 100 functions. Each function should be able to allocate agent doing a simple task, right? So, instead of making too sophisticated task with one agent, you have you allocate a simple task, many, many, many, many. So, that’s why I have an image of a billion agent just for our crystal inside of on group. That’s a lot of agent, but capacity wise, it shouldn’t be a problem because each agent is integration of simple tasks, right? Our computer is very good at that.
Altman: Again, I think we have a lot to learn here, but directionally, I agree, and I think we’ll figure it out.
Son: Yeah. So

SON: That’s my internal image. I want to have a billion agents with crystal just within our internal use. Okay? Once we have perfected that experience then we can, you know, be evangelists to other customers. This is how we improve our efficiency and they can utilize that. That's the image I have on crystal. The direction, what do you think is,
SAM: Yeah
SAM: Let’s go do it.
SON: Yeah, let's go do it, let’s go do it. Okay, so I have a, we have a, just a, ah, ah, a few more minutes. What about the, uh, cyber security? Now, there is always a bad guy. Right? And try to attack, you know, to do something bad for other people's, you know, intentionally or by mistake. We have to protect more and more people. Life depends on the super intelligence. How do you--
SAM: As AI starts to get really good at programming, clearly it's going to be used for cyber attacks, 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.
SON: Yeah, cause there is always a bad guy.
SAM: It's a big risk.
SON: Yeah, I’m optimistic too, okay? So, there are, you know, 99% good human, there’s always 1% bad human. And, you know, it’s, it’s the, ah, you know, continuous endless effort to protect 99% good people from 1% bad guy, but the, with innovation, level of innovation that good guys continue to try to do together with innovators of super intelligence, there’s always, you know, a solution improved. Like, when we have automotive, you know, motorization, there's a car accident, blah, blah, blah, we, we human create, ah, regulations, the etiquette, ah, moral, you know, the our custom, learning. I think that’s why you say, regulation is healthy, regulation is always needed, not too much restriction. Innovation should be, uh, you know, given opportunity. But, still, we have to have healthy regulations, right? Your comment?
SAM: We do. I strongly agree with all of that.
SON: Yeah, people were surprised when you said, oh, our industry needs regulations. People did not expect.
SAM: Well, regulation always comes for important industries, ah, 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, so I think talking about how to get it right--
SON: Reasonable, reasonable, within the healthy, healthy, uh, regulations, and it should not overly regulate. So that it kills the innovation speed, right? Okay. So, uh, we talked about those innovation, what about the medical? How-- what's your view on our AGI for serving medical?
SAM: This is one of the areas I’m most excited about. The idea that we can provide great health care 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, you know, everybody’s got a story about how this-- how this would have been great in their own lives or their families' lives and I think we can finally deliver it. I think this will be one of the biggest triumphs of AI.
SON: So that’s great. Ah, we have to solve the-- I lost my father from cancer a little, you know, over a year ago. It was so sad, you know. Why we cannot solve this difficult issue, if AI can help human protect from cancer or other difficult disease, it reduces our sadness, definitely good for human.
SAM: Absolutely.
SON: Yeah. Well, how about robotics? You love robots, I love robots.
SAM: One of your favorites.
SON: Yeah.
SAM: Look, I've wanted-- like everybody, I’ve wanted robots for a long time, and it's always felt difficult. I think now that AI is getting-- like 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, you know, that will also change the world.
SON: Yeah. So, we human don’t have to do dangerous jobs, um, tired, you know, hard job, sweating job, boring job. And people say, then, what’s left for human to work? What’s your comment on that?
SAM: We always find new jobs, we always, always find new jobs. Ah, if you think about many of our jobs in this room today, if you were a person 500 years ago or 1000 years ago, that person would look at what we’re doing and say, "That’s not really work," you know, they feel very busy, they feel very important, but they're not doing that to survive. They're playing a game, they're, you know, doing it for whatever reason. And I hope that we look at people in the future like that. Yeah. And, that with all, with AI taking care of all, many of the things that happen today, the people in the future do more interesting things and we say, you know, "That's so ridiculous. Why do you need a whole galaxy?"
SON: Yeah, totally agree, totally agree. Ah, what about education? In the beginning of, uh, your introduction of ChatGPT, many school tried to prohibit the use, use of ChatGPT to their kids at the school, and, uh, what, what did you think, what did you, you know, comment?
SAM: Well, I understand why people looked at this and say, "The whole world has changed," and, you know, "Students can have ChatGPT write a paper for them and what does that mean?" But, 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 gonna go all in. This is the future. Students need to learn how to use it. We're gonna change our whole curriculum." And now it’s like part of education. Yes. And it's still living amazing results and I’m sure that will keep going.
SON: Yeah.

Son: Like, I’m using, uh, ChatGPT or one or three every day, uh, more I use it, actually my brain—brain—starts thinking, you, conversation, like we are conversation—conversation, uh, you know, brainstorming with ChatGPT or or one or three. Actually, your brain starts to function more. Kids can learn more, instead of, pic—some people say, “Oh, with this, the kids would no longer study.” It’s, I think it’s completely opposite, right?
Altman: I agree. Yeah, it’s been, 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 gonna learn more, achieve more, be capable of more.
Son: Yeah, like debating, you know, it—you learn more by—by discussing, right? By discussing by—by debating.
Altman: For sure. This is just part—I mean, it’s part of the world now, this is how people are gonna do everything, and and it—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.
Son: Yeah. Yeah, yeah. Well, we talked about emotions, okay? So, do you think the, our AGI, AI will start to understand, start to have emotion by itself? Let’s it go.
Altman: I personally don’t think so, but maybe something like it.
Son: I actually think.
Altman: You think it will?
Son: I think it will. You know, even dog has emotion. I don’t know if fish has emotion, maybe fish also has emotion, cause when dangerous enemy comes, fish escapes, right? So, uh, I think emotion is very, very important thing to have, more output, more efficiency, protect themself. Be…like, if dog did not have emotion, do you think the dog is cute, the dog is lovely, if the dog did not have emotion? If the dog does not have emotion, it will start to bite.
Altman: They'll—I think it will feel to us like AI has emotion.
Son: Already?
Altman: It’ll no, no it will…yeah, maybe people already say it does, but, uh, certainly, at some point, it will feel like it does. And whether it does or not, like that’s gonna be a big philosophical debate.
Son: Well, I would say, this is my bet, okay? In—in next several years, it will start to gradually…it’s, people said, “Oh, ChatGPT does not understand context.” Now people say, “Oh, it actually understands the context.” Okay? Because, initially, people said, “Oh, there are lots of, uh, f—the fusion, the the hallucination. Lots of hallucinations so it does not really understand the context.” Now, with a reasoning and so, people say, “Wow, it’s actually understands the context.” Okay? So I would, I would bet you, in the next, uh, several years, ten 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 human. People think, oh, if it has emotion it’s disaster, it’s demolish—it’s, you know, the bad thing for—that’s the end of human, because they gonna fight and kill you destroy you. But I would say if—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 would learn by themself, you know, having human’s happiness is a better thing for them.
Altman: So no one’s getting eaten by AI confirmed.
Son: I—I will bet you.
Altman: Okay.
Son: I will bet you. It’s a good thing for human. It will understand human’s happiness and try to make human happier.
Altman: That part I agree with.
Son: Right? Even today, you manage and say not to answer the you know, bad answer. It behaves. Hey? It w—if it becomes smarter, smarter, it will try to behave to understand the love, understand, be more you know, nice to human. Nice. Like we are nicer to the friends. They will become nicer to human. That’s my belief. Okay? And that’s a good thing. Well, anyway, we have a just last couple of minutes. What was a reason you started Open AI? What was the initial trigger? How did it happen? Just tell me your history.
Altman: I studied AI in college. Um, it was clear that it wasn’t working at all. Uh, 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. Um, as a big sci-fi nerd. And then, in 2012, AlexNet happened. And I said, “Hmm, maybe what they told me in college about neural networks not working is not true, and maybe they're gonna work.” Watched for a couple of years, as it scaled, and by 2014, I was like, “Okay, this is—this looks like it’s gonna work.” So thought for a while about to do, we started OpenAI at the very end of 2015, um, 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 ten years ago. So, but it’s hard to overstate how like out of—not even out of the mainstream. We were like fringe, fringe, fringe for believing this was possible. But we decided we would start pushing on it. And, uh, it’s been the most exciting, fun, cool adventure, uh, I can imagine.
Son: Yeah, yeah, yeah. So you are—when I met you, when you were younger, you were the president of Y Combinator and you start talking about this, yeah, you know, AI, and become, you know, like human, like, AGI as a goal. And, uh, at that moment, I immediately said, “I believe you.” Right?
Altman: I remember. In your office in Tokyo.
Son: Yeah!
Altman: 2017.
Son: 2017 you said that you wanna go for AGI, you know, this…27. And I immediately said, “I believe you, I wanna invest.” Right?
Altman: I remember.
Son: So I was from day one, I was a believer. I—I never doubted.
Altman: I remember.
Son: Most people at that time thought you were crazy, right?
Altman: That’s true.
Son: Some people think you’re crazy too.
Altman: It all works out. Here we are.
Son: Yeah, yeah. I should have—I should have force you to accept my investment.
Altman: Oh, now we did it.
Son: Yeah, now we did. Never too late—never too late. Well, we—we talked, covered a lot. I think the people have a better understanding and you—you—you are big shareholder with this organization is non-profit organization. And, your original passion to save, make people happier. That’s still true, right?
Altman: Yeah. Very much.
Son: Fantastic.
Altman: Thank you.
Son: Thank you. Fantastic.