115. 对OpenAI姚顺雨3小时访谈：6年Agent研究、人与系统、吞噬的边界、既单极又多元的世界

Why do you think you're doing this earlier than most people? I think the lucky part is that the first thing I did as a PhD student was to do Agent in the language model. You can only surpass your previous boss by having a different bat. I think if OpenAI keeps doing the enhanced

learning, it may be difficult to surpass DeepMind. My mentor is the second author of GPT-1. He was a little suspicious of this at the time. What's

interesting is that Traditionally, people think that when something happens, for example, a big company makes something first, and then the startup company can start to copy it. For example, I made a chat GPT, then I can copy the chat GPT or do something similar. But now it seems that the opposite can

also be established. If you become the CEO of Brokershare, you will have to allocate $5 billion to the AGI industry in the future. How do you allocate this money to both pay back and contribute to humanity? There are different types of super apps in the back-end. There are different ways of interacting. If

you don't believe in this, the world will become very dark. Only OpenAI or Runtropic have the opportunity. But if you believe in this, there will be many new opportunities.

Hello everyone, welcome to Zhang Xiaojun's Business Conversation. I'm Xiaojun. This is a deep conversation program produced by the Language and World Workshop. We hope to explore the new world with you from here. Today's guest, we are very happy to invite OpenAI researcher Yao Shunyu. In April 2025, Yao Shunyu released a very famous book, The

Second Half, announcing that the game of AI main line has entered the second half.

After that, we had a blog talk with him. Yao Shengyu graduated from Tsinghua University and Princeton University and began his research on the intelligence system very early. During

his PhD, he realized that language may be the closest tool to the nature of human invention, so he turned to the research on the language intelligence system. It has been six years since then, and he has done many representative

system. It has been six years since then, and he has done many representative works. Our conversation starts with individuals, and is explored by people, organizations, AI, the interaction

works. Our conversation starts with individuals, and is explored by people, organizations, AI, the interaction between humans and machines, and the world's intelligence boundaries and the landscape of humans and machines.

Hello, I'm Yao Shunyu. I'm

Shunyu. I'm currently working on

OpenAI research.

Today, our guest is Open-Ed researcher Yao Shunyu. His research direction is Agent. Recently, he

just wrote a very famous book, The Second Half, telling everyone that AI games have entered the second half. This is the first time we have tried to have two hosts in this program, except me, and everyone else is also familiar with Guangming. Guangming, you also come to say hello to everyone. Hello everyone, I am Guangming.

Guangming. Guangming, you also come to say hello to everyone. Hello everyone, I am Guangming.

- I am a very obedient student. I feel that I have been studying

in a secret class. I took my undergraduate from Hefei to Tsinghua and then studied in Yao class. In Yao class, everyone would tell you to go to the US to study for a PhD. So I went to the US to study for a PhD. Then I studied for a PhD in the University of Peking. After studying for a PhD, I felt very natural. OpenAI is the

Peking. After studying for a PhD, I felt very natural. OpenAI is the best place to do research, so I joined OpenAI. I feel like my 28 years of life has been very good. You were in Qinghua Yaoban from 2015 to 2019, and you graduated in Princeton in 2014. You

didn't study AI in your major, how did you enter the AI field and then into the Asian field? The tradition of Yaoban is more theoretical computer science. But maybe I still had a

computer science. But maybe I still had a rebellious spirit. I thought the important

rebellious spirit. I thought the important problem might have been solved. If you take a complicated algorithm from n to n to 2.824, this thing has no real meaning. In 2016, I saw a problem in a

class of Li Jian. multi-model work to get back demo. For example, work to get back has a very interesting example. A

back demo. For example, work to get back has a very interesting example. A

king's embedding minus a man's embedding plus a queen's embedding equals a woman's embedding. I thought it was very amazing. And then this can be made into a picture. For example, a king's picture embedding minus a man's embedding plus a queen's embedding. I was very surprised by

his work. But at that time, there was no deep learning teacher or

his work. But at that time, there was no deep learning teacher or resource in Tsinghua or Yaoban. In 2018, Yaoban had a tradition that everyone had to do research in a foreign field. I went to MIT and worked with Wu Jiajing. That's where I really started.

系统性的做deep learning,对。

然后当时我做的其实更多是computer vision, 但是我当时觉得好像vision 你很难实现一个general的AI, 然后intuition就是说感觉language 是更重要或者更central的一个东西, 然后后来进了PhD之后就开始做language,对。 那怎么进入agent呀?

然后后来进了PhD之后就开始做language,对。 那怎么进入agent呀?

对,这个事情其实, I think it's a coincidence. My

mentor did some research before. He asked me how I could become an agent in a simple language game. This was

probably done in 2016 or 2017. You use a very simple RNN and then in a very small text game you can do some dynamic interaction. For example, you learn to know, for example, after crossing the bridge, you can

interaction. For example, you learn to know, for example, after crossing the bridge, you can go to the other side of the river. It's similar to such a very simple thing. After I entered Grass School, I was actually taken over by Computer

thing. After I entered Grass School, I was actually taken over by Computer Vision. But I didn't want to do Computer Vision at that time. Then I

Vision. But I didn't want to do Computer Vision at that time. Then I

went to find people who do language chat. Then I met my current mentor, Karthik.

What do you think is the most attractive thing about Agent or

Language?

I think it's the generality. You can express anything in a language. Or

you can express a lot of things in a language. I think it's very appealing.

I think there's an intuition in music. For example, you want to achieve an AGI. Of course, no one mentioned AGI at that time. But if you want

an AGI. Of course, no one mentioned AGI at that time. But if you want to achieve a very general system, you need to build an agent.

当时我觉得就是说 如果回看AI的历史的话 从很久很久以前 从就是Noel and Simon 他们 In the 1960s, people's initial thoughts were to become agents. At the time, people's ambitions were very ambitious. We wanted to use a summer to solve a mission, and another summer to solve a language. Then we put all these things together and we

went to be agents. Then it became smarter than people. Including the idea of going to see the first picture of Tulin. Everyone naturally wanted to build a person or an agent. But this was too difficult. So I think AI gradually became very fragmented. And the problems we were facing were getting smaller and smaller. Some people were trying to figure out how to solve the problems of

smaller. Some people were trying to figure out how to solve the problems of vision, or the problems of language, or even more detailed, the problems of translation.

And in the end, it became more and more detailed and vertical. But I

think that after 115 years, the birth of the skilling law, and the birth of the research breakthrough, The great thing about history is that we should go back to a more general thinking from this vertical thinking and try to build a more universal system. When you entered the agent system and did research, what

were the most important things that you realized? When you had to activate the language model. There were some gains. I think my biggest gain in the first

language model. There were some gains. I think my biggest gain in the first year was that

Or

基于选择的任务 And I think 95% of people were doing BERT and only 5% were doing GPT. 然后这也是因为当时自然语言处理的主要的任务都是一些 And this is also because the main tasks of natural language

doing GPT. 然后这也是因为当时自然语言处理的主要的任务都是一些 And this is also because the main tasks of natural language processing were some 比如我有一句话,然后这句话是积极的还是不积极的 For example, I have a sentence, is this sentence positive or not? 比如说我很讨厌这个电影,这是一个负面的句子 For example, I hate this movie, this is a negative sentence. 就是做一些非常简单的这种事情 That's why I did something very simple. 在这种事情上BERT确实效果更好 In this case,

sentence. 就是做一些非常简单的这种事情 That's why I did something very simple. 在这种事情上BERT确实效果更好 In this case, BERT really worked better for me. But you will find that if you want to be an agent, you need not only the ability to choose, but also the ability to freely create new actions. Of course, if you are playing a

video game or a video game, your choices are very limited. For example, if you are playing Mario Brothers, it may be left and right. But if you are playing a game in a machine language, your actions are 比如说我在这个游戏里面 我可以用这个剑杀这个怪兽 或者我可以去第三个房间 或者我可以用我的金色的钥匙

打开第一个房间的门 这个事情是Bert永远做不到的 所以我发现这个事情之后 我就再也没有用过Bert 我觉得第二个能力就是说 任务或者环境非常重要 就是当你有一个非常差的任务的时候 你永远不可能学到非常好的东西 从某种程度来说 就当时有很多人在做 Now let's look

at a very simple task, right? For example, is this sentence positive or negative? Or

how do I judge whether the sentence A can lead to the sentence B to be true? At that time, this task seemed very difficult, but now it seems very

be true? At that time, this task seemed very difficult, but now it seems very simple. I think first you have to find a task that is challenging enough, and

simple. I think first you have to find a task that is challenging enough, and then this task can be done in a new way with essence. And then actually, At that time, there was no choice to be an agent or a language agent.

You could only play text games. For example, Zork is a very classic text game. In a world based on text, it's like a interactive script. You can go down, go up, you can go to a room, you can

script. You can go down, go up, you can go to a room, you can do all kinds of things. But you'll find that There are still many shortcomings in this environment. What you can learn is limited to this environment. This environment is not big enough. And if you want

to learn this environment with RL, it will be like learning traditional video games with RL. You can beat this game, but it has no potential for other tasks. You can play very well, but it has no value for other

other tasks. You can play very well, but it has no value for other things in the world. I think We may need a better environment. During your

PhD, you did a lot of work and you are very well-known. Including language

agents, like React, Reflection, and thinking. Including digital automation. Are

these studies broad? What is the commonality between them? How do you follow your interests and do their extension step by step? I think from my perspective, it's a very natural process. When I realized that the environment was in trouble, actually, I think my first important job Wipeshop. I

think we need to solve the environment issue first. If there is no good mission or environment, it's meaningless to make the game as high as possible. In 2015, there was a very good project called World of Bits.

possible. In 2015, there was a very good project called World of Bits.

The idea was to use computers or the Internet as an environment. This

environment is more exciting than games. But due to the limitations of various technologies, this project was not done very well. In 2021, I was discussing with the mentor that it might be a natural thing to do it again.

Of course, I think the technology was not mature at that time. Most people were still studying, for example, can A lead to B, or translation, or can I answer questions from this article. I was still not familiar with the technology, but I started doing it when I was good at it.

Then in 2022, we started doing the environment of webshop. In 2022, the appearance of GPT-3 and Chain

webshop. In 2022, the appearance of GPT-3 and Chain of Thought brought new opportunities. We did React.

I still think that my favorite job is still React. After that, I naturally did more tasks and methods based on these two lines. But I

think my research is on the one hand, how to do something valuable, and on the other hand, how to do something simple and universal based on the tasks and environment of the real world.

Ryaf的提出它有标志一个范式的变化吗?

我觉得这个事情需要可能比如十年后或者五年后再去看 很多时候一个东西刚提出的时候是很难看出来的 当时的学术圈还是不太能接受 就是说我去做一个prompting 然后去把它作为一个research 就是传统意义上你需要去提出一些fancy的 你需要提出一些数学公式

你需要去训练一个模型 你需要去 to prove that you have done a lot of theoretical or engineering work. But if you only use a model, it feels too soft. But I think from a certain point of view, the most valuable

too soft. But I think from a certain point of view, the most valuable thing at the time was to study how to use the model. Because if you want to go to that model, you are actually behind OpenAI or behind these companies.

I think it's too hard. Or

too hard. Or maybe it's not a

It's not a consensus thing. The consensus thing at the time was that I would do the Q&A, or the translation, or some tasks that have been accepted by the community. I think I've always had this non-consensus, that I want to be an agent. Another point is that I've always wanted

to do simple and universal things. I don't want to do something very complicated, but I can only do it in one field.

I

This is a very

good question.

嗯,我觉得这个事情是基于你的context,就是基于你的讨论的背景的,对,就是从历史的角度来说,我觉得从自然语言处理的角度来说,Agent是相对于比如说一个

I can interact with the outside world through the system of creating articles or conversations. For example, using computers, or using the Internet, or using these tools. I think from the perspective of natural language processing, agent is not only able to create new articles or new

thoughts, but also able to interact with the outside world.

But from a bigger AI background, agent is a very ancient concept. You can make your own decisions

ancient concept. You can make your own decisions and interact with the environment, and optimize rewards. All of these systems are agents. From this

rewards. All of these systems are agents. From this

perspective, the meaning of the word agent today is more about how I can use a large language model 能够去做自我决策的这样的agent系统, 而不是传统的, 比如说单纯基于规则, 或者基于在一个领域做强化学习所获得的这样的agent。

因为agent这个词在不同的, There are many different forms of language in the age of AI. You can also say that AlphaGo is an agent. You can

also say that Waymo is an agent. You can also say that Robo is an agent. I think this word is very close to your situation. You mentioned that the

agent. I think this word is very close to your situation. You mentioned that the language agent is different from other traditional agents. What is the difference between the two? Why is the language agent different from the nature of the language? I

two? Why is the language agent different from the nature of the language? I

think the difference between the two is that it can be inferred. Because it can be inferred, it can be used to make a mistake. I'll give you a simple example. I think the reason I made React a strong motivation is that

simple example. I think the reason I made React a strong motivation is that After I finished my first job at Comm, I was thinking about why I could play a new game at once. But now these systems or AI require tens of thousands of steps or tens of thousands of steps or hundreds of thousands of

steps of training to do this. Then I found that it seems that I can think, right? I see a new environment and I will think that this light is black, it may be dangerous.

基于这个常识可能会有怪兽,可能我现在最重要的事情是要点亮灯, 然后基于之前的上下文,灯在我后面,我应该先向后走。

那如果我没有这样的一个思考能力, 我直接从这样一个复杂的语言直接去预测我要去往后走, 这个事情很难,没有推理是做不到的。

所以我觉得最大的区别就是说, - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - I think the earliest AI, which we call "good old-fashioned AI" or "patternism", is actually very simple. I

focus on reasoning. And the way I think about it is, what do I think? I design these rules and let AI do that.

If my temperature is higher than 30 degrees, then the air conditioner should be lower. It's based on such a rule. Then this thing can actually create

lower. It's based on such a rule. Then this thing can actually create a lot of the earliest intelligent bodies. Including the earliest robot, the earliest, for example, proof of mathematical theory, including many other systems, are all created like this. But soon, for example, in 1980, we found that this thing has a

like this. But soon, for example, in 1980, we found that this thing has a bottleneck. No matter how many rules you write,

bottleneck. No matter how many rules you write, 你还是很难涵盖这个世界上所有可能发生的情况。

How do you do

that?

这个心脏病的这样一个系统 那你写了很多很多的规则 但是你还是没有办法去涵盖 所有可能出现的情况 因为人是一个 就是他会说任何事情 对吧 你没有办法去handle 然后你写了这样一个心脏病的系统 你没有办法去处理 比如说肺病 那这个事情就导致了 第一次这个AI的寒冬 对吧 然后我们有neural

network 有了神经网络 然后我觉得第二波 就是Agent的兴起是 deep reinforcement learning. For

example, DeepMind is working on these video games, and AlphaGo. We

also have some open AI mobile games, like Dota. The core of this is that I have an

Dota. The core of this is that I have an endless virtual environment, and I have a reward. I have a very

reward. I have a very 同用的网络架构 然后我就去像黑河一样 去学怎么去把这个reward去improve 然后它就变强 然后这个事情 我觉得取得了很多成功 我觉得可能最有名的事情是AlphaGo

但是我觉得还是有同样的问题 你去做任何一个环境 你需要去做很多 环境specific的工程 You can't do a lot of hyper-parameter tuning or engineering

based on this environment. But the biggest problem is that it can't be modified. You learned a micro-agent, you can't play other games. You can't go from one environment to another. That's not good. And if all the environments

you can solve are virtual, or environments that you can play with your own hands, like a game, you can't find a good real-world application. I

think the third agent is from the big language model. We found that it can do reasoning. And then based on reasoning, you can actually create some new environments, such as coding, Internet, and various digital environments. And then these digital environments have a very big feature, which

environments. And then these digital environments have a very big feature, which is that most of them are based on language and require reasoning. So I

think this time, the main difference between agents is that there are two sides.

On the one hand, we use language models and reasoning to build a lot of 处理各种各样问题的这样agent 但另一方面就是说 agent的环境也发生了一个进化 就是从最早的就是这个符号主义的 就是比如证明数学定理 到下围棋玩游戏 到今天我们去做互联网 去做coding

去做computer 去做这些真实世界的数字环境 所以我觉得是有两条线 大家可能往往会看到方法的这条线 但是会忽视了 I think these two lines are actually similar. I actually have a very basic question. OpenAI has five branches. We

similar. I actually have a very basic question. OpenAI has five branches. We

are all familiar with it. From chat robot level 1 to returner level 2 to agent level 3 and then to innovator and organizer. This is level 4 and level 5. What is the internal logic of these five branches? Why is it that

level 5. What is the internal logic of these five branches? Why is it that there is a chat robot returner and then an agent? Yes, I think...

The logic behind this is that you need to have a language-based knowledge.

Based on language-based knowledge, the earliest application you can make is a dialogue robot. Based on language-based knowledge, you need to be able to interpret.

dialogue robot. Based on language-based knowledge, you need to be able to interpret.

Step 2 is Reasoner. When you have a good language-based knowledge and reasoning ability, you can actually be an agent or an agent that can translate.

And then I think it's obvious that today's agent's most important directions of improvement are: One is to let him have his own rewards, to let him explore himself. and the other is multi-agent, allowing it to form an organization. I

himself. and the other is multi-agent, allowing it to form an organization. I

think these two things are probably in common or can be developed in parallel. I'm

not sure who is level 4 and who is level 5, but I think these two things are very obvious that we need to take the next step. So from

level 2 to level 3, you took this step from the virtual model to the user model. This is actually a very important cross-examination. Or simply from the reasoning point of view, to make the reasoning into an agent to interact with the environment. What are the mainstream structures of agents? Has this formed a

the environment. What are the mainstream structures of agents? Has this formed a consensus? I think, I feel that most of the time, people use a structure

consensus? I think, I feel that most of the time, people use a structure similar to React. That is, you can make a reasoning and then you can create an action. This is the simplest thing. But again, I think the simplest thing is probably the best work. And I think maybe based on a specific task, you will have a lot of workflow or more specific methods. But I

think the most common method is similar to React. Guangming, what are the key abilities that you value the most in improving your agent's capabilities? Someone

mentioned context, or even long-context reasoning, or tool-sweeping, or command-and-control. You were mentioning reasoning just now. If you want to improve your agent's

command-and-control. You were mentioning reasoning just now. If you want to improve your agent's capabilities, which abilities do you value the most? I think this is a good question. I don't think there is a good taxonomy or a system of division. Or a system of division for each person.

For example, one person can divide I think both of these types of separation are reasonable. And

reasonable. And for now, I think what I value the most is to deal with context or

memory capabilities. And then based on it, to do lifelong learning or

memory capabilities. And then based on it, to do lifelong learning or online learning. You've been talking about the environment. Do

online learning. You've been talking about the environment. Do

you think code is the most important environment to achieve AGI? Can you do multiple-round R? The feedback is also guaranteed and can be

multiple-round R? The feedback is also guaranteed and can be verified. Do you think it will be faster to use agent in this

verified. Do you think it will be faster to use agent in this environment? Yes, I think it is undoubtedly one of the most important

environment? Yes, I think it is undoubtedly one of the most important environments. I think coding is a bit like the hand of man. It

environments. I think coding is a bit like the hand of man. It

is the most important affordance for AI to some extent. For the

physical world, the affordance of man I don't know how to translate this word in Chinese, but for humans, the most important thing is to create tools that can be used by hands.

For example, a hammer, a pen, a chopstick. But for AI or digital agents, the most important thing is code. Because other affluence are actually defined by people, for example, your web page or your novel or something else is actually defined by people. Only code is a very natural thing

that is defined by machines. And then I was probably 22 years old, I was very confused at that time, which means that it is obvious that I think the most important thing for coding agents is why no one does it. We did a work called intercode. Everyone was doing it. For

does it. We did a work called intercode. Everyone was doing it. For

example, I have a task, I have a coding task, and I generate a code, and then I evaluate it. But we said, why don't you and then the results are returned to the model. You do this kind of agent task for multiple rounds, and then turn it into an environment, instead of a simple task. And then based on this, we later did the sweep bench,

simple task. And then based on this, we later did the sweep bench, and then sweep agent. But sometimes I think it's very interesting, it's obvious that something is very, very important, but sometimes no one does it. So, for example, if you are a researcher, you think what you do is very important, but no one thinks it's important or is doing it, then it may not be a bad thing.

It's important, but it's not easy to do. There's a strong consensus here. Some people

think that code is the biggest value of this technological revolution. But some people think that it can translate into more tasks. In the entire computer, mobile phone, digital world, it can achieve the tasks that people can do 95% or 99% of the time. Do you think that from code to the entire digital world,

the time. Do you think that from code to the entire digital world, this step of transgression or translation, are you confident? I think

From a broader perspective, you can think of API as part of the code. Any interface based on code is part of the environment. And then I think there is a very classic

environment. And then I think there is a very classic debate. The final AGI is based on API or based on code.

debate. The final AGI is based on API or based on code.

Or based on GUI or based on the environment of human definition Or it's a mix I think this is a bit like Of course, the first point is that many things may not have an API, right? Now it

only has a frontend, it only has a frontend Then you can build an API for it That's a bit like You want to change your car to adapt to all roads Or you want to change your road to adapt to these cars now, right? And then, of course, I think the final

result is "meet in the middle" which means both of them will do it. And

maybe this thing is not that difficult. Now, let's say, an agent can use code, and they can use these screenshots or frontend

Right.

I would like to ask another question about the error. I read your latest article and I was impressed by the fact that you mentioned that it finally got an error. Is it really an error? Because you mentioned that a lot of advanced knowledge has been trained in the model. What is the sign that you can

feel that it is an error, rather than the data contained in the training data?

I think it's possible that if your pre-training includes all the things, RL just activates all these skills. It reminds me of Ilya's saying, "Maybe the ultimate generalization is to overfit the reality." If you can do everything in the world, then it doesn't

the reality." If you can do everything in the world, then it doesn't matter if you overfit or generalize it. But I think Again, it's still generalized. And I think the reason is that it can reason.

still generalized. And I think the reason is that it can reason.

That is to say, when you can learn some skills to think in an environment, and this skill of thinking can lead to a new environment. I think this is the main reason why I say it's a failure. From the previous one, you may have learned more. For example, I play Valkyrie. I have a strong understanding

of this environment or this game. But how do I... I want to add one more small question. We may soon see the most powerful software engineers. Even in 2027, we saw the most powerful software engineers that

can operate almost all tasks and commands on human computers and mobile phones.

What is your fantasy for this day? Is it too optimistic or reasonable?

I think this is not well defined yet. From a certain degree, the ability to write code is already better than most people in the world. Or its mathematical reasoning or logical reasoning

world. Or its mathematical reasoning or logical reasoning ability is already better than most people in the world.

But when you ask whether he can use these environments well, it still depends on what task you want him to do. And then is this task a task that can be reasonably defined? I think many times people or the most difficult problem for humans is not

to reason, but to get this context or to get this I don't know how to translate it. I think the model of bottleneck is not that I lack the ability to push or write code or use the front end. It's that it lacks a complete

context. I don't know if it's a question of intelligence

context. I don't know if it's a question of intelligence or a product or something else. But if you want AI to achieve value, you need to solve this problem.

You were in the second half of the movie in April. How did

you come up with the idea of the second half? What

inspired you? Good question. I was invited to Stanford to give a talk. I thought about what I could say. Obviously,

I couldn't say anything technical. I could only say something philosophical. What are you talking about? And then I

something philosophical. What are you talking about? And then I thought of this. And I think it's a feeling I've had in the past half a year working at OpenAI. People often look at it as a model or a method or something else. But now I think Botanac has been transferred to

how to define good tasks and how to define good environments. Do you think it's the turning point now? From the first half to the second half?

From a certain point of view, I think the main line is from the top to the bottom. I mean the language intelligence. Of course, you can say that audio, multi-model, or robot have many problems that can't be solved. But I think from language to definition of reasoning and agent, I think

solved. But I think from language to definition of reasoning and agent, I think we finally have a very general method on this line. And this method can be used to make a mistake. This brings a fundamental difference. Before, I had a lot of monsters, and I

difference. Before, I had a lot of monsters, and I had to create weapons to fight them. Now I have a common weapon, a machine gun. Now I have to think about where to shoot. I don't have to think about so many ways. I

have a very common way. I may need to think more about I need to use this method to solve a problem. So how to set tasks, how to define problems, what are your

problem. So how to set tasks, how to define problems, what are your thoughts on this in the process of exploring? Yes, I think different people have different flavors. And I have had this kind of preference since a long time ago. I want to define

a reward. This reward is

a reward. This reward is 基于结果而不是过程的,而且它是一个基于规则或者说能够很清晰的算出来,而不是基于人的偏好或者模型的偏好或者一些非常黑核的东西的。

Then when we were doing the webshop, the most difficult part was how to define rewards. In fact, I think the hardest part of any RL task is

define rewards. In fact, I think the hardest part of any RL task is how to define rewards. Because you can always do various environments with Amazon or Facebook. This is very difficult in engineering, but this thing can always be done. The hardest part is how I design the task. This

task has difficulty, real value, and a good reward.

And the reward is not very noisy. It is a reward based on rules or white box, not a black box reward. I think

the fact later proved that this is the key to success in RL today. The most important task like math and coding

RL today. The most important task like math and coding First, it's based on the result, not the process Second, I have a very clear reward based on the rules, not based on strange people or model preferences If the answer is 3, then it is 3. If the answer is 3,

then it is right. If it's not 3, then it's wrong If you do any other design, it seems that there will be hacking If you define rewards based on the process then you may encounter hacking. If you optimize the number of people or the number of machines, then you may encounter hacking. Then

you may have a very beautiful code, but it may not solve the problem.

Then I did other tasks, I think it's also a failure. For example, Swip Bench, including some other, like Kali, or other various tasks. I think one point is based on the result, not the process. The second point is based on the rules of the white box, not based on people or model number. Because OpenAI

has five categories of products, if it's based on agent, that is, based on the definition of the task, to make some categories for possible products, with the emergence of model capabilities, when we want to use model capabilities, how can agent make a category? Is there such a framework in your mind? I now feel that

category? Is there such a framework in your mind? I now feel that there will be different types of applications, there will be different challenges, and these challenges may be Is it true or not? It's hard

to say who is harder or who is easier. From a certain point of view, humans also have such a problem, right? For

example, who is more powerful, Rockefeller or Einstein? It's

hard to decide. As a CEO of a big company and a mathematician, which one is harder? I think this is...

It may be different or different challenges. But for agents, there is another point.

Maybe for people, a very simple or difficult thing, for agents, it may not be so simple or difficult. For example, for people, being a customer service is much simpler than being a software engineer. Right? His salary is also much less. And the documents needed or the various resources needed are also

much less. And the documents needed or the various resources needed are also much less. Now, on the other hand, I think software

much less. Now, on the other hand, I think software engineering is a much easier thing than customer service. Because in

software engineering, you have a better environment, a clearer reward, and more data or more reasons. If you want to do a very robust or reliable customer service, there is a reliability challenge. So I

think we can divide human work into many categories. But humans have many different aspects of challenges. And for machines or AI, the challenges of humans are relatively difficult to respond to. It

may not fully respond to AI. What kind of tasks are more suitable for agents? What kind of tasks are suitable for people and agents to do?

agents? What kind of tasks are suitable for people and agents to do?

What kind of tasks are suitable for people to do? I think from a high level, I think there are different ways to divide tasks. I think

from one way of dividing, some tasks are more about reliability, and some tasks are more about creativity. For example, in 100 times, you need to make sure you don't make any mistakes. If you only make the user happy 85 times, and you didn't make the user happy 15 times, then you might

be fired. One task is to do simple tasks but do them

be fired. One task is to do simple tasks but do them very reliably. The other task is to prove the logic of the

very reliably. The other task is to prove the logic of the problem, write a difficult code, or create a literary script. I can try 100 times, and if I do it well once, I

script. I can try 100 times, and if I do it well once, I will succeed. These two tasks require different

will succeed. These two tasks require different challenges. There is another distinction.

challenges. There is another distinction.

You can go for the depth and breadth of the task. For example, the cursor is a very short loop. For example, I'll change this file. Maybe

I'll finish it in three seconds. For some things, I might need 30 minutes or Three hours or three days. From this perspective, I need long-term memory or long-term memory. And then from the scope of the task, for example, I want to solve this bug. Versus I want to build a

repo like Windows. Then I will have That is to say, the difference in the vastness, right? That is, what a person can do, what a company can do, what a team can do. From this perspective, I think we need multi-agent research. From reliability to creativity, which task is

multi-agent research. From reliability to creativity, which task is better defined by agents? What should be its order and steps? I think

we can do a lot of different things in parallel. And there is actually a very simple... It's a method of designing metrics. For example, when we do coding, we have a very traditional metric called "pass at k". It

means that if you do, for example, if you write "k" on the same code, what is the probability that you will succeed at least once? You can imagine that when your "k" is getting bigger and bigger, your success rate will be even bigger. And then you'll find that a lot of times, coding research

bigger. And then you'll find that a lot of times, coding research will report "pass at 100" That is, I run a task 100 times, what is the probability that I will at least succeed once? But we published a research called "Tallbench" last year Its idea is that, in fact, for another type of

task, such as客服 You need a matrix that is exactly the opposite of it We call it "pass-hat-key" 就是Head 就是一个 密次的那样一个符号 就是说你做case 永远成功的概率是多少 或者说起码 失败一次的概率是多少 就是说我觉得从播放程度来说

有些任务我们需要去optimize past atk 有些任务我们需要去optimize past headk 但是往往我们现在 More importantly, success rate, which is passed at one. Or we value pass at 100 for coding.

We don't really care about the robustness of simple tasks. And I think the reason for this is because everyone is still doing some benchmarks for AI.

Everyone is still doing some tasks for me, instead of doing some actual applications. But if you want to accept this message transformation, I think it's very natural.

applications. But if you want to accept this message transformation, I think it's very natural.

Some applications just need robustness. I just need to optimize robustness. Now I don't think I'm aware of this. But I think if everyone is aware of this, this will make a lot of progress. Actually, the startup company is very worried that the emergence of model capability will swallow the agents that the startup company does. In the long term, what do you think the barriers of a company like

does. In the long term, what do you think the barriers of a company like Curious is? Which agents do you think the model company will definitely do? What things

Curious is? Which agents do you think the model company will definitely do? What things

do you think the startup company has the opportunity to do? Where do

you think its boundary is? I think the thing that the startup company should be worried about is that the model company has no ability to emerge. In

this case, you really can't do anything. I think having artistic ability is a very good thing. It means you have a chance. And I think the biggest chance for a startup is that I can design different interfaces or people and the way of interacting with the digital world. That is to say, ChaiGBT or all

these companies that make models are actually doing similar ChaiGBT products.

The essence of ChatGPT is that you are interacting with the digital world like interacting with people. For example, there is a person on the other side of your chatbot, and you chat with him, or you give him tasks, or you ask him to do deep research, or you ask him to write a code for you. But his way of interacting is a way of interacting like a person,

for you. But his way of interacting is a way of interacting like a person, or a way of interacting like a assistant. If you can use the ability of model integration, create different ways of interaction, then you can create a huge opportunity. I think fundamentally, Cursor is a new way of

interaction that I created. It's not a way of interaction like humans, but a new way of interaction like Copilot. When I write this code, it can give you some hints, or I can help you edit some things. But no

human and human are interacting like this. This is also its value. I think

in the end, the ability of the model will be will generate a super app that can interact with other users. In this case, the biggest opportunity for a startup is to explore new ways of interacting and to have the ability to use the model. These two are indispensable. If you are using an old

interface and new models, then you will be easily replaced by chatGPT. If

your interaction is similar to chatGPT, then why not use chatGPT?

If you make a new way of dealing with it, but the model hasn't changed, and there's no new output capability, then you're very lazy. So for a startup company, the best opportunity is to make a new way of dealing with it, but the model has a new output capability, so you can restore these new ways of dealing with it. XGBT can also follow up with this new way of dealing

with it. Yes, but I think owning a Super F is actually a double-edged

with it. Yes, but I think owning a Super F is actually a double-edged sword for the company. Because when you already have a way of dealing with it, you will inevitably form a path of dependence. Just like in 2020, Google has unlimited resources and money, and has the best research on transformers.

But the most natural way to think about it is, how can I use these things to improve my search engine? When you have a super app like ChaiGBT, your research will naturally center around this super app, and this interaction method. You will explore new products, but whether it's a big factory,

method. You will explore new products, but whether it's a big factory, Google, or OpenAI, you will still have most of your resources around your super app. I think this is a great opportunity for a startup company. Interesting.

super app. I think this is a great opportunity for a startup company. Interesting.

You mentioned the way of interaction. Today, it's still the interaction between people and code, and people and tax. How do people interact with agents in the future? Do you think the super assistant, the HR, is a correct way of

future? Do you think the super assistant, the HR, is a correct way of interaction? If this way of interaction works, do you think there is a chance to

interaction? If this way of interaction works, do you think there is a chance to avoid the current form of CGP? HR is actually similar to It's still an assistant, but it has a language instead of text. I think this is a very obvious and valuable form of form, right? Because people have been interacting with each other

for thousands of years, millions of years, millions of years. This is the most natural form of form for people. This is definitely the most obvious super app. But

then I think the environment is where ChaiGBT is the mastermind. Or it's obvious that this is what the virtual company did in the beginning. I think the obvious thing is that I I think

Assistant or Her It's obviously one of the most important ways of interaction with people.

But I think there will be more opportunities to create new ways of interaction. Do you have any new ways of interaction in your mind? Like the way of exploration that HHPT is exploring now,

your mind? Like the way of exploration that HHPT is exploring now, or the traditional Internet interaction? Do you have any in your mind? I think Canvas is a good attempt. You can use a task to

your mind? I think Canvas is a good attempt. You can use a task to generate a new I think the data of the app

company I think most companies

haven't created a data flow, right?

Most companies rely on the ability to make good models and use them to make good models. If you want to have a data-driven model, you need to be able to find the model yourself. And you can have a good reward through interaction. I think you need to have a good reward to be able to

through interaction. I think you need to have a good reward to be able to separate good data from bad data. I think the most successful case is Meet Journey. Right? I have a very clear reward, which is that people

Journey. Right? I have a very clear reward, which is that people like to pull pictures more. And this reward is aligned with my application.

That is to say, if I do this reward better, then my company will be more successful. Then this model is better. Everything is aligned. And

then in such a situation, I can go to the model myself. I

can do data lunges. And then what you do must be more...

If I'm the Cursor CEO, will you do the pre-training?

That's a good question. I

think I will definitely try to train models.

But I think it depends on the situation whether you do pre-training or not. I think coding is a very mainstream task. All the big companies

not. I think coding is a very mainstream task. All the big companies now will do their own coding. So all the pre-training, post-training, R, and other things will consider this. In this case,

What's the tree structure

around Agent?

如果是基于Foundation Model 然后基于Resonator 然后往上涨Agent的整个的生态数 你在脑海里是一个什么样的结构 我觉得就是有两个方向吧 一个方向是就是Fundamental的Research会怎么演变 或者说这个方法会怎么演变

我觉得另一个是应用 或者说它的交互方式会有什么样的演变 然后从盟主程度上来说 They must have different ideas, but I think it requires different people to explore different aspects. For example, I didn't innovate

on product or fundamental research, but I did innovation on interaction.

I think it's important to have a memory, intrinsic reward, and multi-agent. How can I make an agent... I think this is

multi-agent. How can I make an agent... I think this is similar to what OpenAI and the other innovators and organizations are talking about. As an innovator, you need a long-term memory. For example, I studied Ferman Theory for 20 years. I need a

memory. For example, I studied Ferman Theory for 20 years. I need a long-term memory. 我需要一个长期记忆。

long-term memory. 我需要一个长期记忆。

但是基于这长期记忆还不够 你需要有一个内在的reward 对吧? 因为直到你证明的那一刻,

对吧? 因为直到你证明的那一刻, 你是没有任何外在reward, 对吧? 你也没有获奖,

你是没有任何外在reward, 对吧? 你也没有获奖, 你也没有做任何事情 没有人给你任何feedback。

你需要自己给自己一个feedback。 那这个事情是所有innovator最重要的事情。

不管你是艺术家 还是科学家 还是文学家 还是任何创作者 对吧? 另一方面, 我觉得作为组织,

对吧? 另一方面, 我觉得作为组织, 你需要解决的就是说, How to work together with agents and how to make multi-agent scale. And I think from a certain point of view, now agents are probably like a normal college student. Be a digital intern, right? Maybe

this is the third generation or we say AGI. Maybe it's a normal college student who can do things on the computer. But the boundary of human society is that I think it's

natural that these two things are very important.

In your vision, I feel like there are a few key things that need to be improved. For example, long-term memory. Do you think long-term memory is a problem that can be improved in the short term? Maybe. Of

course, it depends on how short the period is. But I think it will definitely improve. If something is worth enough, it will always improve. If you are optimistic

improve. If something is worth enough, it will always improve. If you are optimistic about technology. You need to talk about this. This is

about technology. You need to talk about this. This is

from the context, the long context, the structure of the model itself, some changes happened. I don't know how much I can share, but my belief is that

happened. I don't know how much I can share, but my belief is that I mentioned the utility problem in the blog, right? Why is our model so capable of reasoning, so good at tests, so good at playing games, but it hasn't created enough economic value? I think the fundamental reason is that it doesn't have these contexts. And in the human society, it's a bit tricky. Of

these contexts. And in the human society, it's a bit tricky. Of

course, we write a lot of things. We use text, Google Docs, Notion, and we record a lot of things. But there are a lot of contexts that are always in the human brain. It's a maintenance based on distribution.

For example, the boss and some of your behavior habits or something that is difficult to summarize in language. These contexts exist in the human brain. Humans

can never write all these things down. This leads to 人是不可或缺的,因为只有人有这种能力,就是说进入这样一个环境,然后去获得这样的一个context,对吧?

就是说如果这个问题解决了,那我觉得可能utility问题就可以很大程度解决了,对吧?

因为这个世界上大多数人并不是这个Steve Jobs,或者也并不是爱因斯坦,他可能只是一个普通人,他数学推理能力或者whatever也没有欧三强,但是他能够去manage context,对吧?

For example, after 7 days of working in this company, apart from the things you see in the text, it also has some context that you can accumulate.

This context makes you have an advantage over O3, right? Because O3 doesn't have these contexts. You have these contexts. Although you are not as smart as O3,

these contexts. You have these contexts. Although you are not as smart as O3, you have these contexts, so you did better than O3. You just mentioned that models or agents need an internal reward system. But it seems that it's not yet available. If you want to give it an internal reward system, will it

available. If you want to give it an internal reward system, will it be possible to change the weight of some models while I continue to learn independently? Then it will become smarter. How far do you think it is? I

independently? Then it will become smarter. How far do you think it is? I

don't know. I think... There will be a day, but it's hard to predict when. Of course, the way he improves himself may change his weight, or

when. Of course, the way he improves himself may change his weight, or maybe he has a long-term memory of language, or maybe a long-term memory of imitating or other things. But he will improve himself. But

I'm not sure when exactly, in what way, and when. Do you want to talk about the internal reward? Like I just said,

is

He got a sense

of security.

Right, so curiosity, or sense of control, or sense of security, there's some inner motivation that makes him do these things, right? Otherwise, you'd be very difficult to explain from a rational perspective why he would do these things. But what's interesting is that I think that when a person grows up, he has a... When you're a

baby, you're actually based on vision, based on physics, based on... It's a physical world, an understanding of the world. What you learn is how to combine your visual, auditory, auditory and your Google capabilities. But when you grow up, you have an understanding of the world based on language,

reasoning or text. How does this world work? How can I open a company?

How can I get promoted? How can I do all kinds of things? You're

not playing a physical game, but a text game. In this text game, of course you have such a memory experience, but it seems to be very different. I think this is a challenge now. Traditional AI, for example,

very different. I think this is a challenge now. Traditional AI, for example, you go to play maze or you go to play some robot imitations, it can define some, for example, based on world models or based on various human-like motivations. But when you play a text game, how do you make a

human-like motivations. But when you play a text game, how do you make a memory experience? This seems to be very different. In your

memory experience? This seems to be very different. In your

research on agents, do you have a deeper understanding of people or any other things? How do you see the difference between people and agents? My biggest feeling

things? How do you see the difference between people and agents? My biggest feeling is that I realized that people can see light because they can reason. I think

this is probably the most important takeaway. And I think this is very interesting because when I was in MIT in 2018, I was in the lab of Josh Tenenbaum. He was a scientist. I learned a lot of cognitive science. The core of cognitive science or computational cognitive science is that

science. The core of cognitive science or computational cognitive science is that although AI has made a lot of progress, it has a lot of problems. We have to see what are the advantages of humans and how humans can do these things. Why can humans do this better? For example, humans can make

these things. Why can humans do this better? For example, humans can make mistakes in a few samples, but machines can't. Why? We have to find these methods in humans and apply them to AI. But then I realized that these AI systems in work still differ

from humans. Like skilling law, RL, and many other

from humans. Like skilling law, RL, and many other things. They are very different from humans. I think

things. They are very different from humans. I think

a better way to approach it is to think about what humans can do and what machines can't do. This is an objective thing. But once you find this problem, you can focus on

thing. But once you find this problem, you can focus on How to solve this problem? You don't have to rely on how people solve this problem to solve this problem. For example, I think people can do things now. For example, I can enter a company. I can work seven days

now. For example, I can enter a company. I can work seven days or be an intern for three months. Then I can accumulate the context of this company. Although I may not be very smart, I am a second or first

this company. Although I may not be very smart, I am a second or first grade graduate student. But I can do a lot of things that AI can't do now. This is an objective fact. How to solve this problem?

um

能做或者机器做不了,这是一个很robust,很客观的事情。

但至于就是说,人是怎么能做这些事情,以及我们要多少程度上借鉴这样的一个方法, 这是一个我觉得更主观或者更noise的问题。

因为一方面,神经科学或者认知科学,它也没有说百分之百解决了这些问题, 它只是说我提供了这样的猜想。

另一方面是,即使它是一个被confirmed的事情, 比如说人,人的视觉其实是个相对被, Again, I think it's a utility problem. I

problem. I think a lot of problems are not like people but they have more value.

foreign

再一个公司打工,然后和老板搞好关系,然后去完成各种各样的任务,那这个事情人就是比AI现在做得更好,那我们就应该适度更像人。

你怎么思考人和agent未来的关系?

要给agent发身份证吗?

我觉得这是一个交互方式的问题,就是说很有可能未来有很多agent,但他长得并不像人,或者你和他交互的方式并不像人,他可能是个平台或者是一个 Or a page, or a game, or something else. Then you probably won't make it rational, right? But of course, I think there will definitely be a lot of agents

rational, right? But of course, I think there will definitely be a lot of agents like this that are rational. If the agent has a long-term memory, is he your friend? If he is your friend, then the person and the agent are

your friend? If he is your friend, then the person and the agent are equal. Should we give him a ID? What is the purpose of the

equal. Should we give him a ID? What is the purpose of the ID? He is a separate individual and has joined us. I think it's possible.

ID? He is a separate individual and has joined us. I think it's possible.

I think these things will eventually start from utility, right?

If something is valuable, it may be produced. For example, many people are lonely and need a friend. This technology can create such an experience. Humanization is a reasonable future, right? But of course,

experience. Humanization is a reasonable future, right? But of course, he goes to make a platform, he makes a recommendation, he makes a game. he may have a lot of different ways to interact with this technology to

game. he may have a lot of different ways to interact with this technology to make you feel that he is not like a person or you can't feel any difference. Then in this context, you won't be able to dehumanize him. So I think

difference. Then in this context, you won't be able to dehumanize him. So I think it will still be based on the economic value of this thing. You mentioned

economic value, do you think AI agent will have a future combination with crypto?

For example, the combination of crypto, this smart contract, and agent, a future agent will help me complete a certain task. He has a public service 价值的计量,然后任务完成后,那就按照智能合约的约定就分配这个经济利益了。

那其实这样是有机会探索出来一个叫value-based的商业模式的, 只是说今天可能咱们还不太能衡量这个任务的客观供应价值到底多少。

对,我对crypto了解不多,但是我觉得可能一个核心的问题是, 就这个技术的演变它会变得更中心化还是去中心化?

And I think both sides have their arguments, right? It's very obvious that the new super companies like OpenAI or Antropic will become 1 trillion, 10 trillion, 100 trillion. They

will occupy a huge amount of resources and compute. They can create a super app or super platform. They will have a huge advantage in centralization. The argument

for decentralization is that 人类社会是一个网络 对吧 然后

它其实是有

uh

网络边缘到中心的可能性或者速度能有多快。

我觉得从某种程度上来说, 过去几百年发生的事情是这样的, 就是说首先这个网络变得更中心化了, 对吧,就是说贫富差距变得更大了, 或者二八定律买卖效应。

但另一方面,其实穷人的或者平民的机会, 可能是更多了,对吧? 如果在古代,比如门阀制度,

可能是更多了,对吧? 如果在古代,比如门阀制度, 九品忠贞制,或者欧洲的贵族制度 那你可能农民就永远是农民 或者印度的种姓制度,对吧? 你有阶级固化。

或者印度的种姓制度,对吧? 你有阶级固化。

那似乎技术发展的趋势是 两者都会加剧,对吧?

就是说,一方面中心化会加剧 因为效率是一个根本性的原因 但另一方面,可能... You mentioned

但另一方面,可能... You mentioned

several attempts at OpenAI in your blog post. I

found it interesting. The first project was to build a gym, a game-based learning environment.

Then there was the World of Beasts and Universe projects. 但这也没有奏效,直到GPT-2和GPT-3出现了,

projects. 但这也没有奏效,直到GPT-2和GPT-3出现了, 才发现缺失的是经验知识。 这个过程,这OpenAI的几次尝试能不能给我们详细讲一讲?

才发现缺失的是经验知识。 这个过程,这OpenAI的几次尝试能不能给我们详细讲一讲?

这也是一个探索的过程。 这是我自己的总结和揣测,不代表...

我觉得OpenAI是一个非常... 是个比较bottom-up的公司,

它最初的可能七八年 就像是一个research lab 有各种各样的想法 有各种各样的尝试 可能每个人想法都是不一样的, But objectively, we focused on the learning of the strong language. Because this was the most

popular thing at that time. DeepMind was founded in 2015. At that time, the most popular company of AI was DeepMind. DeepMind's most successful thing was the learning of the strong language. Before GPT, AlphaGo was the most successful AI project. Naturally, you have to do the learning of the strong language.

AI project. Naturally, you have to do the learning of the strong language.

You only have a different bet to surpass your previous master, right?

I think if you keep doing OpenAI, it might be hard to surpass DeepMind. Even

if you do well, or some tasks you do better than DeepMind, but when it comes to OpenAI, everyone only remembers DeepMind. So from a model perspective, if you want to surpass your previous master, you need a different bet. Turns out, GPT is a different bet. And then...

different bet. Turns out, GPT is a different bet. And then...

But of course, this is still a very unconventional thing. I can

tell you a story. My mentor is the second author of GBT-1. He

stayed at OpenAI for a year and then went to Pune to be a professor.

He was a little suspicious about this thing. He said the results were not very good. You are not the highest score on the list. You spent

a lot of cards or do this thing. There was also a skilling law.

I think you did something very anti-common. Of

anti-common. Of course, you are in agreement with this. Then I think

with this. Then I think you need to look for the next counter-competitive thing. Maybe. After what your mentor said, did anyone have feedback? To be honest, most people didn't think that scale-up GPT was the best or the most promising direction. I think this is possible. Everyone is

doing different things, right? Some people are doing robotics, some people are doing this or that. I think the biggest contribution of ILEA is that although it is

that. I think the biggest contribution of ILEA is that although it is not doing GPT-1 or doing these specific technologies, it is So it's very important that people are willing to do GP3, like Darrell or Tom

Brown. They dare to do GP3,

Brown. They dare to do GP3, which actually gives people a bigger hope. Yes.

bigger hope. Yes.

Of course, the advantage of this is that you don't need everyone to work together.

As long as you have enough people working together, you can do this. Do

you think there will be more GPT-3 moments in the next few years? I think

there will be new skilling dimensions, right? If you have a lot of memory, your test time computer will have a new way of scaling. If you have multi-agent, your test time computer will have another new dimension to scale. I think there will be new skilling dimensions. But when you have a lot of skill

dimensions, how to choose, how to use a certain application to choose the scale of these different skills, I think it's a very interesting question. That question just now, didn't the internal team form a consensus before? When did the enhanced learning become particularly important for OpenHand? I think enhanced learning has always been

very important. Even when we were doing GPT, Jiang Shuman, they were still, there

very important. Even when we were doing GPT, Jiang Shuman, they were still, there were still people doing RL. It's not like I throw away all the RL after I do GBT, but it's like I do 80% or 70% of the resources in my company, and I'm still doing some other things. And I think this is actually very important, right? Because it later proved that the success of

Tchai GBT, RL is also very important. If there is no RHF, no alignment technology, then it can't be a product. So history is not about me walking this path, then I completely abandon this path and walk another path, then I come back and walk another path. It's more soft, right? I'm doing a lot of things, and this is very promising, so I block it down even more. But

I'm still doing some other things. You have a very high-level summary, which is language through intelligent body reasoning, time translation. Is its translation a thing that has been confirmed, or is it still a kind of deduction? That is to say, why...

Why is language so unique and good? Because

it's a tool for people to accomplish many things in this world. Language is also a tool for human development. Like fire or brush. But it's special because it's a tool

brush. But it's special because it's a tool that helps you solve anything that's common or erratic. 当你学会了这门工具之后,

or erratic. 当你学会了这门工具之后, 你就可以去做很多新的任务。 比如你学会了攀岩, 对吧?

你就可以去做很多新的任务。 比如你学会了攀岩, 对吧?

但是它可能不能帮你去做很多新的任务。 但你学会了语言之后,

它几乎总是能帮你去做新的任务。 因为你可以和人交流,

可以学习 可以去思考 可以去推理。

从某种程度上来说 20年以前 大家很多时候没有把这些事情想清楚了。 就大家认为我们有语音,

大家很多时候没有把这些事情想清楚了。 就大家认为我们有语音, 有文字 有图像 有视频 有这些东西 它其实都是一些... It's all data, right? There's probably no

它其实都是一些... It's all data, right? There's probably no difference. I think the biggest difference is that language

difference. I think the biggest difference is that language is a tool invented to achieve grammar. It's more

fundamental than anything else. You're talking about language, and it has the ability to grammar. So, is it a conclusion that language studies finally has the ability to grammar? I think it's... It's my

personal opinion. But I think... Actually, there are many people discussing this matter. Of course, whether it's a spectrum is a relative thing, right? It's not a

matter. Of course, whether it's a spectrum is a relative thing, right? It's not a absolute thing of 0 and 1. But I think the reason I say this is because before this, if you go to train in an environment, you can only do this one environment. But now you can train in an environment,

you can do more environments. I think this is the most fundamental difference.

DeepSync, for example, has a very interesting result. You can do RL on math and coding, but you can also do RL on creative writing. I

think this is a fundamental difference. AlphaGo can only write WeChat, not chess. But now, for example, you can do creative writing on math. I think

chess. But now, for example, you can do creative writing on math. I think

this is a fundamental difference. Do you think that if you train to play this kind of game, it can be used to play other games? For example,

Dota is very strong. Is it strong in all games? I don't think it's easy to say. Even if it's reasoning, it may have different meanings in

to say. Even if it's reasoning, it may have different meanings in different environments. For example, if it's logical reasoning, it

different environments. For example, if it's logical reasoning, it may be easier to move from math to coding. And if it's human-eating reasoning, it may be better at other tasks. But I think the important thing is that you finally have a single model can do all the tasks. We

thought it was impossible before, but I think it's possible now.

You can do RL on many different tasks, and it can transfer to more tasks. But of course, if you only consider the task and task,

more tasks. But of course, if you only consider the task and task, it must be a little bit different. The reason why code and math are so easy to break down, do you have any idea why? Is it because they have a process of thinking? I think it's just because it's the first thing he

started doing. He started doing it because it's relatively simple. It has a very good

started doing. He started doing it because it's relatively simple. It has a very good reward. It doesn't need an environment, it's just reasoning. Now, a lot of other

reward. It doesn't need an environment, it's just reasoning. Now, a lot of other things can also be used. It's just that we started doing this, so people are talking about it a lot now. 一个agent的创业者想问你,agent如何skill

up? 因为现在的瓶颈主要是算力,agent

token用量非常的可怕 单个用户的消耗可能是chatbot的500到1000倍,再叠加几百万用户。

所以你觉得agent怎么skill up?

我觉得可能最重要的点是要找到一个好的应用,对吧? 就是说,我觉得cost本身不是问题,

问题是你的cost不justify你的performance或者你的value。

foreign

I If I do a more difficult one, like investing or deep research, I

might need a bigger model. I might have different ways to balance cost and value. But I think the most important thing is to find something

and value. But I think the most important thing is to find something that has value first. After finding this, cost will always be able to go down.

Do you think that the agent must have a researcher background? What are the advantages and disadvantages of being a researcher? I don't think it's easy to say. I think it depends on the person. I think it's hard to divide people

say. I think it depends on the person. I think it's hard to divide people into two types of researchers. And then there are strong differences between these two types. I think there is a big difference between people. And then I

two types. I think there is a big difference between people. And then I think the most important thing is to find the value. So we call it product market fit or I think finding the right value is the most important thing. Technology is just a means, right? The most important thing is to find the

thing. Technology is just a means, right? The most important thing is to find the right problem. And then, it's a bad thing to have a strong research background, or

right problem. And then, it's a bad thing to have a strong research background, or to handle natural language, or other background. Because you will be too focused on technology. You will find the right thing with this hammer. Now,

let's look at the most successful applications. I think the founders are not doing NLP or AI. For example, Chris is a 4-year-old undergraduate. Of course, Prophecy seems to be a researcher. I

undergraduate. Of course, Prophecy seems to be a researcher. I

think this is very dependable. It may not be so strong with you as a researcher. How do you see Minus, James Park and their founders? I tried Minus, but I haven't tried James Park. I think

founders? I tried Minus, but I haven't tried James Park. I think

Minus is quite interesting. I think it's a good product sense. They

have a good gene for polishing products. This product should be the main product of OpenAI. You will see. I will talk about G-minus a little bit. I think it's interesting that traditionally, people think that the things that happen are like, for

example, I make something in the factory, Then the startup company can start to copy, right? For example, I made a track GPT, then I can copy the track GPT or do something similar. But now it seems that the opposite can also be established. That is to say, you can do something in a

small factory first. It creates a new interaction or product innovation. Then the model company can also borrow or apply, right? I think this is quite interesting.

Many people say that the more the model is made, the more it feels like it's a bonus for the startup company. Because if you create a good model and you don't use it well, then the startup company will use it well. But you can say the other way around. For example, if you

well. But you can say the other way around. For example, if you create a very good way of trading, but you don't have the ability to do a good job of the model or the design, then the big company can do the opposite. If you were the founder and CEO of Minus, would

you go for a vertical direction today? I think

direction today? I think the value of Minus is that it gives people a very general and universal feeling. Uh, but I think there's a very common feeling or interaction agent and you have some killer apps

that are not contradictory. I think a more ideal situation is you have a very common interaction method. The upper limit of this interaction method or the imagination can be large enough. For example, although Cursor is It's an IDE, right? If it only does IDE, it has a lot of

space in the IDE. But if you do a very general, common product form, like Minus, it has a lot of space in the top line. But I

think the thing that is not contradictory is that you can have some killer apps at each stage. For example, it's good at doing PPT, or doing deep research, or doing this thing is very good. It's a bit like, I think, a lot of...

I don't know.

DeepSeq has a lot of space for simple, easy, and first-class interactions. But when you go to maintain it or design it, you can have

interactions. But when you go to maintain it or design it, you can have all kinds of applications that can make it grow. What changes do you think DeepSeq has brought to AI researchers after the New Year's Eve? I

think from an open-air perspective, I think we have discussed a few things. One is that Chain of thought's review, which

things. One is that Chain of thought's review, which is to say, to show such a long chain of thought, seems to be a very important thing. It is a breakthrough in product form.

That is to say, many times, the world is like, there are many technological accumulations, it's like a flood reaching the gate. You need a moment to make this thing explode, to let ordinary people, let most people feel this technology. We

would say there is an iPhone moment, you'll check gpt moment and could you do the moment now this moment could just for you know you're going to have a lot of different job who wants to do some changes and i'll do the check it because it's a little moment you know how to do it well she's a little thing that's not a matter of magic or the general league

of the city now what you're talking about is it's a little bit of a lot of work to do with it to talk about the good chance to get it in a little scene that you know how to do it well she's a little magical to you don't know what you're doing to go to the school she's a little thing of the other you know you're in

the city But I think that's one thing. The other thing is to think about the resumption. Sam also talked a lot on Twitter about how OpenAI is a bit of a no-go. But it's worth thinking about it. And it's something we should do. We

would think that resumption would be a bit behind. Because

this is... It's not like Linux, like a operating system. I have 1000 people, I can give each person a share of the profit, and I can make this thing very good. It has a very good distribution of interest. It feels like making this model is more like I have 20 particularly powerful people, and I have a lot of money. I only need 20 very strong people to do this

thing well. I need a very special organization, a very special resource concentration,

thing well. I need a very special organization, a very special resource concentration, a very special person. In this case, the advantage of traditional open source is not great. Facebook may not do well in open source. In the US, people

not great. Facebook may not do well in open source. In the US, people may ignore this habitually. For some people, doing good open source is a non-trivial thing. Because first you need enough resources, you need strong people, you

thing. Because first you need enough resources, you need strong people, you need a good organizational culture, and you need commercial justification.

The best case scenario is that you are a charity worker with millions of dollars and you go to the world of wealth. This is a small-scale event, but it happens. There is such a person doing such a thing. I

think this is worth reflecting on. I think I will think about it. DeepSick

includes a lot of organizational structure, including its engineering, including its infrastructure. I think there are many places worth mentioning. A AI researcher

infrastructure. I think there are many places worth mentioning. A AI researcher wants to ask you, he said he has limited imagination for agents, so he hopes you can talk about it. And you said your ultimate goal is to create the most powerful agent in the world. I saw you said it before, what do you think it will be like? When did I say

that? You said it in a volunteer interview, in a volunteer community. Okay. Yes, I

that? You said it in a volunteer interview, in a volunteer community. Okay. Yes, I

think traditionally, or most people's imagination for AGI is a model, right? It's

like the smartest person in the world. He has all the knowledge and abilities. He

is smarter than us. He is the most intelligent. I feel that there are different ways of interacting with each other. There are different definitions or strong boundaries. In the end, the boundaries of intelligence are determined

strong boundaries. In the end, the boundaries of intelligence are determined by different ways of interaction, not by a single model. From this

perspective, I think there is a lot of room for imagination. Now everyone

only imagines doing assistant. This is very obvious, there is a lot of room for improvement. There are also many ways of interacting that are not born yet. Just like when the Internet was first born, the earliest Super App, I upgraded Mail to Email. Even Amazon is

already a very innovative thing. I think it's a bit like that time now.

Our imagination is still limited by the way of interaction in the past. I think this will obviously create many new ways of

the past. I think this will obviously create many new ways of interaction to change our world. What do you think the strongest agent should be? I haven't thought about it yet. I think for different tasks

should be? I haven't thought about it yet. I think for different tasks and interactions, different agent systems will be needed to solve them. This is how I feel right now. I think this model can be

them. This is how I feel right now. I think this model can be shared in many ways. But if you talk about a system like this, I think it will be like asking what is the most powerful Internet website or company. This is a difficult question to answer. Because it is a multi-faceted,

company. This is a difficult question to answer. Because it is a multi-faceted, it has many different aspects. I think AI is also likely to become like this. OpenAI may become a Google. It will become a very important part

this. OpenAI may become a Google. It will become a very important part of this new world. But I don't think it means that the world will be... - It's not a single-level. - ...taken apart by a single-level thing like

will be... - It's not a single-level. - ...taken apart by a single-level thing like this. I think if that happens, the world will become very dark. Most people will

this. I think if that happens, the world will become very dark. Most people will have no value. What do you think of the future of agent's ecosystem? I

think it's like when everyone was starting to make apps back then. Maybe

2011 or 2012. If you go back a few years, what do you think the world will be like? I think...

foreign

There are different ways of interaction, and they may train completely different models. Maybe even from pre-training, the skills needed or a lot of things are

models. Maybe even from pre-training, the skills needed or a lot of things are different. For example, maybe another way of interaction is that I want to

different. For example, maybe another way of interaction is that I want to make a friend. My friend may not need to be so good at math or physics. Or if he is so good at math, it's a little fake, right?

or physics. Or if he is so good at math, it's a little fake, right?

Then he may not have a good memory. He may also make mistakes. He

has feelings. He is not very rational. Maybe this thing is also valuable. Maybe someone

will do this. Maybe It's hard to say that it's stronger than ChaiGBT because it's a different application and it has different values. It's possible that there will be an agent-based society. If you think the center of the limit is the context of the limitation, or to put it this way, why are there so many people in this world with values? It's not because their mathematical or

cognitive abilities are better than others. It's because they have some of their own information.

This information is what they have and what others don't have. For example,

there are many middlemen who have this information gap. People

with this information gap still want to maintain their rights or resources. Maybe such people will invent a more multi-agent or

resources. Maybe such people will invent a more multi-agent or more distributed network. For example, information is very important in the trading world. Everyone may only have a small part of the information. In this case, there might be a new form of

the information. In this case, there might be a new form of communication. It might be a multi-agent. Each of us has our own agent.

communication. It might be a multi-agent. Each of us has our own agent.

Between agents, I can exchange information with millions of people, or make transactions, or achieve certain things. I think

fundamentally, these very strong pillars, these very strong nodes, have the motivation to continue to make this thing more centralized. But

centralized. But Again, just as I said, it may become more centralized and more diverse,

and not contradictory, right? I think what we

right? I think what we just mentioned about the evolution of history has two factors. One is the degree of centralization or the scale of the gap. The other is the degree of complete cross-sectionalization or the speed or possibility of going from one edge to the

center. But the third factor is the diversity of the network itself or its

center. But the third factor is the diversity of the network itself or its complexity or its diversity. This thing is getting better and better in history.

Although The biggest companies in the world are becoming more and more powerful in the world. But there are more and more industries in the world. These two things can

world. But there are more and more industries in the world. These two things can exist at the same time. The more important thing is that the big model technology is not monopoly. The top three companies in the stock market seem to be able to catch up to a certain level. If the open market is monopoly, it's more terrible. I don't think it's monopoly for the moment.

But if you can find a product form, In that, the advantages of research can be converted into commercial advantages. Then it will produce a barrier. I

think maybe now, for chatGP, maybe a better thing is memory. I think

this is a place that may produce a barrier. Because if there is no memory, then everyone is fighting for whose model is stronger, right? But after having memory, I'm not fighting for whose model is stronger, but rather I actually

haven't used memory feature that much lately.

I think they've been doing some body sounds recently. I suspect that they've become better at generating or using memories. Including that

they can more effectively extract or retrieve from many user conversations. I don't really understand the details of this. But I

conversations. I don't really understand the details of this. But I

think intuitively, it's something that might generate connections or be related. Do you think that MCP is also memory?

Because I have a lot of context in my personal software and corporate software. MCP is also a way to hack my context. I think from a

software. MCP is also a way to hack my context. I think from a political point of view, yes. From a political point of view, there is a memory hierarchy in the world, right? From an Asian perspective. But the most external layer of this memory hierarchy is always the environment. It's always

like... You consider a computer, right? It has a memory hierarchy. I go from CPU to memory to hard

memory hierarchy. I go from CPU to memory to hard drive. But the most external memory is always the

drive. But the most external memory is always the external environment. For example, I plug in a USB

external environment. For example, I plug in a USB and pull out a USB. Or I upload something to the Internet. Or I make music and turn it into a CD.

the Internet. Or I make music and turn it into a CD.

Right? 外部世界永远是memory hierarchy的最后面一层。

这个是我前年冬天我读了一本 就是冯诺依曼死前写的最后一本书 叫做The Brain and the Computer。

我觉得他写的最让我印象深刻的一句话就是说 Essentially environment is always the most outer part of the memory hierarchy. 我觉得这个事情还是挺哲学的,对。

hierarchy. 我觉得这个事情还是挺哲学的,对。

For humans, you have your memory hierarchy, your working memory, your long-term memory in your brain. But the most external part is your notebook, your Google Doc, your Notion. These things are your most long-term memory, or your most external

your Notion. These things are your most long-term memory, or your most external memory hierarchy. What do you think is the relationship between long

memory hierarchy. What do you think is the relationship between long context and long-term memory? I think long context is a way to achieve long-term memory. If you can achieve 100 billion, or 1000 billion,

long-term memory. If you can achieve 100 billion, or 1000 billion, or infinite long contexts, it's a way to achieve non-tumorary. It's a way that's very different from human beings. But it's possible, right? It's a possible way. Of course, I think there are many different ways. It's hard to say which one is the best

or most suitable. Do you have any preference for the current long context, linear, sparse, or hybrid? I don't want to comment on the method, but I want to comment on the evaluation and task. At least until last year, we were mainly doing some so-called long range arena, which means "hey in the haystack".

I have a very long thing and I insert a sentence in the middle, such as Yao Shunyu is now in OpenAI, and I will ask you this question. I don't want to judge by the method, but I want to judge

question. I don't want to judge by the method, but I want to judge by the task. I think this is a necessary but not sufficient task. You can

complete this task under the pre-condition of long-term memory work, but it's not fully a condition yet. I think it's a necessary condition. But I think now everyone is a

condition yet. I think it's a necessary condition. But I think now everyone is a little bit in the necessary condition and has not created...

更多更难或者更有价值的东西 我觉得这是一个问题 当你没有这样一个很好的 评估的方式的时候 我觉得就很难讨论 各种方式的好坏 你在文章也说了 忽视任务的本身的定义 和评估标准的重要性 那你觉得应该怎么去定义和评估呢 比如说我们怎么去衡量一个agent

你会有哪些北极性的指标 还是要思考怎么去创造 更多现实世界的价值 对 然后当然这个事情 在不同的领域 在不同的应用下 There are very different task designs, very different methods, and very different things. But I think a big trend is that we should think more about the actual value, rather than

these things that are defined as tests or games. Because we

found that once you can define a test or a game, it is not far from being solved. The reason why the real world is difficult to solve is because it is not a game or test. The big feature is that when it's designed, it already has a very good design reward

or a very good design answer. When you already have a very good design reward or a very good design answer, then you now have this general recipe, you have such a method, a common method, then it's not far from being solved. But the real world is also very difficult to solve, because it doesn't have

solved. But the real world is also very difficult to solve, because it doesn't have a standard answer, it doesn't have a Do you think we need to push back more basic

settings in the future? I think it's

future? I think it's necessary. From a human perspective, humans have always been doing this. The most important

necessary. From a human perspective, humans have always been doing this. The most important thing is to break the most basic assumptions. Maybe the assumption that I am most concerned about now is that an assessment of something is based on, for example, 500 tasks. You run 500 times for each of these 500 tasks. Then you add these

tasks. You run 500 times for each of these 500 tasks. Then you add these parallel data together to become your reward. But I think this is completely different from human. That is, you work in a company. The important thing is how

human. That is, you work in a company. The important thing is how much better you will be in a day, 30 days, or a year. It's

not like I put you in the first day of the company in the 100-star universe. I think this is a basic difference between the two. What do

universe. I think this is a basic difference between the two. What do

you think is the difference between the competitive environment and the real-life environment? Because there

is also a saying that some model companies have a high benchmark in the competitive environment. That's not good for real-life. Some model companies are better in real-life. I

environment. That's not good for real-life. Some model companies are better in real-life. I

think it's... We need to consider more about the actual value of something. Because

everyone finds... I think the other motivation I wrote about in this blog is that everyone finds it too easy to brush up a list. You can always brush up a list very high. But in this case, some things have a high actual value, and some things have a low actual value. This is a problem. I

think we need a better way to evaluate. Do you think agent will cause a big explosion in recent years? I think it will. I think we...

I think a good AI product manager is a good product manager and can think first.

Because AI is a fast-changing thing. But I think the

thing. But I think the most important thing is the human nature or the human needs. I think

this is a slow change. I think you can find a good need and you can reverse it from the first point of view. You can

say that I need to apply some kind of technology to accomplish this. I think this is also very important. What do you think of Xiao

this. I think this is also very important. What do you think of Xiao Hong's podcast? I think it's quite interesting. What I remember

Hong's podcast? I think it's quite interesting. What I remember most is that he said that VC is a very expensive way of financing.

It's not when you're not doing well, but when you're doing well. I

think that's a very interesting statement. I think it has a lot of different perspectives on the issue of capital. I think it's very new and interesting for me. Will you consider starting a business? I think most people will consider

me. Will you consider starting a business? I think most people will consider starting a business. Because now is a very exciting time. And now there are many open-ended entrepreneurs starting their own business. I need to do something more challenging, so I naturally start my own business. But I think I should

find a good thing. I still like to think clearly before doing something. What

are your predictions for the future of 12-24 months of agent? I think

first of all, the chatbot system of these model companies will evolve into a very natural agent. It will be a very natural transition. For example, the default

natural agent. It will be a very natural transition. For example, the default chat GPT or the default agentic way of interacting with the Grog or the Chats GPT or the Antropi Cloud. I think chat will still be kept as a subsidiary, but I think agent will become a more important way of interacting.

And then I think there will be new products similar to Cursor. Cursor

is a co-pilot in the environment of coding and IDE. But I think there will be opportunities to do some new environments or co-pilots in larger environments.

Then these two big ways of trading are mutual or different, true trading. One

is, for example, I have a model based on a remote virtual machine or an environment, and then I do a lot of things in it. On the other hand, there are many existing environments, such as existing software or existing scenarios. Then I introduce the ability of agents or AI. I think there

existing scenarios. Then I introduce the ability of agents or AI. I think there will be a big trend in both aspects. So I think things like Devon or Cursor will continue to develop. If we want to push for more agentic capabilities, where do we need to work? Do we need to

work in Pertuning or RL? If I'm an industrialist, I can't do these two things. I'd rather try out some RL processes.

这两个事情是你可以去做,并且

It requires a lot of design, a lot of infrastructure, a lot of engineering, a lot of various things. I think it's far from being good enough now.

There are still a lot of room for improvement. I think there's another very important thing, which is how to build an ecosystem, or how to build the user's intention, or the user's context, or intention. And I think there's still a lot of room for improvement. You mentioned about agent's infrastructure. If in two

years, the number of agents in the digital world has been greatly increased, Do you think that agents need to redesign a new digitalized system? Agents

need virtual machines, computers, browsers, search APIs, identity certificates, economic systems, etc. This type of infrastructure is designed for agents, not for people. I

personally feel that the world will not become so distributed in the next two years. It may be more centralized, with some super apps. For example,

years. It may be more centralized, with some super apps. For example,

now, of course, there are many startup companies in AI, right? But there are only a few of them that are doing well, right? I think maybe in two years, there will still be some super apps. And these super apps will have their own infrastructure, their own environment or interaction. Two things can be done in a hurry. One is based on the local digital environment of users. For example, I have

hurry. One is based on the local digital environment of users. For example, I have a phone, I have a computer, I have a software. I'm already here.

How can I... How can we make it better? Another thing is to create a new environment. For example, I do deep research or I am an operator. I actually

new environment. For example, I do deep research or I am an operator. I actually

create a new environment. I think there are still a lot of things to do.

What about two years later? I don't think anyone can see the world after two years. It's changed a lot. I think you can have some predictions

two years. It's changed a lot. I think you can have some predictions like a science fiction. Or your ideas or your prospects. But it's hard to say. I don't think anyone can predict what will happen in two years.

to say. I don't think anyone can predict what will happen in two years.

Do you have any good points about OpenAI? Do you know what things are its main track? What things are the opportunities for companies to start? Do

you have such a feeling? I think every company, once it has its super app, all its things will surround its super app. For example, when you have track GPT, the way you train the model, including organization architecture, including many things, will surround track GPT to buy, right? I think if you do something that is very different from the chat GPT, there will still be a

chance. By the way, why is your article called "Second Half"? Why is it now

chance. By the way, why is your article called "Second Half"? Why is it now "Mid-term"? Because I think from the methodological point of view, we have just achieved a

"Mid-term"? Because I think from the methodological point of view, we have just achieved a moment like a starting point. That is, we finally have a very common way to solve all kinds of things. If you ask an AI researcher ten years ago, he would think, for example, to do translation, to play games, to

use computers to book tickets, and to do math is completely different. The methods they need are completely different. The people they need are completely different. They

are completely different communities. They hold completely different meetings. They

have completely different papers. These things have nothing to do with each other. But now,

finally, these things can be solved with one method. This is a fundamental method.

You're talking about a longer history of AI, not just a matter of one or two years. I think so. The world is divided by drinking and drinking. I

think everyone has been divided for too long. A researcher who answers questions and a researcher who writes numbers may not have had any communication at all five years ago.

Because these two things are completely different. Now this may be the same thing.

It's a very inappropriate analogy. For example, in physics, for example, Newton's theorem was suddenly proposed. People found that this world can actually be understood in a unified way.

suddenly proposed. People found that this world can actually be understood in a unified way.

Now I feel that we have realized that many problems in this world can be solved in a unified way. I think this is a fundamentally different thing. There were many great things before. I think it's a great thing for

thing. There were many great things before. I think it's a great thing for this thing to happen, right? I think Transformers is a great thing. Pre-Tune is a great thing. There are a lot of great things, but these great things finally

great thing. There are a lot of great things, but these great things finally led to such an event. Just like Newton, there are a lot of great things before him to make him a poudon, right? That is to say, Kepler, and even Aristotle. There are all kinds of great things, but they eventually led to the

even Aristotle. There are all kinds of great things, but they eventually led to the birth of Newton's theory. I think it's a symbolic event in physics, right?

Do you think pre-training is a way to strengthen learning or to build up your skills? Agents are what we want to achieve. Pre-training and RL are both

skills? Agents are what we want to achieve. Pre-training and RL are both part of the technology needed to achieve this. But of course, you can also say that pre-training is a way to build up your skills. Because without pre-training, it's hard to do RL in these words. I think this is a very long-term

thing. Because traditionally, RL people don't care about prior.

thing. Because traditionally, RL people don't care about prior.

He doesn't care about pre-training or the knowledge of the thousand eyes. He thinks

that I have an environment, I have rewards. It's a matter of time. It's a

matter of samples that I can solve the environment. I can theoretically prove that even if it's an Internet, even if I don't have pre-training, as long as I have enough samples, I can still use violence to solve this problem. As

long as you have a good reward definition, and my training and test reward are the same distribution. But maybe the sample amount of this, or he might need to learn for example, 10^34 years, he might never learn it. In the sense of the age of the universe, right? So in this matter, I think pre-training is

the foundation for these machine language RLs. Without this, you can always do it, but in reality you can't. Do you think that tech companies should reopen the trend of pre-training? This is a very important point in the last broadcast. A non-competitor should reopen the trend of pre-training. There is a difference in

broadcast. A non-competitor should reopen the trend of pre-training. There is a difference in cost and value. The reason why there are very few people doing it now is because its cost is very high. But it seems that the additional value it brings is not that big. Because I can use these open source models or APIs, I don't seem to have... It's like a one-sided situation. I have a lot of

cost, but I don't have much added value. Because you have to do a lot of post-training after you've done pre-training, so you can experience the value of pre-training. But I think if one day, just like I said, there are many different super apps in the world, and many different ways of interacting,

and they need completely different model capabilities or models, and the value of these things is just enough to justify the cost of pre-training, I think it will be reasonable. I think it's a value and cost relationship. You just mentioned that

be reasonable. I think it's a value and cost relationship. You just mentioned that division must be combined and drinking must be divided. What do you think about the future relationship between Pertraining and R? Will more advanced knowledge be put into Pertraining? My immature idea is that there may be different applications

that require different types of agents. So it may be different in the way of construction. For example, if I want to play Go, I can just do AlphaGo

construction. For example, if I want to play Go, I can just do AlphaGo directly. I don't need to do anything else. If I have a very straight

directly. I don't need to do anything else. If I have a very straight line, this thing is worth enough. And I have a lot of data, I can create a B-ring. Then I may be able to work mainly on RL. From a

blogger's point of view, I think Google's ads or TikTok's recommendations are similar to this system. If I find a closed environment, I can do things similar to RL,

this system. If I find a closed environment, I can do things similar to RL, and I can bring more value. This is reasonable. But maybe

There are many long-term things in this world that require optimization. It needs to be like a human being. You may not know everything, but you can learn.

You can learn online. You can enter a new company and adapt to this environment to do new things. In this area, the importance of pre-training may be higher.

Because you need more transformation. So I think there may be different technical routes based on different applications. But I think the technical route is a tool after all. You just need your value to match your cost. I think the choice of

all. You just need your value to match your cost. I think the choice of technology is flexible. There is no such thing as a technical route. I think

as long as it is economically reasonable, it is possible. If you are a global super internet or technology company CEO, and today this company does not have its own model, no good research culture, or even no good AI strategy, what would you do as a CEO? I think I will learn first, right? I will learn what this is all about. Because if you

right? I will learn what this is all about. Because if you as a CEO don't understand this, then everything is difficult. I think a company's bottleneck is the understanding of the CEO. If you don't understand this, and you say I'll find some good people, I'll do something, then you may be fooled by them. I think you have to learn first. Then I think you still

have to think about the problem from the value of creativity. Because after

all, you are not a technical expert. You are the CEO of a company, and you have some scenarios, resources, and advantages. I think from the first perspective, a new technology has been created. You need to think about how to use these new technologies and resources to create new values. Of course, you can

try to do something completely different from what it is now, and something of great value, such as disabling GBT. But for companies that are rich and powerful, this doesn't make sense. First, you have to learn this technology.

Second, I think You need to think about what new value you can create. If you become a brokerage CEO, and you need to allocate $5 billion to the AGI industry, how will you allocate this money to both pay back and contribute to

humanity? This is a good question. It depends on how

humanity? This is a good question. It depends on how much energy you have or how much resource you have.

Of course, I think that these model companies like OpenAgeTropic will have value. I think it will be more valuable. I think there is another very valuable thing.

valuable. I think there is another very valuable thing.

It can accumulate user context or build a special environment. Because in the end, I think AI or AGI is a

environment. Because in the end, I think AI or AGI is a system. It needs intelligence. 环境,它还是需要有 user

system. It needs intelligence. 环境,它还是需要有 user

context 或者对 user understanding 那现在可能 我觉得有很多 user data 或者有很多 user context 的公司 就有点像 发明车之前的煤炭 煤矿 可能当时煤矿已经有 就可能 发明汽车之前的 石油公司 对吧 就是当时可能 也许也有一些小的应用

或者怎么样 但是 我觉得现在大家对于这个东西的应用 还没有足够大 可能会有一些 So it's a good opportunity. From this point of view, I think WeChat or these big platforms are still good platforms. Because it has a lot of context. If intelligence can gradually become democratized, cheaper, and

context. If intelligence can gradually become democratized, cheaper, and more popular, then having such a platform, such an environment, such a context, may be a strong barrier. So I think it's still a good investment.

今天顺宇当了很多公司的CEO啊。

我再问一个啊,如果你是微信的老板, 你会怎么在微信里做agent?

我觉得我可能会不急,我可能先观望观望。

就是我好像没有理由要急,然后我会观察,我会学习,我会学习AI。

然后我会观察,就是有没有什么新的交互方式很有意思。

但是我觉得我不会急着去做很多事情。 Because I have a place where I can't attack with one hand, why should I attack with my own hand? Maybe the most dangerous thing is a subversive innovation. I think the real

own hand? Maybe the most dangerous thing is a subversive innovation. I think the real danger is not a WeChat-like thing that defeated WeChat, but a very different thing that defeated WeChat. It's like WeChat defeated QQ. QQ was not worried about a QQ-like

defeated WeChat. It's like WeChat defeated QQ. QQ was not worried about a QQ-like thing that defeated QQ, but a very different product that defeated this thing. I think

maybe... I think we need to be more cautious about the new innovations.

But if it's just these incremental innovations, these small innovations, I think you can do it sooner or later. Maybe the difference is not that big. You don't

have to worry too much. Because everyone says that WeChat is good, but today WeChat has not invested much. If in the future, multi-agents, these long-term memory problems are solved, but this... Agents don't grow on WeChat, that's pretty scary. The original

network doesn't necessarily have value. I think it depends on what kind of network people have. Do you have more agent friends or more human friends? Or do

people have. Do you have more agent friends or more human friends? Or do

you have more interaction with agents in your profession or more human interactions in your profession? Because on WeChat, you have friends and you have friends. For example, I want

profession? Because on WeChat, you have friends and you have friends. For example, I want to buy something or I want to consult something. Or I want to be a lawyer or something. It's just a professional interaction. I think this is due to the importance of the human network. But I think there will always be such a network.

And based on this network, there will definitely be infrastructure and platforms. How do you ensure the security of AGI after realization? Because WeChat used to be a more responsible and safer. If in the future, the power is very strong, many bad people do bad things, and even subvert humans. How do you think the long-term

security problem will be solved? There must be an AI line. I think safety is a very complex issue. When everyone says safety, they may think differently. My

view is that I think the main distinction between safety and safety is that it's normal to pay attention to safety like a business company. Is it enough? Or do

you need to be more responsible for safety? Because for example, for ChaiGP, if he is not safe, he will delete your stuff. Or he will fail this product.

It has no commercial value. So even for the sake of commercial value, it will also pay a lot of attention to safety, right? Every product is like this.

It will rely on its own product and business. It will naturally pay attention to safety. Because if something is not safe, it is not a good product at all.

safety. Because if something is not safe, it is not a good product at all.

People will not use it. It has no value. But I think the main difference between this and this is whether you need a product-specific Or more conscious of this kind of security. But I don't think everyone has fully defined this. And I think the former is easy to solve. If you have a good

this. And I think the former is easy to solve. If you have a good app, you will always have a way to solve security problems. I believe.

Because once a thing has enough value, someone will always need to solve this security problem. Because if solving this security problem is a necessary condition that

problem. Because if solving this security problem is a necessary condition that brings value, and this thing has enough value, then I think it will always be solved. But as for the second one, I think there will be a lot

solved. But as for the second one, I think there will be a lot of uncertainty. I think it's hard to evaluate. Are you worried

of uncertainty. I think it's hard to evaluate. Are you worried about the security problem after AGI? I am worried, but I think the biggest problem now is that AGI has not been realized yet, or we haven't created enough value. I think if something hasn't created

enough value, I'm worried that it's too powerful or too popular. I

think it's like

也许当你能够

If you can deal with complex contexts and have enough autonomy or decision-making power, you may have a sense of consciousness. We think we have consciousness because we are processing information at a high frequency. We are constantly processing information and making decisions. We are thinking about various

ideas. We can choose to do many different things and have different consequences.

ideas. We can choose to do many different things and have different consequences.

If a system can do these things, it can be deemed as conscious.

I think this issue hasn't been properly addressed. You said you want to address tasks and problems. What are some of the most important problems you're thinking about?

I'm actually interested in many problems, but I can't answer all of them. But I

will try to chat with many people and try to understand what's happening. What

are some of the most important problems you're thinking about? Do you have any animal moments recently? For example, the past year. I think one insight is that This technology is probably more common than the current product line. There may be more super apps. I think there is a chance. I think

line. There may be more super apps. I think there is a chance. I think

there are still many small dunks. But I think the big dunks are that there may be more innovative and innovative super apps in the world. The

boundaries of this world may not be defined by one institution, but by different super apps.

你的文章真的是基于deep research,基于你的演讲稿写的吗?

是的,但是我,我就是那个introduction,我基本上就是改的比较小,但是可能后面有些段落,我基本上就善照从行了,对,因为我觉得这个东西它解决的问题是一个initialization的问题,就是说它给了你一个初始化。

If this initialization can make you enter a new stream, it's not that hard to rewrite it yourself. But what I think is important is that it can initialize your mind flow, initialize your new stream. But if you can enter a new stream, you can rewrite it and change it. There's no difference. Or maybe it'll be faster to rewrite it. How long did you rewrite it? Maybe two

hours. As an AI researcher, you've already gained a lot of attention during your PhD. Why? What did you do right? What I want to do is...

PhD. Why? What did you do right? What I want to do is...

There are two lines. One is a simple and universal method. The

other is a task with different and practical values. These tasks often involve creating new values in the real world of digital.

Traditional tasks are created in the virtual world, such as games or exams. Or in the physical world. I think the real world of digital, such as computers, coding, or web, is a place for processing.

It's a huge treasure, and I just managed to dig up some things. And the

creation of GBD3, or the creation of these models, the creation of the push, gives me a chance to do some simple and transparent methods. But I think there are also some reasons for the times. Just having such an opportunity, and I think these things are the most important things for AI. On the one hand, it's a

simple way of communication, and on the other hand, it's a better task environment.

I think it requires a lot of courage and understanding. If you're doing a task that's very high-level, even if you solve it, what can you do? I think you need to do some very difficult tasks. And I

think another important point is to look at the boundaries of many things.

If you only do R, or only do reinforcement learning, or only do natural language processing, or only do one

-

Did you gain any know-how during the process of creating React? Did you encounter any obstacles? I think the hardest part is finding tasks. I think all the tasks I've

obstacles? I think the hardest part is finding tasks. I think all the tasks I've done, all the methods I've used, I think the hardest part is finding tasks. Because

I think it's very obvious that one day such a thing will become very valuable, right? An intelligent machine can do actions by being able to push it. But I

right? An intelligent machine can do actions by being able to push it. But I

think the really difficult part is You can define

it as What is the most extreme task? I can add something to my previous question. I

think my experience is quite special. Most of the good ways to launch a project are because it has a specific task. And this specific task is a very common way to launch it. For example, PPO may be doing a specific task at the beginning. Or Transformer is doing a specific task at the

beginning. For example, I think the attention type of work is actually very

beginning. For example, I think the attention type of work is actually very affected by translation. If you were doing translation at that time, then you might not be able to do something great. Because if the attention mechanism is most suitable for its task, it is a translation-like thing. Because it needs to naturally pay

attention to different places in this series. And then this attention may not be linear.

And then it needs a method like Transformer. So most of the time, people find that the great method is because you have a task and you have to solve it, and then this thing is just right. But I think my experience is special because a lot of times I think about something and I think it's very common and good. But I need to find some tasks to prove that it's

very common and good, or it will be very valuable in the future. It may

not be worth enough now, but you need to find some simple tasks to prove that it's valuable. I think this is very difficult. Just like the idea of creating a product-market fit, doing research requires a method-task fit. That's the hardest part. You have to think about it first. What is

fit. That's the hardest part. You have to think about it first. What is

your most aggressive task? A common use task. I think nowadays, even if it's aggressive, it's not aggressive. I think anything is possible. Before I

graduated, I thought a lot about how to create an Einstein or how to create a scientist. Because I was a more academic person back then. In Princeton, your idol is of course Einstein. I naturally

then. In Princeton, your idol is of course Einstein. I naturally

thought of the most interesting task. Can I discover the next comparison? Can I discover the next great scientific theory? I think this can

comparison? Can I discover the next great scientific theory? I think this can undoubtedly mark the realization of AGI or ASI. If AI discovers the great common theory. I think after I came to the U.S., I came to

common theory. I think after I came to the U.S., I came to the company, I think the human organization is also a very interesting thing. And I

think if I can create a new company, if I can create a company that is worth a trillion dollars, I think it's a very interesting thing. It's also difficult to create a business value. It's not as difficult as

thing. It's also difficult to create a business value. It's not as difficult as the invention. Why is the human organization also interesting, instead of the

the invention. Why is the human organization also interesting, instead of the human product? The product is of course interesting, but what's interesting is that

human product? The product is of course interesting, but what's interesting is that I think this is very

interesting.

Why is he the organizer of the five branches of OpenAI? I actually thought that innovators and

OpenAI? I actually thought that innovators and organizations were more equal or more related. I

asked a question in the group. Why is it harder to be a CEO and a scientist? I think it's hard to say. And the

path of your research may be different. I think they are all important. I don't have to worry about who is the 4th or

all important. I don't have to worry about who is the 4th or 5th. I think they are all important. But I don't think you have to

5th. I think they are all important. But I don't think you have to achieve one thing before you can achieve another. I think you can explore at the same time. On your journey to growth, do you think your mindset is

same time. On your journey to growth, do you think your mindset is similar to your peers? Or different? I think there are different places and different places. I think my path is quite unobtrusive. I didn't skip

different places. I think my path is quite unobtrusive. I didn't skip classes. I went to the US to study in

classes. I went to the US to study in a big environment. I didn't do anything surprising. But I have my own

anything surprising. But I have my own opinion on the value of something. I

think people tend to do things that are highly certain. Including doing research, doing

highly certain. Including doing research, doing business, doing anything. But I think it's a good time to do something higher than that. Because

there's a huge opportunity. If there's no such opportunity, the best way is to do incremental things, to do things that are strong in certainty, and then accumulate step by step. But there's a very high thing. If you dare to think, or you're very bold, or you

thing. If you dare to think, or you're very bold, or you have a lot of imagination, then good things will happen, I think. What inspires you the most in your growth? Books, movies,

think. What inspires you the most in your growth? Books, movies,

music? What is your mindset? I think reading books is very helpful. I like

reading magazines. I read all kinds of books. I think this is very helpful. I read all kinds of movies. I want to go to

very helpful. I read all kinds of movies. I want to go to all kinds of places. I feel like I've been a more general person since I was a kid. I want to try to be very universal. I

want to try to understand a lot of different subjects and do a lot of different things. But then I realized that even if one person is smart, and has experience, the knowledge he can understand or do is only a small part of the knowledge of the human society. Maybe

a better thing is to create something that is more universal and simple than you. I think I have always had a desire or pursuit for this kind of

you. I think I have always had a desire or pursuit for this kind of universalism. What does universalism mean? Because it can be simple enough.

universalism. What does universalism mean? Because it can be simple enough.

Do you know how to play the

game?

Interesting.

I have a lot of favorite rap singers. I think rap is interesting because everyone has a different style. It's a bit like the first of the five, the second of the five. Everyone has something unique. So it can show a lot of value. I think this is why a lot of people like rap. You have

of value. I think this is why a lot of people like rap. You have

your own personality, your own flow, your own way of thinking. You can create something different. It's not necessarily the best thing, but it's something

something different. It's not necessarily the best thing, but it's something different. I think this is very attractive. What is the similarity

different. I think this is very attractive. What is the similarity between him and you? He is an AI. It's interesting. I remember when GPT-3 first came out, everyone thought it was amazing. Then I thought about the first thing to do, which is to see if I could make a good rap lyrics and have content. I found it very difficult. It seems to be very difficult

today. Maybe it means that a rap singer is a job that is underestimated

today. Maybe it means that a rap singer is a job that is underestimated by people. Why is it difficult? Because it's a sweet word. Isn't that what

by people. Why is it difficult? Because it's a sweet word. Isn't that what Predict Next Token is about? I think, first of all, a good thing, or a good flow, or something that sounds comfortable, is a reward that's hard to be quantified. But I think a unique flow or something like that is a reward that's

quantified. But I think a unique flow or something like that is a reward that's hard to be quantified. Sometimes, a thing, like this rhythm, or this flow, or this style, if it appears too much, then it's not good. It's good to be unique. It's a reward that's hard to be quantified. The second point is

unique. It's a reward that's hard to be quantified. The second point is that a true great singer has a lot of unique thoughts about life. I

don't think AI can do that. Because he doesn't have life yet. Is there something that can be used more commonly than language? I think in a specific field, there will definitely be better expressions than language. For example, in the game of Go, using natural language is not the best way to think. But I think

language is not created to deal with specific tasks. It's to be able to solve all tasks and to solve people's cognitive abilities. And

then it forms a comprehensive expression. So it's not for specific tasks.

Often you will find that it has a lot of tolerance in a specific task.

But it's a universal thing. Of course, it's like we can create a new language. AI can create a new language, and it may be more efficient. But I think that in the end, English or the language of

more efficient. But I think that in the end, English or the language of the future will still be a mainstream language. Because humans

already have a strong prior, a strong knowledge. And people have such a value or motivation. He has this motivation to make the language of the machine and people look alike. So that we can better understand it, control it, monitor it, change it, and manipulate it. It seems to be

a very natural choice. What is your inner drive? What is your vision?

Who do you want to be in ten years? I think I still want to use a very simple word. You hope you can create something different for this world, right? I think if you can explore new things, I think it's interesting. I think if you go to a company like XAI or Thinking

Machine, or a company like Chatbot or Assistant, I think it's still possible to make a lot of money and be successful in the business. But if I do a company like

the business. But if I do a company like If someone else can do it, then

it's okay to let them do it.

If you look at it from the perspective of the human race, if many people can do this, then others can do it, right? There is no difference.

There seems to be no change in this society or in the overall situation. Of

course, I remember that I mentioned this matter, and then someone said that this is very fake. Because in the end, you will find that there is nothing to replace this matter, right? Relatively, it was not mentioned in time, but someone will mention it. There's no such thing as "you're dead"

or "you're not here, so no one else can mention it" But I think this is a reasonable statement If you can clearly see that someone else is doing this or someone else is doing something similar to this Then you can choose to roll with them I think if you want to roll with them, you think you will be more efficient or you can do this

better I think this is reasonable But you can do something different to explore I think you have to create value for this society. But I

think this generation is very lucky. This technology is very versatile and great. I think there is enough space for it. I think coding is

great. I think there is enough space for it. I think coding is a very obvious thing that everyone is doing. Maybe my value is to show the initial signal. If others can do it, I think others can do it too. And

initial signal. If others can do it, I think others can do it too. And

then the other point is that you want to make your life more interesting or more interesting. more happy, then you can do something you like. But it's

more interesting. more happy, then you can do something you like. But it's

hard to explain this with language. It's a taste or preference issue. What do

you suggest to the startup of agents? I think I've said this many times, it may be a bit old-fashioned, but I think it's about figuring out what your value is. I think technology is a tool. Of course, understanding technology and the trend of

is. I think technology is a tool. Of course, understanding technology and the trend of understanding technology is very important. I think creating value is the most important thing. Or thinking about what kind of value you bring to your users is the most important thing. What is

your recent taste-based bet? I hope you can see it. I'm doing

something. My last few questions are quick. What food do you like in the world? I like coconut. What place do you like in the world? I like Istanbul.

world? I like Istanbul.

一个少人知道但是必须知道的知识点。

我挺建议大家去看智能简史这本书的,我觉得有很多很有意思的知识点。

比如说,为什么大多数动物都是左右两侧对称,并且有一个像嘴一样的食物入口,有一个像肛门一样的食物的出口。

然后为什么气体是同一个口,而食物和水是两个口。

This is very interesting. It has some fundamental reasons. What fundamental reasons? For

example, you will find that if you want to do navigation, you have to move in this world, the left and right-angled structure is the best. It's

like you find that all the transportation tools in the world are left and right-angled.

Because you can move forward and backward in one direction, and turn left and right in the other direction. It's the same structure as the left and right-angled structure of cars and planes. As for food and There are other reasons for this. Out of all the books you've read, which one do you recommend?

this. Out of all the books you've read, which one do you recommend?

I think the book "Smart Explanation" is very interesting. I read it last year and I think it's very interesting. And then I would recommend various kinds of self-publishing. I think "Publishing" is very interesting. It's like you're reading someone else's

self-publishing. I think "Publishing" is very interesting. It's like you're reading someone else's life. You're experiencing someone else's life. What are the articles that you think

life. You're experiencing someone else's life. What are the articles that you think affect the AI process? I think there are many. What about the "sin"? I don't

think there is any sin. These things are all a process of accumulation. Backprop, Transformer,

GPT. I think this is a process of gathering. I don't think there is a single job that is the greatest. Based on your current knowledge, what is the most important "BAT" in your opinion? BAT has different super app platform, with different interaction methods. If you don't believe in this, the world will become very dark. Only

interaction methods. If you don't believe in this, the world will become very dark. Only

OpenAI or Antropiq have the opportunity. But if you believe in this, there will be many new opportunities. What is your MBTI? I want to say INFP, but I'm not sure. I don't remember the meaning of these words. Have you

listened to our blog before? I listened to the first half of your and Xiao Hong's episode. I think it's quite interesting.

Hong's episode. I think it's quite interesting.

今天的节目就是这样。

这里是商业访谈录,是一档由语言及世界工作室出品的深度访谈节目。

你可以到公众号关注我们的工作室获取更多的信息。

我们的公众号是语言及世界,languageisworld。

我们希望和你一起从这里探索新的世界。