Kimi Founder Yang Zhilin: K2, Agentic LLMs, Brains in Vats, and the Beginning of Infinity

-

Hello, everyone. I'm Xiaojun. Today's guest is the founder and CEO of Yuzhi Amiant, Yang Zhiling. A year ago, we published an article titled "Yang Zhiling's Re-opening of the Model-Based Entrepreneurship, A Snow Mountain that is Unseen" Then, after a year, we had another conversation.

Hello Zhiling, let's say hello to our audience first. Hello everyone. When you first started your business, the title of our interview in 2024 was "Mountaineering to the unknown snow mountain". Now it's been a year, and it's July of 2025. How do you feel about it? I feel like it's been a long time since you mentioned this word. It's been a year. AI is one day, and the human world is one year. So I don't know how many days I'll be in the human world in the year of AI.

Yes, I feel like a lot of things have changed. But I think that feeling...

- -

-

-

-

-

-

-

- The beginning of infinity. - Right. -

He said there are two words to be carved on the stone. One is "problems are inevitable", and the other is "problems can be solved". Basically, you can think that before the start of the movement, this society was a static one. People didn't pursue innovation. You might use a lot of magical ways to explain the phenomena you observed. But these explanations may not be good. For example, you see this

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

-

-

-

The so-called beginning of infinity.

AI. For example, we use K2 to do a lot of model training or data processing related work. These things used to require artificial code writing, or some students may not know how to write code, but they couldn't do it before. But now, for example, a lot of data processing or model analysis, even model training, you can use this

What do you pursue if the snow mountain is endless?

-

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

-

-

-

-

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

I think that in essence, it's about making the model try and reflect on itself. I think that reflection is the key here. Reflection is basically two kinds of abilities.

you can think of it as a new "proposal" that you can keep proposing in the process of solving a problem. This proposition will be self-evident. For example, if you propose this proposition, you can ask yourself whether it is right or wrong.

-

- Pass at K

-

-

-

-

多轮的agent的强化学习的方式 或者说你通过这种强化学习技术 训练出来这种agentic的模型 那它的特点就是说

-

通过交互得到的外界给我的反馈

外界給我的這個新的狀態的更新是有關係的。 它有了一個過程。 對,然後這兩個東西它都指向了同一個東西, 就是所謂的test time scaling, 就是你可以在測試的時候, 或者說在推理的時候做更好的規模化。 意思就是說,比如說我們之前如果在做chat, 或在做對話的時候, 你更多的是我只是單輪的輸出一個,

-

-

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

I think there are more model companies that are doing this kind of one-sided product. I think this is also very interesting because at the beginning, if we look at the beginning half a year ago, or last year, there were a lot of products that were based on this basic model. And then you might do some hand-to-hand prices on it.

or design tools to make the model more usable, and then build a product. Enjoy the power of the model. Yes, and what it's basically doing is to reverse the process of training the model. Because the model training process is also through various, for example, you can think of it as an in-house environment tool. Its controller may train

这样的一个模型,但它可能没有直接开放给你,那你是逆向出来,你更接近去你和它的分布,到底你用什么工具它效果会好,到底用什么样的system farm效果会好,到底用什么样的context engineering它的效果会好,是一个这样逆向的过程。 但是你发现就是说如果模型公司去做一方的产品,它的逻辑是完全不一样的。

You don't need this reverse process. It's more of a positive approach. I can design these tools first. I can design my context engineering methods first. And then I can train this model in this environment. So your model will naturally perform better in your environment. So I think these are two different ideas. But the second one is

它的上限也許會更高, 就是你可以更好的去整合這個工具, 和整合這個模型, 然後你的模型有可能有一些解決不好的地方, 你可以去調整這個工具的設計, 你可以把它設計得更好, 然後同時你又可以端端端的去做訓練, 我覺得這個可能也是在開發方式上,

-

-

It's the ability to rely on each step. But now I think it actually looks like

-

-

-

But the performance on Azure is very high. I think these two things

-

-

No matter if you have new ideas or you do the experiment and get the results of the experiment and then you can use the ideas from the next version or you can optimize a certain infra's performance. This requires a strong agentic ability to do it. So I think innovation is mostly produced

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

现在已经看到很多这样的趋势了,当你有一个agent之后,你就可以把它拓展成一个multi-agent的系统,你可以从一个agent focus出来很多个不同的agent,让它去做不同的事情,然后它很多可以transign,然后up-put的并行,然后再合并起来,然后再分成几个不同的task,你可能有的去写测试,有的去写文档,有的去设计,就是整体的。

-

我覺得 reasoning和agent相當於說 對 就是可能會是 innovation和organization的一個前提吧 我理解 reasoning是 agent的前提 是因為它能夠讓語言通過推理 在agent中得以放話 對 我說如果想去解決一個 可能最複雜的agent的問題 你肯定如果不會推理是很難 但是你假設沒有推理的這種方式 你還是可以一定程度上 做一些agent的任務 所以不需要推理的任務

-

-

Innovation.

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

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

But you really made a strong reasoning. I think there is still a lot of room for improvement today. So how do you see the division of L1 to L5? What kind of degree do you think it is? I think it is a few important technical milestones. But first of all, it is not necessarily a trend-based relationship like we just said. We expect that something will be solved immediately. Then you can solve the following problems. It may be that you will improve at the same time. I think that for example, reasoning is that you have to

you really need to solve some open issues. Or you need to do a good innovation. You need to introduce a new model structure, then your training ability may need to be higher. So I think these abilities may need to be further improved. Let's review. In 2023-2024, your key decision was that you decided to start a business in February and started financing the team.

Then in the second half of the year, Kimmy was released. Then the long text version of BAPT was released. What are your key decisions in the years 2024-2025? Good question. I think it's very important. Maybe technically, you can use pre-training and, for example, SFT as the focus of such a development method to transform into pre-training and strong chemistry as the focus. And I think this actually requires a lot of

foreign

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 101010101010101010101010101010101010101010101010101010101010

or even in the training process, there are some side effects. We found that you can use the reward directly from the end to the end, and you can train this very well. I think this is not very clear in the early stages. And then in this process, we may have accumulated some strong mathematical mechanisms and some know-how in the algorithm.

I think one of the main points is that we hope it is a very good basic model. We hope it is a very good base model. So if you want to have a better base model, we have to look at the bottlenecks of pre-training in the entire field. And we found that because you actually

high quality data is slow to grow. And multi-modal data can't improve the quality of the text. You can think of high quality data as a constant. So what we hope is that we can maximize the use of each piece of data. This is called token efficiency. You want to have as much data as you eat.

you can grow more and more, which means you can get more intelligence. It's different from what we thought before. For example, if you do a lot of performance optimization in the training system, you make it train faster, this thing is also very valuable. But training faster itself does not improve the intelligence of the top line, because your token is still so much. You train faster, you just say that I will complete this training in a shorter time, but its model effect will not necessarily be better.

So this is the optimization of training efficiency or compute efficiency. We are now looking to improve its token efficiency. We are using a data as several parts. We are very concerned about the optimization of Mewon. Mewon's token efficiency is relatively high. You can think of a optimization like Adam that has been used for 10 years.

Most of the models are trained with Adam, but its token efficiency is not good enough. It doesn't think about each element independently. For example, a vector parameter will consider the dependency between them. This way, you can have better learning efficiency. That is, you can get more intelligence by learning the same data. For example, in the early days, if you were in the case of Compute Optimal,

-

-

So the compression rate will increase faster and the loss will decrease faster. Is this original? No optimization? No optimization. It's something that Keller proposed. We have done a lot of optimization on it so that it can be used and trained in large-scale language models. For example,

-

-

-

-

-

-

I think this has some requirements for the way it's being modified. But I don't think it's the best way to modify it today. I think there's still a lot of room for exploration. Going back to what we were talking about earlier, we said that we hope the K2 model will become a good base model. We also hope to improve its torque efficiency.

These are our corresponding designs, including adding more parameters through larger speed. The token efficiency will also be higher, because even if you write more data after adding more parameters, you will also absorb it better. Because you have a bigger

Yes, you can verify that it has better token efficiency through experiments. So this is very important. The second thing is that we hope it has a good agentic capability. For example, it can be used to study and to simulate tools and environments. I think that for an agentic model, the biggest challenge now is actually the model's decoding. Because I think the limitation of the current RL technology is that

不管你的训练任务还是你的评价指标, 它很多时候都是单点的。 比如说你就训SweepBench, 它要提升SweepBench, 然后我觉得它是一个基本上很确定的东西。 但是你的指标提升上去之后, 并不意味着你的模型的泛化会变得更好。 对,然后所以我们也尝试去可能 解决一部分这种泛化的问题吧。 就是我们不希望说它过你合到某一些工具, 或者过你合到某一些环境, 或者过你合到某一些...

- -

-

-

AI can do a lot of alignment research.

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

-

至少在公开可查的资料里面也是比较早去做的。 因为我觉得这个空间会越来越大。 就是当,就很有意思,就是说当你的, 越往上爬的时候,你发现就是你的空间是在变大。 首先它token在变多嘛, 你完成同一个任务的token是在变多的, 所以它的问题的复杂度是变得更复杂。 就像刚刚讲的,就是问题不可避免, 但是问题总可以被解决。 就是说你这个不可避免的问题看起来会,

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

-

So the whole thing, this cycle is actually relatively long. But if you look at it, you only look at the model itself, its training, from the beginning, press the training button to the end, that actually doesn't take that long. But your overall development actually needs to be more pre-planned, do a lot of things, in order to finally ensure that your training is a more smooth process. When does the BAT start? It has to do a lot of accumulation. It's like,

-

So when did you make it?

Why did you choose this?

-

What kind of technology do I need to add? What kind of model do you expect? We are thinking about what the next generation model will look like after K2. This is a question of continuous thinking and decision making. You may see that there are many new things in your toolbox. What are you going to use them for? There is a process like this. Are your research and training teams separated? What is their process?

Because you started to study these techniques a year ago, is this something that a team is doing? Yes, it is a team doing it. Because it is difficult to separate these things. For example, in the actual training process, you will encounter this problem. If you didn't know it before, you couldn't solve it. So in fact, these two things are not separated. Have you encountered any challenges in the process of K2?

-

模型的上限是有害的。 对,然后我们就是说等于是又回过头去revisit, 去重新看这个问题,然后去修复它。 因为这个东西是你在小规模实验上没有办法预测的, 因为小规模上我们没法复现这个结果, 对,它就是不会有这个爆炸的问题。 然后其他的基本还好,因为其他的我们都在小规模上做了很多实验, 然后它基本上也是可以迁移的。

- -

I think this is the main thing in the paper. We still want to share it with the community. I'm curious how you define "agent" and how you classify it. Good question. As I said, it can be a tool that can interact with the world from a simple brain. The most important feature of an agent is that it can be a multi-tool. I think there are two types of agents. One is multi-tool and the other is tool. Multi-tool means you can do it many times.

This is a test time-scanning method. The tool is actually a way to connect the brain to the outside world. For example, if you use a search engine, you can connect the model to the entire Internet. If you can write code, you can use the brain and the digital world. Because all digital world's automation can be described with code. You can let it have this kind of automation ability.

So I think these two are the characteristics of agents in my opinion. There will be more tools in the future, right? But you will have a long-term distribution. If the model is well-designed, it will not only use some common tools, but it may also use very personalized tools. For example, suppose you have a

-

-

-

-

-

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

It's just a good result, not the purpose of designing this system. So designing a plane is to be a means of transportation, not to fly like a bird. Yes, so with the agent system, it's more about general purpose intelligence. You are aligned with this goal, but it's just like humans. How to improve the transparency? Is there any way to do it?

Yes, I think this is a very difficult question. I think today's agent's words have a risk of being limited to some benchmarks. But now there may be fewer good benchmarks. So I think this will be the next challenge. But I think there may be some solutions. I still think if you can train AI with more AI,

-

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

-

But if you only fit those tasks, you only get those tasks. In fact, many users or in some OOD scenarios, they don't have a good feeling about it. I think in this field, the benchmark is not enough or the benchmark is not working. And then the agent's code is not working. Why is the field of mathematical code relatively easy to be solved? Actually, it's not. If you do a strong learning,

-

-

So every time you move forward, your translation will become better. It might be a

It doesn't necessarily have a goal. You will always have better solutions. For agents, the mission and environment are very important. How to set a good mission and environment? Do you have any thoughts on this? Going back to what you said,

-

I noticed one thing. Usually, they think that in task design,

You design a task that is challenging, and this task may come up with a new, fundamental solution. But your KR actually uses some medium-sized tasks. What are your thoughts on this? Will this lead to transparency?

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

What is the relationship between Coding Agent and Tune Agent?

Coding is a more abstract term. It's a paper-based term. If you use two tools well enough, you can do most of the work. In the end, we hope to be able to do more than just coding. Including the model we're training now, we're not saying that we can only do coding because it still has some limitations. Is coding a given environment?

It's a part of the task. But it's probably a very important part. It's a relatively easy task for an agent, right? It's easier to verify, so it's easier to learn. But it still has the same challenge. For example, I just talked about the problem of versatility. I think even a coding agent will face the same challenge. Coding is a very important part. It represents the digital world of digitalization. Suppose you want to create a new tool today.

-

This tool can also be used to implement codes. Codes have a unique role in this. But it's not enough if you have a coding agent. For example, today there are many people who are not programmers who use Cloud Code to do a lot of their tasks. If you are a lawyer or a product manager or a designer, you will also use Cloud Code to do some things. It's because your model is to a certain extent

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

benchmark

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

So I think this is a very important direction. Is long context and long memory important? Long context is also very important. Because now, there are a lot of tasks that you can't solve in a context like 128K or 256K. You may need to go to a million level or even more. But you need to be at this level, your brain is still very easy to use.

-

These two things are naturally conflicted, so you need better architecture. But some of the architecture you may find that the effect will be improved in a longer context. But in a shorter context, it may not necessarily improve or even decrease. So there will be a lot of these problems in the architecture. But these problems may be gradually solved. I think there are some solutions.

So this is also very important. This is probably a support for architecture. And I think the way of training RL now will definitely have a lot of room for improvement. For example, when you want to train a very complex multi-agent system, you may not have enough reward for just using the end-to-end. But how to generate the reward in the middle, whether it can get rid of some artificial design, I think it will also be a very interesting challenge.

-

-

The community can contribute some things. For example, you can do a lot of things in the reasoning side. You can let this model serve you for free. And it can be used by more people. But if you want to make the model better, you can only do it yourself. If you look at the base model, if you use this model to do a lot of post-training,

especially post-training agents, there are new opportunities. For example, if you want to be a legal agent, and you are a startup, you can use GK2. Under your tool collection, you can find a specialized agent who performs well in the situation you are concerned about.

I think there are opportunities like this. But it's more likely to be used to restore more of the lower applications. It's very difficult to turn these things into your main model, the improvement of the base model. This is still very difficult at the moment. I think it's probably like this. I think this problem can be observed in a dynamic way, that is, the relationship between open-source and closed-source. Will you choose to open-source for a long time?

I think this is something we want to do for a long time, but we don't necessarily do it on our own. We hope to share with the community the know-how of technology you just mentioned. And I think this can accelerate our

-

So is this a belief of a technology system or a market game? I think objectively speaking, both.

and it may be beneficial. But ultimately, we hope that we can use this technology to reach a better level of safety and speed. How will the open-end ecosystem be developed? How many will be left in the end? I don't think there will be many. Overall, there will be a few. Overall, if you look at the past two years, I think the trend is still clear. The market will be more focused and focused on retail.

You may have hundreds of them at the beginning, and then tens of them, and then a few. I think a few may be the most stable number. I think it's a probability. Are you on the open side or the closed side? Like I said, we may continue to open, but not everything will be completely open. I think... Choosing new is one, right? We may have to observe this in a dynamic way. But we definitely hope to be able to share more technology for a long time. Why are Chinese companies all open? Well...

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

I think, like I said earlier, there are a few key technical points. For example, how to improve the Asian culture. I think that's a very important issue. And we hope to see if there are any new things in technology. Including the optimization system. I think before we make it, but I don't know if it's true now, but at least before we make it, no one is doing it. So it's a...

It's also a non-competitive model. Because everyone used Adam before, I think this is also a very important one. Because it's really good for token efficiency. So now we can have a better base model.

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

If you are making an agent product, you need to combine the model with the tool and context.

-

-

There are more ideas. For example, the search engine system and the recommended engine system. But they are all adapted to different times. I think the complexity of the system now is that you want to make this model more universal. It will bring a lot of complexity. On the one hand, you can think of it as becoming simpler. On the other hand, you can think of it as becoming more complex. Simply put, you just have to put everything in the same model.

You don't need to maintain so many models, you don't need to do a lot of routing strategies. From a concept point of view, or from some engineering point of view, it's getting easier. But on the other hand, it also becomes more complicated. If you want it to be very common, then you hope that this model can work in all kinds of scenarios.

agent

agents, you have to put them in the same model, then there might be some problems with the fighting. Maybe your tool definition is different.

- Right. -

It requires a lot of steps to complete the task. Even if it's a programmer, it's not just writing code. Or even if it's writing code, it's not just doing 3D benching. So you have to do something that's very common or something that's really usable. As you progress, it will have higher requirements for the common sense. I think its system complexity is mainly reflected in the fact that you need to make it

-

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

-

and agent reasoning chat, all of these things need to be combined. I think there is a challenge like this. And now, you're not just doing SFT, you're doing I/O. When you do I/O, this challenge will probably be further increased. This is a comparison. It's just that you only do... In general, pre-training is relatively easy to do. You just need to put all the scripts together.

-

-

-

It's actually a way of interaction, but there might be more in the future. For example, when you have a multi-agent system, how will it interact? It might change with the boundaries of the ability. For example, when you look at coding, you have Copilot, then Cursor, and then Clock Code. But every generation of interaction will change. And you realize that the interaction is actually changing with the model.

-

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

Why hasn't AI products formed a data wheel yet? Because the scaling of computing is too powerful.

I think this is one aspect. For example, if you scale pre-training and then you scale IELTS, the efficiency of IELTS scaling is much higher than pre-training. Because it's on-policy training with a floating-duty system, so the efficiency of scaling is higher.

So when you have a high scaling efficiency, you can see that the scale of the computer and the scale of the flops is greatly improved. So you will see that the other improvements are very small. This is one aspect. On the other hand, the so-called data wheel relies on the feedback from the external environment. We don't want it to have too much noise. But now, somehow, it's not completely done. This problem is very good. Because your noise

The sound of the large model is more sensitive. It's a little different from the traditional recommend system. It's very sensitive to sound. Now it looks like the scaling of the Flops is more effective. But when will this balance change? It's also possible that you can use new interaction

to reduce the noise of the signal you collect. Yes. But... But you need to create a new kind of interaction. Yes, but you need to adapt to the development of model capabilities. Your interaction cannot exceed the model capabilities. You should design a good interaction within the scope of the model capabilities. But I think this is worth trying. But today, I think you can scale the FLOPS range or improve its learning efficiency.

it's more convincing and more effective. If we say that user data can't provide the smart model, then it seems that there's no need to do 2C products today. We can just focus on improving the smart model. It depends on how you look at it. You may still need to have a certain amount of data. You probably can't use

- Evaluation.

-

What are your thoughts on this product?

I think more people are still thinking about how to make models because I think, like I said, the products that have been trained for models are basically almost done. Model products. Yes, I think it's still in place today. We will continue to do it in this way. Some people will say that Kimmy is going to do... Of course, you don't agree that you are going to do China's OpenAI. But they will see it this way. Are you going to turn China's OpenAI into China's Anthropic? Is there such a transition?

I think it's hard to define it like that. Because China and the US are not the same. And today, we still have to think about this issue from a global perspective. I think China may not be very successful in itself. I think it's actually a little bit simpler.

We hope to continue climbing the mountain. I think climbing the mountain and making time is a good friend. And then, together with the community, we can accelerate the technology's promotion. I think this is what we want to do. Let's talk about the double-edged sword. What do you think? Is API a good business? I think the double-edged sword is clear. One is an API, the other is a product. I think we will try these two. But I think today,

The main priority is to make the model better. I think this is the main goal. But when you make the model better, for example, you are leading in some aspects, then it is true that there is

-

For example, if you can make the model to the level of Opus, or even better, then the space will be bigger. So we will spend more time to make the results better. Some users say they like Kimi, but they are worried that Kimi will not make much money. What should they do? Can you make money? Yes, you should invest first. I think whether you can make money depends on your model. I think the market is a

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

This is the most interesting part. What are the unexpected things about this journey? I think it's okay. We have expected all the difficulties. But what are the specific difficulties? The most interesting thing is that you never know what new technical problems will arise. And how these new technical problems will be solved.

I think this is the most interesting part. Basically, it's hard to predict. If you can predict it, it's not that innovative. But I think this is the most interesting part. You will always encounter new problems. How is your life now? What's your life rhythm? Now, I may be sleeping a little late. How many seconds? It's not the same. It may be different every day. But it's okay. I spend a lot of time to see how to train my model better. So your time is mainly invested in model training?

I think so. The concept of Dam Wing is very abstract. I think it's very important to have a strategy for your technology. I think this is the most important part of the company strategy. Your next step is

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

做一個可能第一個open的agentic的模型。 我覺得這些可能都是一些技術的bat, 或者技術的決策。 我覺得這個東西基本上會至少決定五六層的公司的走向。 但你要做很好的決策, 還是需要很多證據支撐嘛。 就是我們還得做很多實驗, 所以你得了解很多具體的實驗的結果。 就是到底這個東西是什麼樣, 那個怎麼樣, 接下來可能會怎麼樣發展。 這個東西它你不能拍腦袋, 就是還是得知道更多的信息。

I think this takes a lot of time. The second one is the technical aspect. Which of these characters are the most complicated?

I think the most important thing is what you're going to do next and what you're not going to do next. Yeah, it's the same as what you said earlier, which one of the two is the longest time-consuming? Yeah, it's okay. Because I think the key to this is the process of collecting data. You do an experiment and see if it's solid. Then you combine this with a certain technical understanding and then you can judge. A lot of times, as long as you have

-

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

-

If you have 20 experts who are doing domotai, you may not want this to happen. Then you may think that the domotai you learn is a stupid domotai. We hope it's a smart domotai. What are some important milestones in your life? Maybe it's the flexibility. Yes, I think so. I think the flexibility of agents is the most important thing. I think it's a problem that hasn't been solved yet. Have you solved some of them?

-

your I/O task will be more complex. But we don't want this to happen. If you don't observe it, it will be lost. We are trying to find more solutions. I think this is very important. And as I said, we will continue to support long context. We will see how to do it. If it is high in IQ, it can have longer context. I think this is

- -

In some scenarios, the results are not that good. But this problem can be solved at a certain point. Where do model companies and agent products companies see their relationship and boundaries? I don't have a clear answer to this question. I can only say that it seems that the product of the one-sided is good. It can do vertical integration because I can put the model in here for training. So my model and tool are integrated. It's not that I do it separately and then go back to the project.

But because there are so many agent spaces, I feel like a product may not be able to do it. So if you can find some space, for example, the implementation of your tool needs a lot of

I think there's a chance. I think there's a chance.

-

Yes, you can still do vertical integration. But if your tool is something that I can only do for one purpose, and even if I can't do it for others, for example, if I have some online services today, then my tool, my order or transaction tool, is something that no one else can do. Then you may be...

can produce some unique value. Of course, I think there is another possibility, that is, when your first-party product or your general agent, its traffic or business model is mature enough, many of these exclusive, originally monopoly tools, it will also be willing to intervene, because its overall commercialization efficiency will be higher. But I think the improvement of commercialization efficiency takes some time. So within this time window, you may actually have space for exclusive agents.

Yes, but I think the reason why it can work in general is because of its overall commercialization efficiency. So today, even including many content platforms, you may eventually connect the content to this general agent, your commercialization efficiency will be higher than today. This is possible, but it may take a long time. Is Minus your client or your real opponent?

In terms of production? I think it's still early, it's hard to judge what it is. And maybe the product itself will be advanced. So I think in the short term, it may be more of this kind of cooperation that is probably greater than competition. But in the future, it will indeed be advanced. So for example, it's a bit like Cloud and Cursor's relationship. Cursor may also need dynamic change. It may also have to

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

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

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

It's okay, because a lot of things can be predicted. Because I think the ability to predict is more important. Because your virtual model needs to have a cycle and cost. So we hope to predict as much as possible in the early stages. The only thing that was not predicted was the problem of MaxLogic. But that really can't be done because it can't predict on a small scale.

-

- Do you know this? - No, I don't. He said that as the model gets bigger, the ability to talk and the emotions are getting stronger. But the ability to interpret, especially the performance of the data, is slow on the shoulders and then it goes down when it gets bigger. So he thinks that using more modern models, like doing math problems, tends to be not honest. He thinks this is the fundamental defect of Next Topic Prediction. Yes, but it can... So it needs to be paired with the scaling of the talk.

-

So in the end,

- -

- Will there be a GPT-10? - What did I say? The reason why it has a deep impact on you is because when you first saw it, you said it wasn't GPT-4, but GPT-5. You said it was GPT-10. I don't know what you said. - Okay. - It has a deep impact on you. Okay. I think... But I think we have already achieved a lot of things we had planned for today. For example?

For example, you can complete a few hours of tasks. I think this is a fast-growing process. But I think overall, you are gradually seeing more variations and then you break through the process of transformation. I think this is relatively clear. What do you think about the word "evaluation"?

-

There are a few things that are connected together. If you look at our research or the process of creating new knowledge, it's like a process of learning. There's a theory called experientialism, where people get new knowledge from experience. Later on, many people have different opinions. Actually, humans have been on this planet for many, many years.

But maybe 300 or 400 years ago, no one said that the earth was a ball. You've been in this experience, but you don't know what the truth is. So experience can't give you the knowledge directly. Instead, you proposed this guess. You observed some things, but you proposed a guess that I think this ball may be the original. Then I thought of various ways to verify it. Including the neural network that is now training. You may observe that you have tens of thousands of NEC indicators. You may observe that some NEC indicators may not be right.

Then you can't give them new knowledge. You can only say, "I came up with an idea. Why is it like this?" Then I design some experiments to verify it. So this is what we think is better. Of course, all good scientific methods are like this. But these things seem to be in common. Because you find that you manage a team in the same way.

uh

Of course, now you have a lot of good technologies. When you are doing I/O, you actually want to add a part of SFT. Because SFT is a very good, well-known, so-called PTX loss. You don't want this model to fly too far. But you have to control your hand. You can't have too much SFT. If you have too much SFT, your students will lose their subjective mobility.

and then I couldn't innovate. I'm doing some practical work on this point. It seems to have some effects, but the core is to master the balance between SFT and IELTS. SFT is that you directly give it, that is to say, this thing should be done like this, and IELTS is that you give it a reward. If it is done like this, it may be good. It may be more likely to respond to the target.

I think these two things may need some balance. For example, RL is the main one.

Right.

It's a very simple thing. For example, if you just want to make all the benchmarks very high, then you may have a lot of overfitting problems. Because everyone will do their best to make the score, but after you finish it, you find that the model hasn't been made well. There may be such a problem. So the definition of reward is very important. Or the definition of reward also requires you to understand

-

You have to make a reward. Everyone seems to have good results, but in reality, they didn't achieve what you wanted. This is the risk. Using SFT to manage your team is a loss of creativity. So these are the things that need to be balanced. But I'm still learning. I'm not saying I did it perfectly today.

- -

- -

-

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

I don't dare to say it's that deep. You can only say it's in your own story. You keep feeling what kind of person you are and why you want to do this. I think you should keep thinking about these questions. What kind of person are you? Why do you want to do this? I just think it's interesting.

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

I asked Kimi about this. She told me that

He said that this is the magnifier of human civilization. I think it makes sense. It's like the beginning of infinity. From the beginning of the human race, humans have been finding new ways to break through the boundaries of knowledge. But the next step is to rely on AI. Because AI is a huge leverage. For example, if you want to spend decades or even decades to learn a new knowledge, you can learn it overnight.

and then you can make new breakthroughs. AI will become a meta science. I think this is very important. It's the amplifier of human civilization. Is it possible to destroy human civilization? I think at this point,

I think the risk is not that it doesn't exist. But we can do a lot of things. Whether it's a safer team or better social mechanisms. I think these things are... For example, when he can do something, he might create some new work. We need some ways to complete this transition. In the end, the technical improvement will be

-

-

-

Everyone is different. This is what I think. But I think that in terms of creativity, a large part of it is possible to be done by AI. I still enjoy this process, but I have to admit that one day you may have a lot of creative work to do by AI. But I think the latter two are still centered on people. But if he takes away creativity, he will take away productivity. It doesn't matter because people can enjoy the results of production. If we have a good mechanism.

还需要改变社会的机制。 对,但它也是一个缓慢的过程, 它不会是说你一两年做完, 它可能是要一二十年逐渐的去调整。 你会频繁地跟Kimi聊天吗? 当然,就我要测试模型。 你会跟他聊一些很深刻的话题吗? 有时候会。 或者很自我的话题? 有时候会。 有时候会聊什么,除了刚才说的那些? 也还好,就是还有一些是工作时程的问题。 你觉得今天Kimi的成功概率增大了, 还是失败概率增大了?

-

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

调整的能力,就像比如说知识创造也是这样的一个过程,就是你不可能创造出来的知识所有东西都是对的,就是你会发现有的东西它也是错的,对吧,但是它在一定时间内可能是对的。

But it may be wrong within a certain time. But when it is wrong, you have to make adjustments. For example, like Newton did a lot of things. It was the best theory at the time. But it's not perfect. It's wrong. It's completely wrong in some cases. You need some other explanations for this universal force. You need some relative explanations. Explanations through the twist of time and space. I think it's the same.

I think the evolution of an organization or the development of a company is a process of dynamics. Any middle point is right at some time, and some is wrong at some time. This is also what Kimmy told me. Any middle point will become a target of criticism. Anyway, it's similar to what I meant. I feel that you will always have this limitation of time. And then maybe more importantly, how do you go through this process? I think it's about investing in something that may not change. For example, your talent and skills. And then the other thing is to go through this process

-

-

It may require dynamic adjustment. It also depends on your current performance or how strong your PMF is. It may have different time points and different strategies.

- Okay, -

So you know how to make decisions emotionally?

- -

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

AI is moving very fast. You have improved your ability, so your excitement level has increased. So I think it's basically at a relatively stable level. I'll ask you the last few questions. A global food you like. Food? Ramen. Why? It's delicious. A knowledge point that few people know but must know. I don't seem to be good at answering this question. Recommend two books you've read this year. There's a book I've been talking about.

Backpropagation. Transformer. GPT-3.

Innovation.

L4 to L3. What's the latest game? I don't know. I feel like my brain is already confused. I've been talking for a year. I can't figure out the limit of the carbon-dioxide.