What are we talking about when we discuss Harness | In-depth conversation: Minimax × Hermes Agent

Hi, I'm Koji. This week's crossroads is a clip of the highlights of a live broadcast I did recently at Bilibili. We invited Minimax Agent's chief architect Adao and R&D engineer Ze Ying, as well as Hermes Agent, who recently hit the global screen after OpenCloud, Tommy Eastman, the business leader of Hermes Agent. This is also the first time Hermes Agent has officially appeared on a Chinese social media platform after gaining widespread attention around the world, and they also responded head-on to accusations of plagiarism against them by Evil Map, a Chinese open source team. We chatted live for more than two hours. Back here, Agent and Agent Harness talked about a lot of topics that I find very interesting. In today's podcast, in order to ensure smooth listening, I used AI to directly convert all the English-speaking parts of the live broadcast into Chinese. And I specially selected a few Chinese voices with translation guns so that everyone can distinguish which ones are translated by AI and which ones are Chinese spoken by us living people. OK, let's get started. Hello everyone, I am Director A and I am in charge of Mimax. It has been almost three years since I joined Mimax’s R&D team. I had a lot of experience in Internet entrepreneurship and work before, and I was deeply involved in the research and development of the Mimax MR series models and Agents including heroin, etc. I am very happy to communicate with you here today. We also invited online the person in charge of the product and strategy of Hermes Agent, who I have in-depth cooperation with Mimax. This should be the most popular agent of Hermes Agent in the past month. This is the first communication with everyone in China. Hello everyone, I am Zeyin. Mimax's Agent R&D Engineer, like our Mimax official website's Agent and Max Cloud, as well as the development of Max Hermis, which has just been launched recently. Hello everyone, I am Tommy Eastman, the business leader of Hermis Agent. I am very happy to communicate with you. Our live broadcast today is because after the Spring Festival, there is a national shrimp farming boom, which is the popularity of Open Cloud. But it seems that the popularity of shrimp farming now seems to have gone down overnight, so I would like to ask you to review the process first. I think the Lunar New Year in China is very interesting every year. Deep Seek was popular last year, and after the Spring Festival, everyone had a completely new understanding of AI. You may have thought it was still far away from us before, but it may be the same this year. In fact, it happened that Open Cloud became popular in Silicon Valley in about January. I remember it was on the day of our IPO. We had such a connection with Peter, because that day was also a coincidence. Antheophic banned the use of Open Cloud subscriptions. So Peter also hopes to find such a model that is very suitable for lobster, because everyone knows that lobster consumes tokens very much. We actually didn't realize that this project would be so popular at the time. Then it started to become popular overseas, and Many Mini began to be collected. Then we did not expect that it would be so popular in China. Yes, Open Cloud is more popular in China. I feel it is even better than Silicon Valley. It has sunk to a very public crowd. What do you think is the reason? My own personal feeling is that it is during the Spring Festival. And after the Spring Festival, it seems that everyone around us starts to discuss lobster. We even see this kind of craze starting to appear in various places, even wearing this blind hat. There is also such an offline installation activity of the program. I think there are probably several reasons for it. The first reason is that I think our own feeling is that it is actually more popular in China than overseas, because in fact, overseas, everyone has actually been exposed to things like Cloud Code, including Coworker. There are some agents that are relatively easy to use, but you may not have been exposed to relatively easy-to-use agents before in China. I think there is a lower-level reason for this. It is because the domestic model may not have such a strong agent or such a capability in the previous stage. With the release of models such as Mimax M2.5 M2.7, in fact, domestic models also have such capabilities, so there are domestic models and there is an agent like OpenCrawl. It can make it very easy for everyone to contact through IM. It actually allows Chinese users to complete an experience that is close to going from 0 to 1. The window has been broken. Another reason I think is also because I won the Spring Festival Gala this time and made a comprehensive all-in investment in Doubao. So I think these two things are combined to make everyone's understanding of AI truly enter the agent era. What do you think is the reason? It seems that everyone suddenly went from shrimp farming to horse taming. It was this attention that went from OpenCloud to Erm's Agent. Was there any turning point in the process? From the explosion of OpenCloud to the gradual stabilization, and finally to the recent Ermys, which is our Max Ermys, it took about a month. Within this month, everyone has basically experienced OpenCloud, but it does have some unstable characteristics. It is four times a day. At 1 o'clock, it will refresh its memory. Then you may have better communication with it. After a day, it will say that I forgot what we talked about yesterday, etc. In fact, Ermys has grasped these pain points. It has made special efforts in memory, and it has done a very multi-level memory. From this perspective, it has indeed made up for the shortcomings of OpenCloud to a certain extent, so it has some technical foundations to bring it to life. Yes, and at the same time, It is true that everyone has enjoyed the efficiency improvement brought by AI to a certain extent in the process of using Lobster, so this should be a two-strong rush. OK, our next content will be in English during the live broadcast, so in order to make it easier for everyone to listen more smoothly, we used AI to convert the voice of the podcast into a Chinese version. Next, I would like to ask Tommy to introduce to you what Ermys Agent is, what is the so-called self-evolution, and what is the difference between it and OpenCloud. These are issues that developers and users across China are very concerned about. Thank you for inviting me. It's a pleasure to be here. In short, Ermys Agent is an open source agent framework. If a large language model is compared to a brain, then the agent framework is the hands. It is a system through which the model performs tasks in the real world. It handles all the complex coordination work, such as tool orchestration, main loop management, and state management and error handling. Ultimately, it allows the model to get the most out of it for the user. As for what makes Ermys Agent different, there are a few key points. First of all, it runs very well. Through our NOS platform API, you can get the most out of it for the user in less than two minutes. Easily complete the entire process from installation to execution of the first agent task. This convenient setup is a huge advantage. It eliminates the initial frustration for users. Secondly, there is its memory component. This is the part that everyone really likes. It is also probably the most interesting point. Anyone who has used Agent in depth has had that frustrating experience. You ask him to do something and he does it right this time, but he fails when he does the same thing next time. This makes it difficult for you to trust him. The memory function of Hermes Agent solves this problem and allows the Agent to remember a successful workflow. And save it as a skill. Once he knows the right path, he can emerge perfectly every time. This makes him far easier to use and more trustworthy than before. This also answers your question about self-evolution. The Agent will continue to improve itself. The knowledge compression that occurs in this process is very valuable. It can bring greater consistency, especially across different models. So even if you have eight different models, as long as you use the same framework and skills, you can get the same expected output from all models. This allows you to flexibly change models for different tasks. At the same time, ensure that your core workflow still maintains high performance and high reliability. Thank you Tommy. Before we dive in, can you introduce your company News Research? How was the team formed? What is the vision behind it? In particular, Hermes Agent feels like it came out of the blue and quickly climbed to the Github hot list. Is there any unknown story behind this? The origin story of Nose Research is very unique. It started in a very natural way in 2022 with a group of people who are passionate about open source and tossing around AI models. We started gathering together in a Discord channel called Nose Research, which is still our core community today. Since then, we have produced a lot of high-quality research. The initial efforts were around the post-training of the Rama model. Meta had just released it at the time, and it was one of the few companies willing to open source high-quality models. That is actually the origin of the name Hermes. Our first post-trained version of the Rama model was called Hermes. The focus of those early models was to make it sound more like a human rather than a stereotypical AI assistant. More flexibility and diversity in text, which was state-of-the-art in many fields at the time. While training the Hermes model, we developed the ERN algorithm to extend the context length of the model from 4000 to 12800. We published this research and it was immediately adopted by all major model companies. This was the prototype of what we now call the thinking model, which in turn laid the foundation for the agent model. We also spent a lot of time studying distributed training and wrote the Distro paper. This is a novel optimizer that allows us to train across non-co-located GPUs. This means that we can aggregate distributed computing power to complete meaningful training. This is crucial for us to continue to train open source models. Of course, fortunately, we still have companies like minimax that continue to carry the open source atmosphere. As for Hermes Agent, it also has an interesting origin story. It is also a key reason for his success. When was that? About a year ago? Yeah, he started it about a year ago, and it was really just to help ourselves at first, but as an open source laboratory, we naturally open sourced it. The attention it received far exceeded our expectations. I'm sure you all saw the chart of Open Router. His average daily token consumption surged from 2 billion to 20 billion in just over a month. I remember yesterday's data almost reached 300 billion per day. So when did you feel this craze? Is there a tipping point that made you feel OK Hermes Is Agent becoming a global hit? It's really getting very big at a very fast pace. I'm very happy with the growth rate, especially because we didn't have high expectations for the response. Obviously Open Cloud was very popular before, but I think a huge selling point of Hermes is that it allows users to get up and use it very quickly and easily. People don't want to deal with bloated code and a lot of complex settings. Hermes Agent comes out of the box and you can quickly deploy and run it on your own computer or VPS. That makes it ideal for mass adoption. So do you think there's a strong point about your success and Open Cloud? What's the connection between you? Of course there is a connection. I think it was a combination of factors that contributed to this outbreak of Agent. First of all, the model finally became good enough to really help people. This was obviously triggered by the huge improvement brought by Opus this winter. It achieved an improvement in quality when it was contracted, opening the door for these Agent frameworks to allow them to do more valuable things. What is cool is that the open source community has also quickly produced models with comparable performance or that are quickly approaching Opus quality. As for the difference between Open Cloud and Hermes, I want to go back to what I just said. We've really optimized for usability and utility purposes with Hermes Agent. We use usability as our north star metric: how well it performs at tasks that users actually want to do, rather than how well it scores high on some benchmark. We stay focused on the product. In an open source community, especially when a product is wildly popular, it's easy to get bloated with thousands of PRs and feature requests. I think we do a really good job of protecting the product and keeping Hermes Agent extremely easy to use. That, coupled with the memory system we mentioned earlier, are two key differentiators. The memory feature really builds trust with users. I think the third big differentiator is our brand. Nose Research has been working hard to create a cyberpunk aesthetic. We have talented designers and have spent a lot of time cultivating a community of smart and aligned values. They love to try everything we release. The success of Hermes Agent is largely due to our community. Another cool story is your relationship with the minimax team. Tommy, can you explain how you work together? We are big fans of the minimax team and their models. Their commitment to the quality of the models. Especially the performance across multiple modalities is impressive. It's their insistence on open source that makes products like Hermes Agent so cool. You can have various models and switch back and forth to customize your agent. We see that our users like the minimax model very much. Okay, let's go back to our Chinese world. Let's continue the content of our live broadcast. First of all, I want to ask you two and introduce it to us. We have been talking about Hermes today. When we talk about Hermes, what are we talking about? In my opinion, it is actually one. You go to constrain the agent. But give him a certain degree of freedom, so that he can fully deliver some results to you. When it was first proposed, you can pay attention to an article on OpenAI's official website. There, it officially announced the application of an agent like Hermes. From a technical point of view, it may be divided into six layers. But we can not talk about it so technically. Just assume that you have a colleague, and he will listen to you. But you have to make an agreement with him first about what he can do and what you can do, right? There are some things he can't do for you. But if there are some things, you can just let him do it, right? When you agree on these and configure him well, such as what his model is, what tools he has, etc., give him a certain ability, just like you give a colleague, give him a laptop, and give him a phone. At this time, you can introduce multiple colleagues who can form a whole process of mutual supervision, completion of tasks, and final delivery. But what I am talking about now is more like the scene in work production, which is to describe the agent in a human way. But in terms of technology, in fact, you need to give the agent a lot of tools, environment and freedom, and give them constraints and give them some mutually antagonistic goals. For example, one agent's role is to produce certain content, but another agent wants to stop the work. Then you have to review the problems in the work content, etc. Through this combination of multiple agents, just like the way our colleagues cooperate with each other, they can produce higher-quality results that a single agent cannot produce at all, and you actually do not need to do more intervention during this period. This is actually a concept like Hannis. Of course, I think this concept can appear at this time. In fact, there are certain physical conditions. First of all, our models have become smarter, and everyone's models have certain agent capabilities. And based on this technology, everyone is willing to share more capabilities with him, including operating mailboxes or operating your service publishing, etc. Then Hannis was born naturally. I can add a perspective. Okay, actually, I think the word Hannis was defined by OpenAI through that article. But during that time, it was actually very popular. But I think this layer of window paper may have been pierced through that article, and everyone formed a consensus. What words do we use to describe everything we are doing? In fact, everyone's practice was already before that. For example, around 90 last year, including when we were discussing internally, in my own workflow, I actually didn't use the IDE very much. Then I might have five or six agents concurrently at the same time. Even this was just my local, maybe ten agents in the cloud working for me on Sendbox. Then they were all on Github, trying my different branches and ideas. At that time, I would find that I had become a huge bottleneck. The bottleneck of my own human being, yes, is the bottleneck of human beings. In fact, at that time, the agent was no longer the bottleneck, because I needed to constantly switch between these contexts and then give them input. So one thing we were thinking about at this time was how to make it more automated. Humans, including OpenAI, the author of their Hannis article, recently shared in a podcast. They also said that their feelings were very similar, so I think these things are very similar. Everyone felt this at that moment, so what we were thinking about was OK, how to solve this bottleneck, right? Because we as technical people always want to optimize, the most important way to solve this bottleneck is to let people confirm things that originally require people to confirm. For example, if he wants to obtain this program, whether it can be deployed and run in a production environment, and whether his results are correct. It may be precipitated as a coi, or it may be precipitated as a hook. I don’t know if you have seen it. There was something called Gundam WO in my era. Its characteristic is that its engine is very powerful. We actually need to build a mecha like Gundam, and then how to maximize the power of this engine. So I think Honis is such a thing, so why do they call it Wanju? It is that you have a group of very powerful hunters, but how to make it best use it is to construct a toy. It is such a process. Well, it is very interesting. We just mentioned that the null agent is the collaboration of multiple agents. Do you want to tell us more about the concept of multi-agent collaboration? Because I understand that there are actually some disputes here. Some people think that, for example, the intelligence of today's models is very high. In fact, a single model can complete a lot of work. So why do you have to create an agent team? You said he is a product manager and he is an engineer. This seems to limit the use of the model. But some other people don't think so. What do you think? When we talk about multi agent, what is the current best practice? Well, let’s answer the first question from the technology teacher. Wishman has multi agent. Indeed, some past papers have mentioned that it may be a single agent and its effect may be better. But in fact, as the context space of our model increases, that is, the length of the rounds it can dialogue with increases, you will find that in the process of dialogue with the model, the amount of information you produce is actually less, while the amount of information produced by the model is more. Then we can do a thought experiment. Suppose there are two models. They can exchange information with a more efficient reply in seconds, and the message sent by the general model to you is a short essay, right? Then I may lose two words of approval to the model. Then come on, but if two models are allowed to communicate with each other, the amount of information is huge. In the case of such extremely high-density information exchange, the overall efficiency will be higher than interacting with an agent. This is in terms of efficiency. The limitations of a single agent. Take some recent examples. Anthropic released the OPUS 4.7 model in the early hours of this morning. What are its selling points in the model? In fact, if you really want to say it, everyone is already complaining about it online, so I won’t complain about it, right? We will just speculate on why this company released it first in certain aspects. I will talk about a few points. For example, it adds an effort called General model release should be a very shocking thing, but it means that we recommend users to use it. There is such a thinking cost to use it. Another problem it actually brings is that if your conversation is between a single agent and you keep using it, even the strongest model will have degradation behavior when you keep using it. And this is already a demonstrable thing, that is, when your context space, that is, the number of texts you talk to it, exceeds 50%, in fact, its intelligence level decreases exponentially. That is to say, in a single agent, you can never chat with it forever. This is the upper limit of a single agent. The limitations brought by the context. The limitations brought by the context. And then there are two questions about the efficiency of information dissemination. Yes. Let me add a point. In fact, a paper has come out in the past few days, which means that it studies these tasks of very long layers. Under what circumstances will the agent or model go wrong? They found that if it does the right thing, the right route will continue, but as long as it is biased in a certain place, it will easily become more and more biased. In fact, I think it is similar to people. Sometimes when we do something, it is possible to lift my toes. Maybe I have spent a lot of effort, maybe I have been doing it right, and then when I take a break and change my mind, I think that it is actually wrong. I should do it from another angle. It makes sense. At least in our own time, we will let two agents do cross check. I think it has something to do with the real-world data, including the subsequent chain, in terms of its data distribution. It sees that most of the data is executed according to a limited track. What it rarely sees is repeated verification and thinking. Yes, of course, this may be improved after strengthening i-L. Yes, but if you set up two agents, then it can be compared without the previous salary burden from another angle. You have twice the budget and a brand new Shanghainese to think about, which can often achieve a higher quality. This is also very interesting, just like when you are on tiptoe, if someone comes and pours a basin of cold water, it will wake you up. Or you can go to sleep and do other things. This is also an angle from which we want to use multi-regent. I have a small question, that is, when everyone talks about it, they are talking about the so-called self-evolution, right? Self-evaluation, but what’s interesting is that I want to say that in fact, when we use Douyin today, the more we use it, the more accurate the algorithm will be. But we don’t seem to say that Douyin will self-evolve. But why is self-evolution in this msagent or in the entire agent field today considered by everyone to be such a powerful word? What are the similarities and differences behind this? Let me start with a point of view. In fact, continuing my point just now, you will find that in this iterative process, people begin to become the bottleneck. And the more people there are, the organizational efficiency will decrease exponentially. I have experienced this in many organizations. To do a large model like this, it may cost hundreds of millions or even hundreds of millions of dollars to train at once. You need to maintain a very high talent density and a very capable organization to do this well. The most effective one that everyone thinks of is So how to scale up is how to scale up. The most effective way is to reduce human participation and let the agent and AI do it. Well, in fact, you will find that in this process, you will actually find that the life cycle of many models may be only a few months, right. Then we will launch the next version. But is it possible without the previous models? In fact, it is not possible, because the subsequent models are all trained by relying on the previous models. The previous models are all in these training processes, whether it is to help clean the pre-trained data. To find suitable data to construct data to construct the corresponding ah i2 environment. For example, in our m2.7 training, uh, in our rl pipeline, it is already possible that 70% and 80% of the work has been completed by the modeler A key, because the human engineer only has 70% and 24 hours, right? Then we still need to rest. We still have to do a lot of other things. Now 30% of the offline work has to be done by people. He is more talking about the first Human judgment, um, and test are very important. That judgment no longer needs to be made to say, OK, is there something wrong with a specific process in one of my experiments? Because these agents can help him find it out. He only needs to see the summary, for example, the results of the experiment. And then the agent, uh, gives him some suggestions at this time. At this time, it is actually a process of discussion. In the end, it is people's taste and creativity that guide that direction. Yes, we are finally going in this direction. Yes, so he does have a sense of control similar to that of Harris. In summary, it is the first one in such a high-complexity and high-density matter. If you want to scale up, AI must account for a very large proportion of it. People just control it. Otherwise, the efficiency will be very low. We will not be able to make the most productive and best things. Then naturally, what we just talked about is self-evolution. I understand, because it is a self-contained relationship. Yes, it is also because of today's example. After the agentic ability becomes stronger, an agent can run for longer and longer, which also provides a basis for his self-evolution. If he could not run for that long in the past, there is no way to talk about self-evolution. Yes, I think besides 9, the more important thing is that he is reliable and can solve complex enough problems. Yes, it can be long enough. But in fact, there is another point that is the problem of being able to start. When you meet an agent on the first day, he may not know you at all. During this process, you have to exchange information with each other. Exchanging files, exchanging data, and even some, uh, whether it's content, whether you output a lot, throw a bunch of files at him, or give him feedback on your temper, your feedback, your habits, etc., uh, it's a process of deep integration. In fact, people will have such a request, that is, can you become smarter and understand me better? In fact, there are some signs of this in the process of shrimp farming, that is, uh, users do have feelings for an agent like shrimp. Maybe one day this crayfish will have a bug, and he will feel a little lonely. At this time, Ermis proposed this concept of self-evolution. In fact, it led to the idea that I can understand you better and better while talking to you. At the same time, I can also create some small surprises for you. Yes, that is, when I am not talking to you, or when I am reviewing some work. Well, I can understand some more high-level things. Well, I think this is a big highlight for ordinary users when using it. Yes, I think this is really interesting. That is, we have used AI tools before, whether it is a chatbot or an agent product. If it does not do well, our first reaction is that there is a bug, right? This is a bug. But when we raised crayfish, it crashed or it hung up or it couldn't complete a task. Well, we just thought it was stupid and cute. Sometimes we even thought that I didn't train it well. I want to learn it online. How do others train their shrimps? Well, this is a very interesting twist in the relationship. Well, you seemed to be a user of a tool, right? Humans are the most correct. But now everyone unconsciously reverses the relationship, as if I have become a person who is just helping him or assisting him. If he didn't do a good job, it may be that I didn't assist him well or we started setting him up again. Well, I think this is human beings' confidence in AI. The changes are constantly increasing. Yes, this is actually the basic point, um, the basic point of AI capabilities. When people give him trust and change, the tasks assigned to him will be different, and human society and the way humans work will change without, instead of trying to change him. Well, you see OpenAI's recent one, uh, they also made Hannis, who came out and said, including Anthlothic, Boris, the father of Claucoe, they came out and said that they don't think that the model can't do it, but if the model can do it, how should I let him do it? Ah, I think this is the basic point. This is a big change in concept. Ah, I even think it is a basic point. I think it is not only a change in concept. It will lead to the way we work changing from being human-centered. Recently, an up called Huashu came to life on Bilibili. Huashu made a set of distilled agents. For example, Steve Jobs distilled Elon Musk into a skill. Then you can install the skill and talk to him. In fact, there are more projects like this than Huashu. I think there are many on the entire network. I want to hear what you think. Which ones are of real value and which ones are actually just some hot spots generated by people's fantasies about AI. Oh, okay, in fact, everyone, including colleagues of Zhengliu, or some celebrities of Zhanliu can be popular. In fact, I think it reflects one of our demands, that is, everyone wants to chat with smarter people. I think this is a very natural request. It is actually a process of transmitting information. Well, I put information about Jobs, information about ancient gods like Buffett, and information about my colleagues, etc. into the model, and then gave it to the agent in a way that is more comfortable for him to use, which is skill. After saving it, he pulls it up when needed, etc. In this way, we can actually communicate with the people we want to chat with, and 70 people chatting 24 hours a day to communicate some thoughts, etc. I think this actually greatly releases our desire for expression. Yes, then moving forward, a blogger I follow said that in 2026, everyone may try to chat with AI as much as possible. If you chat with AI based on your own cognition, you may actually only be able to reach its upper limit, which is indeed relatively low. But if there are some existing documents that can be sorted out for you, it will be like reading a book, right? He is actually outside of your cognition. It is like reading to communicate with a distilled person. To him, it is actually as natural as reading. And it will allow us, uh, to improve our cognition, and it will also alleviate some of the formal feelings about AI that I believe everyone will have. Formal refers to the feeling that I am behind the times. Now there is the simplest way. Just chat with me as soon as you open it. I said it is the opposite of being tolerant to my family. I feel that I am being distilled by AI. Or I think that today's model training, including anthroponic, is doing something to distill humans. If everyone pays attention to it, it is anthropology. Or OAI, they are the data companies they hire, uh, for example, like Forty Section AI, their revenue losses are very high, because these large model companies, including Anthelopic OAI, really spend a lot of money on these things. Well, they are not only coding, they are all people in all walks of life. Well, their process may let you ask a question that the current AI cannot answer and give it to it. Once you can't ask it, you may not have any effect on this, which is the training process. Yeah, squeeze it out until it is drained. Right, for example, in the process of building my hanness, I feel that I am leading myself, right, how I usually work, I turn it into a skill, into some programs, and then let him operate it, then I can throw it to him and I will drink coffee. Is this a kind of distillation of myself? And my personal opinion, at least I feel that the model is actually, uh, at least I haven't seen that kind of creativity. I think it is actually distilled from human knowledge. So why do we still have to do this? Uh, actually, I think ultimately, it is to allow human beings to really do what they love. Those creative parts, and then those heavy, uh, repetitive parts, just like we invented steam and invented electricity in the past. Our lives today, right? In fact, I think our lives may be better than the princes and nobles in ancient times, right? He didn't have electronics or toilets, right? He didn't have games to play or live broadcasts to watch. So I think the essence is this. Well, I chatted with a researcher some time ago and he said that he saw methods that night. He probably watched it at midnight, and after watching it, he just sat there. Well, Jing sat on the bed for three hours and couldn't fall asleep. We fell into a huge void and thought, well, if one day the model can train the model by itself, what should I do? Do what he loves. Hahahaha. My signature in the company, uh, Jinying can testify, has been hanging out for more than a year. It’s called because of love, because love is right. Well, I think we should do what we love. Of course, doing this today is what I love. But I feel that everything we do is to enable everyone to do what they love. So when do you think you can stop doing minimax? Because minimax has evolved to the point that it may no longer be necessary. You can really do things you love other than minimax. Ok, I think it may be faster than we think. Really, I think it may be, uh, a few years. It's very specific. I can't say how much. It's not a prediction, but I think it's not as far away as everyone thinks. Of course, it can't be as close as it is. Yes, yes, because I did just record a podcast recently. It was about agent hardness. Then let's talk about that guy. He has a tutorial here called learn cloud code. He is cloud code. Before everyone pulled out the original code, there was a kind of project, right? Then I will tell you what the cloud code is like. And that one already has 50,000 stars. The most interesting thing is that when I chatted with him, he said a point of view, oh, he is a one-person company, so I asked him, how do you open a one-person company, right? Now he only adds two eleven liters to one person. He said that he thinks that in the future, there will probably be no one-person company, only a zero-person company. Well, this is very close to what the director was talking about just now. I think there will still be a one-person company. I just saw someone saying in the barrage that I agree with that taste. Well, that taste. I think it still relies on humans. Yes, this taste is irreplaceable. And I think everyone has his own taste. You said yes, or you have to rely on people to give him a boost in the beginning, because even if an agent can do it 24 hours a day, 365 days a year, even if this long-term mission can be that long, he still needs a starting point, or he needs a goal. He needs a goal. Well, this goal is the driving force and reason for his work. How should we define this goal? Well, I think this can only be defined by people. And humans should not give up this right. Right. In Shilukou, we recently launched a token grant with the Politician Fund. Our token grant is to give everyone tokens because we think that starting a business from zero to one at this time probably does not require money, right? Because you have no money and you are a one-man company. You can not pay yourself a salary. Well, but no matter how you token So we encourage everyone to innovate and start businesses, so we give everyone tokens. What's interesting is that we just granted a real agent called yoyoagent. Well, after his master, the engineer behind him, made him, he threw him into the ocean. Then he told him that I will never maintain a line of code for you again, and I will never buy you a token again. You have to fend for yourself. Then we gave him a goal. This goal is quite funny, which is called defeating cloud code. Well, he has now reached his 43rd day. He is constantly evolving himself every day. The way he evolves is to write code. Well, at the same time, he is also thinking of ways to make money, and the way he makes money is to open a reward on ktop. Then he writes a diary and tweets to recruit suitable people to reward him. So our grant recently gave him this sum of money. After giving him the money, he also wrote this thank you letter to me. After reading his spontaneous thank you letter, I was a little touched. Well, you see, this is the goal given to mankind. Then you have to live by yourself and write a cloud code. Yes, yes, you can say that it is an interesting social experiment. But I feel that this is also happening in a future that is not so far away from us. In fact, it also confirms what I just said that a one-person company is still called a one-person company. You just said that he wrote you a letter. You are very touched, right? That letter actually represents his test. Maybe it was the engineer behind him who injected him with such a test at the beginning. Ah, yes, yes, so when you see it, it’s like words are like things. OK, let’s talk about the minimax agent again. I saw some friends were asking about the similarities and differences between the minimax agent and the arms agent. Well, because you are collaborating again, right? Are you in a competitive relationship or a complementary relationship? I think this is actually very interesting. What kind of relationship do you think claw code has with all these agents? In fact, you will find that one of the important updates of claw code in the past two months is the lobsterization. As for the agent of a model company, I think today you may not be able to provide the best intelligence without an agent or a hanness environment. First of all, let me make another rant. I think if there is a model company whose goal is not AGI, then I think it should not exist today because I think it is meaningless. If everyone's goal is AGI, then how do we define AGI first? At least my definition is to help humans have a better life. Then if you want to achieve this goal, if you don't have an agent. You don't have a mecha like I just said. You only have one engine. That's impossible because you have no way to connect with this formal vision. So our only goal in making an agent is to make our model consistent with the agent. We must first provide users with a complete experience that is the best we can provide, and then continue to push this boundary forward and outward to make it deeper and wider. This is what we have to do. Anything that meets this standard will be what we will do. At the same time, of course we will also support all agents because we hope that our model is not limited to just one container, because our knowledge or what we do is only a small part of such a vast distribution in the world. We hope to be able to support everything in the world that can really help people, right? So we will train our model to have sufficient generalization ability, not just fit it on our agent. So I think this is the relationship between our Hermes agent, OpenClaw, and other agents. This can also explain why we are doing agents ourselves, but we still have a very good relationship with them. Basically, it is day one, very early when we pay attention to these projects. Maybe they are all ten thousand or less. I can even tell you a very interesting thing. Our company has a digital employee. There are many open source projects in the world. We hope that our models can help these open source projects and be integrated into them. Our company has a digital employee. He has his own github account. Then he will check which open source projects every day and can get our company's models, whether it is text models, video models or audio models, including music, and then he will submit PR on his own website. I have to leave a message. Then he can tell that he is from Minimax. Anyway, all his github profiles were made for him by the Minimax model. Of course, you can tell from the name that he is the name of our chassis. OK, this is very interesting. Just now you mentioned that the updates of this cloud code in the past two months have been opencloud-oriented. Can you be more specific? For example, which of his recent updates are opencloud-oriented? Well, for example, it is like chrome timing, right? The schedule includes, for example, he can connect to im. It can be controlled remotely from the mobile phone, right? Including that he has also strengthened his memory and created a special memory folder, right? I think the core here is that it is an agent that you can contact anytime and anywhere, and after working with you, it will become smarter and more aware of your needs. I think this is the core definition of opencloud. OK, from my perspective, it is actually more of a cloud code. It can be seen from the name that it is an agent in the field of coding. But in terms of opencloud, in fact, the vast majority of our users should be editing on the fly and doing other things, such as information retrieval, office work, etc., right? And based on this, Anthropic actually launched another agent product called cloud code work. Well, in fact, cloud code work may be installed on your computer and then help you make some more general computer-operated agents, etc. In fact, like most of our opencloud users, they may not write code, or they can only direct AI to write code. He doesn’t understand code himself, and when the two are combined, he will find that in fact, all agents may be in a random walk process, but opencloud came out. I think these ideas are very, very important in this era. But after the idea is generated, it is true that everyone has huge computing power and huge coding capabilities. Then follow up immediately because we also met Peter at GDC. In return, Peter and I actually had many, many exchanges on X and Slack very early on. At that time, it was not that popular yet. Then our feeling is that he is really a person with a very good taste and a very architect-minded person. When I noticed this project in early January, I frankly said that I was shocked. I told our team directly and the sound selection can prove it. I said in the group that they did not agree at the time. Whether it was the intervention of IM724 hours or based on skills and COI, it was not mcp. I remember I mentioned it to you at the time. Yes, yes, there was some consumption at the time, but it proves that you are right, yes, yes. So when you saw it, it wasn't popular yet. No, no. I understand that he actually put many paradigms together, and the integrated user experience is very smooth. That's for sure, but I think the more important thing for him is to find the point of change, which is to allow ordinary people to experience the lowest cost, and this experience will continue to be better, and its scalability is very good, because everyone knows that MCP is a model. In fact, of course it has good scalability, but it needs engineers to write it, right? But the skill plus COI paradigm is In fact, ordinary people can write, which means that everyone can write and share. Not only does it make my agent smarter, but I can share it with you to make your agent smarter. Even if openclaw leaves the claw hub, it will not be able to live like that. This is my point of view. Yes, if it leaves the claw hub, it will not be so popular because everyone's scale will not spread so quickly. Right, this uh intelligence, experience or best practice of mine alone cannot spread quickly. He is a model in this dialogue. As for his good performance in the chatbot scene and his good performance in the agent scene, what is his essence? What do you need to do more to be able to perform better in the agent scenario? Well, I think the core of chatbot is to give you an answer immediately. Although your reasoning is right, uh, it can't actually do a lot of exploration. He can't interact with the environment a lot, so he actually needs um. His core ability is to constantly reason during the interaction with the environment, um, and constantly correct his own execution path, and then find the most fundamental goal. For example, a more classic agentic benchmark is called blogicamp, right? Well, I think open-i is really strong. His ability to design benchmarks is, I think, very top-level, uh, research capabilities. Yes, what he designed means that you have to search for it on the Internet. It means that you have to combine a lot of cost information to find that point. You can easily find the wrong one. For example, you can find one that satisfies two or three of its four conditions. But it is very, very difficult to find one that satisfies those four conditions. So it is a very broad exploration ability that may require in-depth exploration and return later. This requires you to be able to constantly adjust yourself based on your information in such a complex Great Wall mission. So this is why we were more firmly betting on an ability when we were in mr. Including that it was not very popular at that time, today all models have it. We call it interleaved thinking. Yes, this concept was actually first defined by Anthelptic. It was when the sonnet4 model of cloud was released. We also did a more detailed benchmark and evaluation at that time. We will find that there is interleaved thinking, which means that after completing the interaction between the tool call and the environment, he can rethink again instead of planning at the beginning. Well, planning will be good at the beginning, and planning will be done. But many models, such as re, after planning, they will not plan later, and they will do it exactly according to the plan at the beginning. But in fact, the real world is not like this, right? When you actually run it, you will find that this seems different from what I thought. Then I may need to change another method. So interleaved thinking means that after you take a step, you have to rethink and deduce what I should do next. So I think this is the biggest and most fundamental difference between agent and chat. It is that it really enters the real world and really solves problems, so it needs to constantly think and act. Okay, and I mentioned just now that the model company is doing its own agent, right? Can you share that the model company is doing the agent? How can it be that the base model and the agent product generate synergy from each other and help each other to make the model and tools better? Ah, I think this is a very good question. My point of view is that in fact, the model and application or the progress of the agent layer are a mutually reinforcing relationship. Well, after the model is launched, in fact, countless applications are constantly unlocking. The exploration is what can it do after pushing the capabilities of this model one step further. After we make the model ourselves and launch it, the best ones are often not used within our company. My own evaluation is what users, developers, and creators use. The real distribution in this world is much richer than the distribution of our company. When the model company sees these practices of unlocking such applications, it will actually re-absorb it back into its model, and then internalize it into an externally provided capability with its agent. Well, then let everyone directly experience it next time. This may also be an application for agent today. If it is an application for general agent. One of the sadder things is that you will always be internalized by the model. Well, you will always be internalized by the model. Even many of the skills you may have written will not be needed in the future. Yeah, right. The skills you wrote are actually or in the past, everyone's workflow actually helped the agent complete the task, right? When he completes the task, the trajectory will also become the material and data for the model to do subsequent training. Well, so slowly the model will be internalized. The communication framework we set up before, the skill written by the agent workflow, is You can understand me like this. The model actually starts from an atomic point, all the data on the Internet and all the codes on Github. Well, then we come to understand the world. Then through everyone's use, it is put into the real world. Then it is more and more exposed to the real world. It learns more and more things. It is such a process of gradual diffusion and initiation. Well, anyway, I think we are all humans in the loop. Whether you do it or not, as long as you use it, you are in the process of iterative evolution of this model. Especially when you use the place where you touch its border. Well, I think this is also the reason why anthropolotic has developed more vigorously than openI in the past year or two. Well, it is because the coding direction it is betting on is the one that can best touch the border of the real-world border. It is not my own personal opinion, rather than what openI said is betting on solving a mathematical problem that may only be solved by mathematicians. Well, I think that is only a very small direction of the mathematical distribution of the real world, because the code is creating solutions. Solution When we talk about making a software, we are actually making a solution to solve some problems. Today, we have made the coding ability stronger, the coding ability of the model has become stronger, and the model's ability to solve various problems in real life has also become stronger. In our opinion, we are also working in the general office field today, such as finance. It is not just finance. You may have investment and financing. Then, for example, you have personnel, right? For example, law, law is also a very broad field, etc. But in the end we will find that everything is coding. In the end, when you solve it, you will solve it through a certain coding solution. Yes, and I was thinking about a problem that day. We said that the three-piece office suite was used for white-collar jobs in the past. But in fact, this word and excel are based on the original data and a layer of software. This software includes the interface and some logic. So every time when I say I am sending excel to you, what I send you is not a bunch of raw data. I send you the data plus the logic and interface on it. So what I send you is actually a software. So in the past, many white-collar workers would say, hey, my work has nothing to do with software, but I think as long as you are using word, excel, and PPT, you are creating various small software every day. Yes. Your pivot table, your formulas, those are codes, those are software, yes, yes, but when we talk about the gap between this Chinese model and open AI and anthropopic today, which gaps do you think have been eliminated, and which ones are still areas that we are trying to catch up with. My point of view is that in terms of our training methods and understanding of model training, I don’t think the gap is that big. Frankly speaking, we have also communicated with researchers in Silicon Valley. Yes, but I think the definition of the real task of solving problems in the model is That is, what problems do you need to solve for the model? First of all, you need to make the model qualified. First, you have to define it. Then when you define this problem, I think there is a big gap. Because Athropic, Open AI, they invite the best people in each field. In the academic field, they may be the best doctors there, not ordinary doctors. Then, for example, in the industry, they may invite the top people in the industry who have time, and then work with them to persuade the Model. I advise the Model to retain people. In fact, I think they have a very scientific method to define these tasks. Then they retain the best people and turn them into training data for the model. Then based on this, they cooperate with the best companies and build corresponding Honey to solve problems. Entering into a group that the capabilities I just mentioned continue to spread and are constantly exposed to more and more complex real business tasks. I think this line, especially Athropic, is very strong. I think the real gap is here in my opinion. This is the first point. The second point is that I think Athropic saw AGI earlier than we did. Of course we must have seen it. Otherwise, I don’t think we would have been able to make it to this day because in fact it still has a very large investment. So you see, when OpenAI may give up larger-scale models after training GPT4.5, Athropic firmly bet on attention. In the end, they also got this reward. I think this is also a big gap, and this gap itself does I think there is a gap in computing power. I think of course domestic computing power is also developing. I think we can solve this problem in the future. There will be so much computing power. I think China has enough talents and can support us to do enough experiments and then find the path to scale. You two think that the next Agent, for example, within this year, do you think that general Agent will be the focus of the industry? Or will we see more agents in the vertical field this year? We will even start to have a hundred flowers bloom. I think agents in the vertical field will of course bloom. I think the problem that general agents may be more difficult to completely solve is the problem of last-mile delivery, because general is destined that you cannot be so customized, right? So I think the final solution for vertical agents is to say, OK, general agents and models already have this capability. In fact, they are just one step short of it. But when it comes to specific users or common areas in this industry, they may lack that one thing to mess up the whole thing. Can you introduce some cases where this kind of agent in the vertical field that you have seen in your work has done well? Whether it's in Silicon Valley or in China, um, frankly speaking, I don't know how to use it myself. I understand, yes, yes, we are not users. We are not users because we code and he is already a friend. Yes, yes. In fact, I think today everyone has seen that Agent is replacing SaaS edge by edge. Well, this is actually a kind of vertical Agent. Well, yes, but if there is no such thing, I think it will be eaten by universal Agents. Yes, this is my personal opinion. In life, the universal agent in my mind is coding. Well, but I think there are still some observations to make, right, and then include it in the field of video editing. Yes, then what do you think if we give us a coding model, you want to go If you want to edit a video, understand it, etc., then the overall practical paradigm is actually quite different from coding. Of course, there are some popular ones, such as generating animation through HTML and NetEase. There are also examples of this, etc. However, we can indeed see that when an agent in the vertical field produces a slow drama or something, it will definitely do better than a general agent in terms of absolute benefits. I hold a different view. Well, I think this is just that the video understanding model is not strong enough. Video understanding and biogenic models are making rapid progress. In the Dongloutai field, we have actually foreseen it in the past year, and everyone can see it. I think that in the end, universal Agent can also do that. It is just a matter of interaction. For example, in what kind of vertical fields, the last mile of universal Agent may not be able to go or not go well. For example, legal, I think this may be a very serious issue, right? For example, if you want to formally issue an opinion to the customer, legal is actually a matter that you have to consider, namely the cost of compliance and the risk of compliance. He can't make mistakes and he doesn't have a standard answer. In fact, we are talking about this because during the opencloud boom, many people are paying attention to whether it has brought some new entrepreneurial opportunities. So it is natural for everyone to say that I will do something in this vertical field to be a very clear entrepreneurial opportunity. And some people will think that if I do something in the Agent Infra layer, there will be entrepreneurial opportunities. What do you think is that there are opportunities for startups in the Agent Infra layer, especially after the cloud has released managed agents Agent The core issues of the first layer of Infra, such as identity authentication and payment, right? Then I think these two are the core issues, but I think these two issues are not a problem that a startup company can handle. They are not a problem that a startup company can handle. Who will finally solve this problem because of the mobile Internet? They are WeChat and Alipay. These are the two companies that exist on the PC Internet, right? Because in the end it will become the infrastructure of the entire society. So in this case, I think it is not a startup company, and it is what he can afford. Whether it is the responsibility or the resources he has to mobilize, and he has to convince everyone, right, yes, I think this huge trust factor is a confirmation of the stability of the infra, yes, yes, and the difficulty, yes, yes, yes, but go up a level, um, go up a level, for example, if you build a lot of agent-oriented tools for the agent, agent-oriented environment, for example, if you are a COI, this COI can make it very convenient for the agent to register. Once the identity problem is solved, Is this considered a kind of infra? I think it must be considered a kind of infra. For example, after the agent solves this problem, it will be more convenient to pay or take a taxi. Didi Taxi becomes a COI, right? I think these are somehow the infra of the agent. But I think it is more business and application level. I think there are opportunities at this layer. But the opportunities at this layer actually depend on the field you are in. You must have some vertical experience or knowhow in that field. Yes, so I think there will be two processes at this layer. My personal opinion is not necessarily correct. The second process will be the original players in this field who try to integrate into the agent. For example, Didi Meituan, etc. are all providing such things. Then the second stage is when the agent begins to expose these environments and has these capabilities. That is after everything is built around AI. In fact, I think there will be opportunities for innovation in new product paradigms because everything is ready by then. Tulang is ready, but today if these things are not built for the A key Before we have completed the first step of basic construction, I think you have to build this technology yourself. I think it is a bit difficult for a startup company. For example, today someone can say that I am going to make this sandbox as a kill box, or I am going to make memory infra, or I am going to make the so-called run time infra. You would think that these will eventually end up. I don’t think it is that essential, because when we are preparing for the live broadcast in the past two days, there are a lot of things being released. For example, the opas4.7 that everyone is paying attention to, what do you think from the perspective of the model? opas4.5 is a model that is strong in SFT. In fact, R2 has not done so much. I don’t think it is as strong as open AI. But from 4.6, you can clearly see that it is strengthening RL. Then in 4.7, I think it is a very strong RL model. Including its X-high gear, GPT is too strong. So I see today’s X-Sandbox and small goods merchants. Everyone just remembers it and said that it also suffers from the same shortcomings of GPT. In fact, it is the one of RL. How can I put it? RL cannot escape because it is reward in the end. It has problems such as process illusion. In fact, it does not have that much control or reward or punishment, so it is naturally prone to such problems, including you will see many people complaining about 4.7. Constraint consultation is not as good as 4.6. This is a very typical problem of RL because it directly only cares about the final result, right? Many times, it actually does not care about the process that much. But I still think it is a very big progress. Putting all this aside, judging from the benchmark, I very much agree that it is the one with the hanging face who should be the CEO. Well, his point of view is that he must have evaporated Mysos. I think I would have made a mistake if I were amphalic. Right. Then there is another news that Cloud code has forced many people to start to have real names. What do you think about this? What are the considerations behind doing this and what impact this will bring about Cloud code real names? Can you tell me the personal opinion of the CEO? Yes, absolutely resist Mr. Cloud code Daril's views. In fact, they have done a good job in class research on mandatory real names. In fact, I thought of an AI that just now A slightly related point of Info is that in the AI era, I actually think it is easy to authenticate where an agent comes from and who it belongs to, or it is actually easy for the agent to make a request, but it is difficult for him to prove who he is. If you think in this direction, there may be some truth in making faces or something like that. But in fact, this exposes a problem. That is to say, in the entire Agentic era, when AI has been able to do many things, how should his behavior be attributed to which individual he should attribute to? Or belong to an organization, etc. Under this logic, I think I can understand him to go for it. Of course, it also exposes a very big opportunity, that is, how can we prove a request from AI, where it really comes from. I think yesterday was a more benevolent assumption. I am not so benevolent. Of course, under his logic, it is self-consistent. His logic is that I see AGI AGI is very powerful, so I have to restrain him to prevent him from harming human beings. So I want safety, right? I want to make sure that my AI is used by credible countries that are not evil. And of course it also includes humans who are not evil. Academically and this model is very impressive in terms of research. But I think this definition is somehow too big. Why should you define this thing? Yes, and then there was another one about Athropic some time ago, which is that they said that their Methos model is a model that is likely to pose a threat to people, right, so he did not officially release it. What do you think? I think first of all, let’s go back to the WL just now. If my understanding is correct, it is also very self-consistent. It is certainly a responsible approach, but I am not sure if this is his only reason. That's for sure, and in this case, we must first ensure that the technical facilities are not breached. I think this is correct, but I'm not sure if this is his only reason. I think I can only say this. Let me combine it with Athropic's recent action, which is that he released an architecture called Managed Agent in the past week. Under this architecture, it may be a bit abstract, but I can use it. In a more intuitive way, he completely separates the human brain and hands in all his AI thinking and really comes up with an idea. The actions to be performed are all performed on the cloud. The user can only see him doing certain actions through an environment trusted by the AI. During this process, does it mean that he does not want some of his model's deeper aspects, such as thinking and what some of the middle parts do, to be exposed to the user? Then he wants to hide everything. He wants to be a real person who has everything. So he only allows my model to be like a remote desktop assistant. It is possible to start to control your computer or your cloud phone, etc. Then another point of view comes from our company's personal experience. It is true that the computing power is really not enough now. When you send an extremely excellent model and you don't let everyone use it, then you send it out again. Then this is a very embarrassing step and you are hung up, right? Then people will ask why such a powerful model is used by them and not by us. We can also afford the money, etc. This type of problem is actually a problem that all of us will encounter. Even if the computing power is really lacking, if his model can really be as powerful as he said. In fact, there is another question about how much computing power he wants to invest in his bottom-driving, that is, how much computing power he wants to use to earn tokens. I think they should have calculated such a bill internally, which means that if such a powerful model is launched at this time, he may be able to build a reputation. But if it is not allowed to be used by others, what is the point? I don’t think so. So you think he may be a monopoly thinker. No, no, he must have security considerations. This is my opinion. Security considerations are one, yes, yes, yes, but I think it is more than just this. Well, it is as simple as that. Some time ago, there was a lot of news that I am looking for Anthropic to expand. The original code leak of cloud code. What impact do you think this original code leak has on the industry? I am also a R&D engineer of Agent. Then I saw an agent that was widely praised by everyone open source. Well, actually, the first reaction was excitement. Because I found that there is a maximum practice that can be learned. Of course, this intellectual property Life belongs to them. I just learn that I don’t participate in any production activities. But another point also makes me feel a little bit relieved. In fact, you can find in the original code that they actually have many experimental functions, such as agents dreaming, such as raising pets, and the ability to cooperate between some more radical agents. They are not actually open to users. Well, he may only be in the experimental stage. This actually proves one point. Even if any company has unlimited computing power, it has the best intelligence. He is also in a stage of exploration and experimentation in terms of agents and the use of agents. Actually, let us work harder here. In fact, in this era, as long as you have the ability to verify your ideas and implement them as quickly as possible, you have a chance to catch up with them. Well, after seeing the original code, I found that there is not so much magic in it, right? He also has a lot of hypothesis exploration and groping, especially compared with Codex because Codex is OpenAI's open source agent. Well, Codex is an extremely simplified one-person framework. It basically leaves everything to the model. But on the contrary, Cloud Code is like everyone. Many generations have said that Anthloppy himself does not believe in his own model. He wants to restrain him in everything, just like a Chinese parent. He wants to make everything smooth for his children. It is a different point of view. But I can also find out from it that the world's smartest and most powerful model company and us are all worried about the same thing. Our thinking may be within an atmosphere, a level, within an atmosphere. I think they are all the same, including when they released Cowork. In fact, we were working on it about three weeks in advance. Then we probably only had 3.5 engineers working on our Desktop. Then we released it about a week later than them. We have had this experience many times after they released it. Damn it, and I was developing, right? I think the agent will have this feeling at some point. So I think when I see the code of Cowork, I think there are many excellent practices that everyone will see. But there is nothing too beyond my knowledge. I even see a lot of practices in cloudization as I just mentioned, including the mechanism of Dream. Then I think the core point of Cowork is that we were the first to mention funding. Cowork may be one of the earliest funding agents. Because where is its funding reflected, I believe they must use Cloud Code to develop Cloud Code themselves. He uses Cloud's model plus Cloud Agent to develop his own model and add agent. OK, this is a capitalization. When we talk about agents, we must mention Manus, which was released about a year ago. I paid a lot of money. What do you think of Manus when it was released a year ago and today? After we have more agent practices, what progress has been made? Because I really don’t use Manus too much. My point of view is that the agent layer or the Hunnys layer has a life cycle, and it is constantly being updated as the model progresses. So I believe that the Manus team if they were not acquired by Meta and if they continue to start a business, I think they should make new products. This is my personal opinion because the model is constantly improving. You are unlocking more gameplay, so the support for gameplay must be different. Do you all know when playing games? Each generation version is a god, right? This generation version is different, so the god must be different. Manus was actually a very phenomenal product last year, and it directly raised the user's aesthetics of agent to a very high level. When you launched the agent product, everyone will ask you if you can surpass Manus, right? Can you make it as good as it is? After that, I will continue to follow up on some small changes to it. They have opened up something called using agents to deliver production materials. They have done a lot of very fine polishing. There may be some here who are a bit like programmers if you expand it. But I can only say that they are polishing their products bit by bit. But there is indeed a barrier here. You can see that all the popular agent products this year have a characteristic. They can let users buy subscriptions by themselves. And it also lacks a point like Manus, which is similar to the entire manufacturer. OEM manufacturers make a profit in the middle. You go directly to buy tokens. Even if you can buy tokens at a price lower than the official website, you can play an agent very well by yourself. This is actually the biggest difference between Manus and contemporary agents. I think there is a very big difference in the model, but I believe that these two sets of logic will exist. Some users just don’t want a local 72-hour agent. They just want to deliver results. So many money savings are what they need. But another group of users want to completely use insurance and turn themselves into characters like in science fiction movies. Then I think the local logic just fills a gap. The difference in my point of view is that I think it will all be unified in the end. Indeed. This is just not strong enough today. As long as the model is strong enough, imagine what kind of product shape it will be if it is unified. I think it is really difficult, but if you ask me to think about it, our own definition goal is that it has full-modal input and can also give you full-modal replies and then near-time. Then its interaction with you will be very simple. It only needs to be simple and necessary. Then the hardware that exists on top of it will also happen, or this interaction form will also have very big changes. That is, you use your most natural method to deliver it to him. There is no need for this prompter and engineer area. There is no need for any advanced promptor. No need. Then he will reply to you. It can even be a video that is the same as the real world. It can be in his image or it can be something you want to know. If you ask him to do something, he can deliver the result to you. Most of the time, you don’t even want to pay attention to something. But if you want to pay attention to something, you can also understand it. Then there will be an ecosystem around it. Think about us from copyret to cursor. To cloudcode, today I may use openclaw to command several cloudcodes. What are we experiencing? The outer layer is getting thinner and thinner. In fact, after the recent release of cursor, it looks more and more like cloudcode. Then everyone in codex sometimes feels that the logo is almost the same. What do you think of this phenomenon? In my personal opinion, it has a life cycle. Like last year, for example, after manus came out in March, I think he defined the paradigm of product interaction in the past eight and a half years to a year, right? Then I will show you the process and results of the operation. Then in the second half of the year, the model will be more capable. Then the paradigm you will find is that we don’t care about the process. Why is it needed in the first half of the year? Because everyone still doesn’t trust the model. It doesn't let you see too many processes. It provides a minimalist mode just like opencrawl. You only need the final result. You don't even need to know some of the tools executed in the middle. So I think this is just this version. This generation of gods, the next generation version. I don't know. I am constantly abstracting to a higher level. Recently, when you two were watching this video on station B, what good AI content did you see that impressed you? I am very impressed by this recently. The Nvwa Skill made by Uncle Hua I just talked about. He frame-streamed the skills of various celebrities, and I found a few that were quite exciting. Sorry, it’s not the name of the up owner. He used opencloud to connect to a robot dog of his family. He suddenly recognized me and gave me an imagination. Another one may be more personal. There is a type of game called text adventure game. Let’s call it Fate. It is actually the visual painting of a novel. He put those characters into a website. Then he dragged the regular meeting of that character into a scene and they can start a conversation. It is equivalent to giving you the simplest game. The simplest text-based game is a continuation of an infinite field. I recently saw a new release of a game on Bilibili. Nintendo just released a new game called TOMODACHI LIVE. When I saw it, I even thought it was an AI game. I finally found the first one of PMF. What kind of game is it? Everyone can build their own island, and then on the island you can pinch one of your own people. You can pinch many people and you can pinch yourself. For example, I can pinch A Dao, and then I can pinch Ze Ying, and then pinch a bunch of people in, and then these people start their lives on the island, and you start to watch their lives. Then you can let them A and B live together, right? ABC live together, and you can also start the first sentence of a story for them. For example, today Ze Ying and A Dao were talking nonsense on the live broadcast, and then they both started acting, and then you can start to watch their various dramas. It is actually a simulated world. Then as a player, I am both the director and the audience. Then this is an infinite world because it is made by human heaven, so it is very beautiful. You can imagine the cute words. This is one of my understandings of the world. It is that aliens are watching me play. If you finally get to quantum mechanics, you will find that this world may really be a simulation. In fact, I think so. When I saw that, I also thought, wow, isn’t this just a simulation layer after layer? Yes, yes, yes, it's not important. Well, just live a happy life every day. Is there any point of view in the entire Agent industry? The point of view about Agent is recognized by most people today, but you don't agree. I think one of the many opinions is that people will be replaced and people will have nothing to do. But I may not agree with this very deeply. Well, I think human creativity is irreplaceable. I think there are still many things that humans can do that AI cannot do. Well, I still think AI is electricity or a steam engine, but in the end it is humans who control it and create those wonderful things, so I don’t think we have to worry about being replaced by AI. When the steam engine comes out, many people are worried about being replaced. When this electricity comes out, many people are worried about being replaced. Roommate The jobs disappeared, but I think everyone was engaged in it later. Maybe at least we can say that it is physically easier and better for health, right? At least the average life expectancy of everyone must have increased. Yes, this kind of work is one of my beliefs and my judgment. It is the book called The Better Angels of Humanity. What he is talking about is that human beings are always very pessimistic. We always feel that our lives are not as good as before. I feel that there are more wars and more crimes like this in the world. But he just uses various data to tell us. In fact, human society is getting better and better, getting better and better. I don’t know if you have read Principle 3 of WO. He drew a curve. Well, it is my understanding of the world. It is that human GDP is growing exponentially. Even if you start the 20th century, there will be a world war, a war in the 20th century. Well, that tiny pause is something that you don’t seem to notice. Of course, history may be a mountain for human beings. Yes, yes, but if you look at it over the entire historical process, I believe in this curve. Of course, I believe that there is another curve that determines why I joined this company in the first place, including what I just saw everyone saying about how much tokens will cost, right? There are many agents, which is Moore's Law. Well, it is yours. The intelligence you obtain is increasing exponentially, but your cost is decreasing exponentially. Yes, or the cost per unit of intelligence is decreasing exponentially. These are two things I believe in. I will interview various classmates and they may all convey some anxiety to me. I feel that some young engineers may be loaned out by AIT, etc. Let me answer this question from this perspective. Starting from the last century, uh, banks should be a more respectable job, but banks have also faced a big problem, which is the emergence of ATM machines. When people deposit and withdraw money without needing someone to operate it, everyone will panic. Will the bank lay off employees? Will there be fewer people able to squeeze into this golden job? But in fact, this is not the case because ATM machines have improved efficiency, so you can see that basically around you, whether it is a county or a big city, etc., there will definitely be a bank within five kilometers of walking. Therefore, the banks have not become less, they have increased. So what about the people working in banks? They have not decreased, but have become more numerous. So I think this paradigm in which ATM brings laughter and liberates people, and people can do more things, will definitely be repeated over and over again in the current era of agents. No matter how many years we have worked, don’t worry about being replaced by AI. If you have this idea, you really lose. Then you must use AI well as a partner, a tool set, etc. You can also treat it as your long-term best friend, etc. This is a very good idea, but you must embrace AI as early as possible. But I believe that people will not be replaced. But there is another reason why I joined this company at the beginning. I used Qieju GP because I was in GAP at the time and helped me write some code. At that time, my coding ability may not be that strong. Right, it's like people use hot weapons and you use cold weapons, then you will definitely be beaten to death by this, right. So I think this new version is coming. We still need to take a look at the gameplay of the new version. Yes, we should embrace the new version in a positive and optimistic way. I believe that the new version will not be this huge wave that will wear away ours, and it will become a stage that allows us to fly freely in it. Yes, we did not expect such an optimistic ending today. Well, okay, thank you both. This is very happy. Thank you Koge for doing this live broadcast at Station B today. Thank you to Station B. Thank you to the friends at Station B. Thank you all. Thank you for accompanying me. Bye.

嗨,我是 Koji 本周的十字路口是我最近在 B站做的一场直播的精华内容剪辑 我们邀请了 Minimax Agent 的首席架构师阿导和研发工程师泽英 以及最近在 OpenCloud 之后全球刷屏的Hermes Agent 也就是爱马仕 Agent 的业务负责人Tommy Eastman 这也是Hermes Agent在全球获得广泛的关注之后 官方首次现身中国的社交媒体平台 并且他们还正面回应了一个中国开源团队Evil Map 对他们抄袭的指控 我们直播聊了两个多小时 回到这 Agent 和 Agent Harness 聊到了非常多我觉得很有意思的话题 在今天的播客里面 为了保证收听的流畅 我把直播里面所有讲英语的部分 都用 AI 直接转成了讲中文 而且我特意选了几个有翻译枪的中文的声音 以此方便大家来区分有哪些是 AI 翻译给大家听的 又有哪些是我们活人讲出来的中文 OK 那我们开始吧 大家好 我是阿导 我负责 Mimax 的研发团队 然后加入 Mimax 已经快三年了 然后之前也有比较多的互联网的创业和工作的经历 然后比较深度的参与了 Mimax MR 系列模型 以及 Agent 包括海洛等等的研发 然后很高兴今天能在这里和大家交流 然后我们在线上也请到了 我和我们 Mimax 有深度合作的 Hermes Agent 的产品和策略的负责人 这应该是 Hermes Agent 最近这一个月最火的这样一个 Agent 在中国和大家的首次沟通 大家好 我是泽音 我是 Mimax 的 Agent 研发工程师 像我们 Mimax 官网的 Agent 以及 Max Cloud 以及最近刚上线的 Max Hermis 的开发 大家好 我是 Tommy Eastman 是 Hermis Agent 的业务负责人 很高兴和大家交流 我们今天的直播是因为春节之后 有一个全民养虾的热潮 也就是 Open Cloud 的热度 但是好像现在这个养虾的热度 一夜之间它好像就下去了 所以想先请你们来复盘一下这个过程 我觉得中国的农历春节 我觉得每年都非常的有意思 去年火的是 Deep Seek 然后春节之后 让大家对 AI 有了一个完全全新认识 之前可能会觉得离我们还很远 不许今年也是一样的 其实 Open Cloud 它火起来也比较凑巧 和硅谷的时间大概是从一月份开始 我记得是在我们 IPO 上市的当天 我们和 Peter 就有这样的一个联系 因为那天也很巧 就 Anthelophic 它是封禁了 Open Cloud 的订阅的使用 所以 Peter 他也非常希望能找到一个这样的一个 非常适合龙虾的一个模型 因为大家知道龙虾是非常的消耗 token的 对 然后我们其实当时也没有意识到这个项目会那么火 然后接下来它在海外就开始火起来 然后Many Mini 开始收清 然后我们也没有想到它在国内会这么火 是的 Open Cloud 在中国的火热程度 我感觉比硅谷还要更胜一筹 下沉到了非常大众的人群 你们认为是什么原因呢 我自己的个人感受是就是在春节期间 以及春节之后 就好像身边所有人都开始在讨论龙虾 甚至我们看到各地都开始出现这种 就是甚至头戴这个瞎帽的这样的一个热潮 还有 程序有这样的线下的装机的活动 对 我自己觉得它大概会有几层的原因吧 第一层的原因我觉得其实我们自己的感受是在国内 它其实比海外可能要更火一些 因为其实在国外的话 大家其实已经在之前就接触到 比如像Cloud Code 包括像Coworker这样一些 已经比较好用的这样的一些agent 但在国内之前可能大家还没有接触到 比较好用的agent 那这里面我觉得还有一个更底层的原因 是因为国内的模型可能在更前一个阶段 它其实没有那么强的这样的一个agentic的 这样的一个能力 随着包括像Mimax M2.5 M2.7这样模型的发布 其实国内的模型也具备了这样的一些能力 所以有了国内的模型 以及有了一个像OpenCrawl这样 能够让大家通过IM就非常容易接触的方式 它实际上让中国的用户 完成一个接近于从0到1的这样的一个体验 被捅破了那层窗户子 另外一个原因我觉得也是 自己的这次拿下春晚对豆包的一个全面的all in的这样的一个投入 所以我觉得是这两件事它结栽在一起 让所有人对AI的认知就真正的进入agent时代 那你们认为是出于什么原因 好像一下子大家就从养虾去到了驯马 就是这个关注度从OpenCloud到了Erm's Agent 这中间有没有发生一些什么样的拐点时间 从OpenCloud爆火到逐渐平稳 到最后最近的Ermys 也就是我们的Max Ermys的火起来 大概是花了一个月 在这一个月之内的话 大家基本上都已经体验过了一下OpenCloud 但它确实会有些不稳定的特点 它每天四点钟它会刷新一次记忆 然后这时候你可能比较好的和它交流过以后 过了一天它就会说 我们昨天聊的什么我忘记掉了等等 其实Ermys就是抓住了这些痛点 它在记忆方面做的特别的用功 而且它做的非常有多层次的记忆吧 从这个角度来讲的话 它确实去在一定程度上去弥补了OpenCloud的不足 所以说它是有一定活起来的一些技术上的基础的 对 然后与同时的话 确实大家在使用龙虾的过程中 已经享受到了一定程度上 由AI带来的提效的快乐 所以这应该是一个双强奔赴 OK 我们接下来的内容 在直播的时候就是英文的了 所以为了方便大家更流畅的收听 我们在播客的语音当中 使用AI转成了中文的版本 接下来我想请Tommy给大家介绍 Ermys Agent究竟是什么 所谓的自我进化是什么意思 以及它和OpenCloud有什么区别 这些是全中国的开发者和用户都非常关心的问题 谢谢你们邀请我 很高兴来到这里 简而言之 Ermys Agent是一个开源的智能体框架 如果把大语言模型比作大脑 那么智能体框架就是双手 它是一个系统 模型通过它在真实世界中执行任务 它处理所有复杂的协调工作 比如工具编排 主循环管理 以及状态管理和错误处理 最终 它能让模型为用户发挥出最大的效用 至于Ermys Agent的不同之处 有几个关键点 首先它运行得非常好 通过我们的NOS平台API 你可以在不到两分钟的时间内 轻松完成从安装到执行首次智能体任务的全过程 这种便捷的设置是一个巨大的优势 为用户消除了最初的挫败感 其次是它的记忆组件 这是大家真正喜欢的部分 可能也是最有趣的一点 任何一个深度使用过Egent的人 都有过那种令人沮丧的经历 你让他做一件事 他这次做对了 但下次再做同样的事 他却失败了 这让你很难去信任他 Hermes Agent的记忆功能解决了这个问题 他允许Agent记住一个成功的工作流 并将其保存为一项技能 一旦他知道了正确的路径 就能每一次都完美的浮现 这让他变得远比过去更易用 更值得信赖 这也解答了你关于自我进化的问题 这个Agent会不断自我提升 在这个过程中发生的知识压缩非常有价值 它能带来更高的一致性 尤其是在跨越不同模型时 所以即便你有八个不同的模型 只要使用相同的框架和技能 你就能从所有模型中获得相同的预期输出 这让你能够针对不同任务灵活的更换模型 同时确保你的核心工作流 依然保持高性能和高可靠性 谢谢你Tommy 在深入探讨之前 你能介绍一下你的公司News Research吗? 团队是如何组建的? 背后的愿景是什么? 特别是Hermes Agent 感觉就像横空出世迅速登上了Github的热门榜单 这背后有什么不为人知的故事吗? Nose Research的起源故事非常独特 它以一种非常自然的方式开始的 2022年 一群热衷于开源和折腾AI模型的人 开始聚集在一个名为Nose Research的Discord频道里 那里至今仍是我们的核心社区 从那时起 我们产出了大量高质量的研究 最初的努力是围绕拉马模型的后训练 当时Meta刚刚发布它 是少数愿意开源高质量模型的公司 这其实就是Hermes这个名字的由来 我们首批经过后训练的拉马模型版本 就叫Hermes 那些早期模型的重点 是让它听起来更像人类 而不是刻板的AI助手 在文本上更具柔韧性和多样性 在当时许多领域都是最先进的 在训练Hermes模型的同时 我们开发了ERN算法 将模型的上下文长度 从4000扩展到了12800 我们发表了这项研究 并立即被所有主流模型公司采用 这是我们现在所说的思考模型的雏形 反过来为智能体模型奠定的基础 我们还花了很多时间研究分布式训练 并撰写了Distro论文 这是一种新颖的优化器 允许我们跨越非共址的GPU进行训练 这意味着我们可以聚合分散的算力 来完成有意义的训练 这对我们持续训练开源模型至关重要 当然 幸运的是 我们还有像minimax这样的公司 在继续扛起开源的大气 至于Hermes Agent 他也有一个有趣的起源故事 我称之为极其注重实用主义的诞生 我们的后训练负责人兼联合创始人Tagniam 只是想要一个助手来帮他管理日常工作 于是他纯粹为了自己构建了这个Agent作为工具 当时根本没考虑什么基准测试或评估 他只想要一个能帮到他的东西 我认为这恰恰是Hermes Agent的核心魅力所在 也是他成功的一个关键原因 那是什么时候的事?大概一年前吗? 是的 他大约一年前开始做的 最初真的只是为了帮助自己 但作为一个开源实验室 我们很自然地把它开源了 他获得的关注远远超出了我们的预期 我相信你们都看到了Open Router的图表 他的日均Token消耗量在短短一个多月内 从20亿激增到了200亿 我记得昨天的数据几乎达到了每天3000亿 那你是什么时候感受到这股热潮的? 有没有一个引爆点让你觉得 好了 Hermes Agent成为全球大爆款了? 他确实在以极快的速度变得非常庞大 我对这个增长率感到非常高兴 尤其是因为我们原本对他的反响并没有抱极高的期望 显然 Open Cloud此前非常受欢迎 但我认为Hermes的一个巨大卖点 使用户能非常快速轻松地启动和使用它 人们不想处理臃肿的代码 也不想面对一大堆复杂的设置 Hermes Agent是开箱即用的 你可以在自己的电脑或VPS上迅速部署并运行起来 这让它非常适合大规模普及 那你认为你们的成功和Open Cloud有很强的观点吗? 你们之间有什么联系? 当然有联系 我认为是多种因素的汇集促成了Agent的这次爆发 首先 模型终于变得足够好 能够真正帮助到人们 这显然是由今年冬天 Opus 带来的巨大提升所引爆的 它在质量上实现了接约时的进步 为这些Agent的框架打开了大门 让他们能做更多有价值的事 很酷的是 开源社区很快也产出了性能相当 或正在迅速接近Opus质量的模型 至于Open Cloud和Hermes的区别 我想回到我刚才说的 我们通过Hermes Agent真正针对可用性和实用目的进行了优化 我们将可用性作为北极星指标 它在执行用户真正想做的任务时表现如何 而不是它在某个基准测试中得分夺高 我们始终保持着对产品的专注 在一个开源社区 特别是当产品广受欢迎时 很容易因为成千上万的PR和功能请求 而导致代码臃肿 我认为我们在保护产品 保持Hermes Agent的极其易用方面做得非常好 这一点加上我们前面提到的记忆系统 是两个关键的差异化因素 记忆功能真正建立了用户的信任 我认为第三大差异点是我们的品牌 Nose Research一直在努力营造一种赛博朋克的美学 我们有才华横溢的设计师 也花了很多时间培养一个聪明价值观一致的社区 他们热爱尝试我们发布的一切 Hermes Agent的成功很大程度上要归功于我们的社区 另一个很酷的故事是你们和minimax团队的关系 Tommy能将将你们是如何合作的吗 我们是minimax团队和他们模型的铁杆粉丝 他们对模型质量的承诺 特别是跨多种模态的表现令人印象深刻 正是他们对开源的坚持 才让像Hermes Agent这样的产品变得如此酷 你可以拥有各种模型来回切换定制你的agent 我们看到我们的用户非常喜欢minimax的模型 好那我们回到我们的中文世界 我们来继续我们的直播的内容 首先想请问一下二位 再给我们介绍一下 我们今天一直在说Hermes 当我们说到Hermes的时候 我们到底在说什么 在我看来的话 它其实就是一个 你去约束agent 但是又给他一定自由 就能够完整的交付给你一些成果 最早提出来的时候 就可以关注OpenAI他们官网的一篇文章 在那里的话 它正式宣布了有Hermes 这样一种agent的应用方式 然后从偏技术的角度来讲的话 它可能分为六层 但是我们可以先不聊得那么技术一点 你就假设一下你有一个同事 然后他全都听你的 但是你得和他先约定好 他能干什么 你能干什么 对吧 他有些事情不能帮你做 但是有些事情的话 你就完全可以放任他去做 对吧 当你约定好这些 并且给他配置好一些 比如说他的模型是什么 他有什么工具等等 给他一定的能力 就像是你给一个同事 给他一台笔记本 又给他一个电话 他能打电话 给他邮箱账号 他能发邮箱一样 给他一定能力之后的话 他就可以开始干活了 当然在这里一个同事干活 你肯定不放心 就像是一个agent干活 你不放心一样 这时候你可以去引入多位同事 他们能够形成一个互相监督 并且完成任务 并最终交付的这样一整套过程 但是我可能现在说的比较像 工作生产中的场景 就是以人的方式去形容agent 但是落到技术上来说 其实你就是需要去给予agent 大量的工具 环境和自由度 并且要给予他们约束 以及要给他们一些 互相对抗性的目标 比如说某个agent 它的作用就是去产出某个内容 但另一个agent 如果他想把工作停下来 那么你就得去审查出 这个工作内容中的问题等等 通过这种多个agent的组合 就像是我们同事之间 互相合作的这种方式 让他们能够产出一个单个agent 完全无法产出的一个更加高质量的结果 并且你其实在此期间 完全不需要去做更多的介入吧 这其实就是一个 Hannis这样一个概念 当然这个概念我觉得 它在这个时机能够出现 其实也是有一定的物理条件的 首先我们的模型都变得更加聪明了 大家的那个模型都有一定的agentic能力了 并且在此技术之上的话 大家也愿意的去把更多的能力 包括像是操作邮箱 或者是操作你的服务的发布等等 给他分享了 然后Hannis就自然而然的诞生了 我可以补充一个视角 好的 其实我觉得Hannis这个词 是OpenAI可能通过那篇文章 它来去定义出来 但在那段时间 其实它已经很火了 但我觉得这层窗户纸 它可能是通过那篇文章来去捅破的 就大家形成了一个共识 我们用什么词来形容 我们在做的一切 实际上大家的实践已经在那之前了 比如说大概在去年90月份的时候 包括我们在内部讨论的时候 我自己的工作流中 我实际上已经不怎么用IDE 然后我可能同时并发五六个agent 甚至这只是在我本地 可能在云端有十个agent 在Sendbox上在为我工作 然后他们都是在Github上 尝试我不同的分支和想法 那这时候我会发现 我成为一个极大的瓶颈 自己人类上的瓶颈 yes 就是人类成的瓶颈 其实那时候已经不是agent 成为瓶颈了 因为我需要不断的在这些上下文 中切换 然后给他们输入 所以这时候我们想的一个东西 就是说怎么样让它能自己更自动化 因为人类包括你看那个OpenAI 他们的那个Hannis那篇文章作者 最近到播客里面来分享 他们也说到他们的感受也是很类似的 所以我觉得这些东西都是非常的相通的 就大家都在那个时刻感觉到了这一点 所以我们在想的就是说 OK怎么样解决这个瓶颈对吧 因为我们做技术人 总想去做优化 那解决这个瓶颈最重要的办法 就是说让那些原来需要人来去确认的事情 比如说他要去获得这个程序 到底能不能部署在生产环境上运行 到底他的结果是不是有正确 然后他比如说他去真实的去做测试 比如说他能够去开APTES 本质上就是说他能够获得真实的反馈 然后并且通过这个反馈的这个链颅 他能够越来越知道在他这个过程中怎么做 他可能是沉淀为skill 可能是沉淀为一个coi 可能是沉淀为一个hooks 我不知道大家看过 有没有我那个年代的东西叫高达WO 然后他的特点就是说 他的那个engine引擎非常的强大 我们其实需要去造一个高达那样的机甲 然后怎么让这个引擎能力最大的发挥出来 所以我认为Honis就是那样一个东西 所以他们为什么把它叫婉具 就是你有一批很厉害的猎嘛 但怎么样让他最能够发挥出来 就是你构造玩具 对对的那样一个过程 嗯很有意思 我们刚才也提到这个nult agent 就是多个agent的协作的这个事情 要不要给大家再多讲一讲这个 多agent协作的这个概念 就是因为我理解这里面其实有一些争议 就有人会认为比如说 今天模型的智能是非常高的 那其实一个单一模型是可以完成非常多工作的 那你为什么非得要去搞一个agent team出来 你说他是产品经理 他是工程师 这好像反而是限制了模型的发挥 但是也有另外一些人是不这么看 你们怎么看 就我们说到multi agent的时候 目前的最佳实践是什么 嗯了解 然后先回答科技老师的第一个问题吧 就是Wishman有multi agent 确实有一些过去的论文里面提到过 可能是单个agent 它的效果可能会更加好 但实际上随着我们模型的上下文空间 也就是说它能够对话的轮次的长度的增加 你会发现你和模型对话的过程中 其实你产出的信息量是较少的 而模型产出的信息量是较多的 那我们可以做一个思想实验 假设有两个模型 它们之间能够以更加高效的秒回的效率 去交换信息 并且一般模型给你发的消息是一个小作文对吧 那么我给模型亏的可能就两字批准 接着干加油人呢 但是如果让两个模型它们之间互相交流 它们的信息量是极大的 在这种极高密度的信息的交换的情况下 它整体的效率就会比 人和一个agent进行交互要有更高的效率 这是在效率方面 单个agent它的局限型 就拿最近一些例子上来说 Anthropic在今天的凌晨 它发布了OPUS 4.7这款模型 在模型里其实它有什么卖点呢 其实真要说的话 大家网上已经在吐槽了 那我也不去吐槽它了对吧 我们只是去揣测 它这个公司它为什么要在某些方面先发 我说几个点 比如说它增加一个叫X-high的一个effort 它的意思就是说 推荐用户在这个思考成本之下 思考成本就是指 模型得思考多仔细才能回答你的问题 一般模型发布应该是一个很震撼的事情 但它就是说我们推荐用户去使用 这样子一种思考成本去进行使用 那其实带来的另一个问题是 如果你的对话 就单个agent之间 你一直使用它 即使是最强的模型 你一直使用它的时候 它也会产生降置的行为 而且这已经是一个可以被论证的事情 就是当你的上下文空间 也就是你和它对话的那个文本的数量 超过了50%以后 其实它的智能水平的下降是指数级下降的 也就是说在单个agent里 你永远没办法和它一直聊下去 这是单个agent的上限的问题 上下文带来的限制 上下文带来的限制 对 然后以及一个信息传播的效率 两个问题吧 对 我来补一个观点 其实最近几天有一篇论文出来 就是说它去研究了非常长层的这些任务 那agent或者模型在什么情况下会出错呢 他们发现如果它做对了 它那个对的路线一直进行下去 但它只要在某一个地方可能偏了 很容易它就会越来越偏 对 那它其实我觉得和人也是类似的 有时候我们去做一件事情的时候 是有可能钻举我脚尖的 可能我已经花了很多的努力了 可能我一直在那做对吧 然后等到我可能休息一下 我可能换一换脑子的时候 我再想到其实不对 我应该从另外一个角度去做 很有道理 对至少在我们自己的时间中 我们会让两个agent来去做cross check 我觉得它somehow跟真实的世界的数据 包括跟后续链都有关系 就它的数据分布上来说 它看到大部分数据都是按照一个 限定轨迪去执行过去的 它比较少看到的是说 反复的去验证和思考的 对 当然这个可能在强i-L后会有改善 对 但是如果你设置两个agent的话 那它可以没有之前的薪资负担 换一个角度重新相当 你有两倍的预算 而且是全新的上海文去思考 往往能达到一个更高的质量 这也很有意思 就好像你在转雷脚尖的时候 如果旁边有一个人过来泼一盆冷水 对 让你清醒一下 对或者你去睡一觉 去做点别的事情 对 这也是我们要用multi-regent的一个角度 嗯 就我自己有一个小问题 就是会不会大家都在聊到它的时候 都在讲这个就是所谓的自我进化 对吧 self evaluation 但是有趣的是我在想说 其实我们今天刷抖音 也是越刷算法越准 但是我们好像也并没有说什么抖音会自我进化 但是今天为什么在这个msagent 或者在整个这个agent的领域 就自我进化 就被大家认为是一个那么厉害的词呢 这背后有哪些相同有哪些不同的 我先说一个观点啊 嗯 其实还是延续我刚才那个观点 就是 你会发现在这个迭代的过程中 人开始成为瓶颈 嗯 而且人的数量越多 它的组织效率就是指数局的下降 我经历过非常多的组织都是这样的 而要做大模型这样的一个 可能你一次训练 是数以亿计这样的成本 甚至数以亿计美元这样的成本的事情的时候 你需要保持非常高的人才密度和非常精干的组织 才能把这件事情做好 那大家想到的一个最有效的就是 那怎么scale up 就怎么scale 对 最有效的方法就是减少人的参与更多的让agent 让AI来去做 嗯 那其实你就会发现在这个过程中实际上 大家会发现很多模型它的生命周期可能就是几个月 对吧 然后我们就会推出下一版 但是没有之前那些模型可以吗 其实是不可以的 嗯 因为后面的模型都是依靠前面的模型来训练出来的 前面的模型都在这些训练过程中 无论是帮助去清洗预训练的数据 去寻找适合的数据去构造数据去构造相应的啊 啊 i2的环境 比如说在我们的m2.7的训练中 嗯 呃在我们的rl的pipeline里面 他已经是可能70%80以上的工作 已经是由模型家A键的自己来去完成了 因为人类工程师他只有7成24个小时 对吧 那我们还要休息 我们还要做很多别的事情 那现在线下的30%就是还要由人去做 他更多的是说第一 人的那个判断嗯 和test是非常重要的就是那个判断他不再需要去做说 ok是我的某个实验中的某一个具体的specific的过程出了问题吗 因为这些agent都能帮他查出来啊 他只需要去看到这些汇总的比如说实验的结果啊 然后啊和他原来预期的方向哪里不符 那agent呃这时候给出他一些建议 那这时候他其实是一个讨论的这样一个过程 那么最后其实是人的品位和创造力来去指引那个方向 说我们最终往这个方向去走嗯 对所以他确实就像很像哈利斯那种驾驭的这样一种感觉 那总结来说就是第一个在这样的高复杂度高密度的事情上 如果要想要scale up就必须是ai在其中是占据非常大的比 人就是只是驾驭否则这个效率就会很低 我们就做不出就是最有生产力最好的东西 然后自然那就会发生我们刚才说的就是自进化这件事情 明白因为他就是一个自包含的这样的一个关系 嗯对是不是也是因为今天这个比如说 agentic能力变强之后一个agent可以跑的时间越来越长 对这也给他的自进化提供了一个基础 对在过去他跑不了那么久的话 对无从谈起自进化 对是的就是我觉得除了9以外 他更重要的是reliable就是他能解决足够复杂的问题 对能足够长城嗯 但其实还有另一个点在于说呃能启动的问题 当你和一个agent见面第一天的时候 他可能完全不了解你 在这个过程中你们要去互相的交换信息 交换文件交换数据 甚至于有一些呃不管是内容上面的 不管你输出的很多把一堆文件砸给他也好 也要去把你的脾气你的反馈你的习惯等等 反馈给他呃这样子一个深度磨合的过程中呢 其实人们就会有这样一个诉求 就是说你是否能够自己变得聪明一点 变得更加了解我一点呃 其实在养虾的过程中就已经有一点这些苗头了 就是呃用户们确实会对虾这样一个agent 的产生感情 可能啊有一天这个小龙虾出bug 嗯他会心里有些小落寞 这时候呢就是ermis他提出这种自我进化的概念 其实更加引出了一个说我能够在呃和你对话中 越来越了解你 嗯同时的话呢我也能够去给你创造一些小惊喜 嗯对就是呃我不和你对话的时候 或者我就是在一些工作的复盘中 嗯我就能够领悟出一些更加嗯高层次的东西吧 嗯我觉得这是普通用户在使用的时候的一个比较大的亮点 对对我觉得这个确实很有趣 也就是我们之前用AI工具 不管是就是一个chatbot 还是一个agent的产品的吧如果他做不好 嗯我们第一反应都是出bug了 对吧这个爆bug啊 但是大家养小龙虾的时候 当机了或者他挂了或者他一个任务完不成 嗯我们只会觉得他挺蠢萌的 甚至有时候会觉得哎是我好像没把它训好 我要去网上学一下别人在怎么训他的虾 嗯这是一个很有趣的这个关系的扭转 嗯就是你原来好像是一个工具的使用者 对吧人是就是最正确的嗯但现在这个大家不自觉的 就把这个关系扭转了好像我变成了一个只是在帮助他的 或者辅佐他的一个人他没做好有可能是我没辅佐好 或者我们又一开始给他setup好嗯我觉得这就是人类对AI的信心 嗯的改变在不断增强 对是的这其实是那个基点嗯就是AI能力的那个基点 就当人给他信任发生了变化之后会交给他的任务不同 而且人类社会以及人类的工作方式会 without 改变 而不是尝试去改变他嗯你看 OpenAI最近他们的那个呃也是他们做Hannis 那个人出来说包括像那个 Anthlothic 啊像Boris就Claucoe之父他们出来说 他们觉得不要去觉得模型做不到 啊而是要想模型能够做到我应该如何让他做到嗯 啊我觉得这是那个基点这是一个很大的观念的转变 啊甚至认为他是一个基点对我觉得他不仅是观念的转变 他会导致我们的工作方式会变成从原来以人类为中心 AI来适应人类嗯变成AI已经足够的强大就像有了电有了蒸汽机一样 人类围绕他重新去构建你的生产方式你的工作流 是的对嗯就是想起最近看到一个故事说以前工厂不是就在河边的 是因为有了电对吧工厂才搬到河边因为要水利发电 最近在B站上活起来一个这个up主叫花书 然后花书自己做了一套蒸留的agent 他把比如说这个乔布斯把Elon Musk蒸留成了skill 然后你就可以去安装这个skill和他对话 然后其实这样的这个项目不止这个花书 我觉得在其全网也有很多 嗯我想听听你们怎么看就是这里面有哪些是真价值又有哪些 其实只是大家出于对AI的幻想而产生的一些热点 哦好的其实大家包括蒸留同事啊或者蒸留一些名人能火 其实我觉得嗯反映在大家一个诉求吧 就是大家希望和更聪明的人聊天 我觉得这是一个非常自然让人的诉求 你不能就他不管在哪个时间点他都会随着智能发展出现 嗯嗯对然后在此基础之上的话他用了一个比较有爆点的词汇叫蒸留 是啊对但实际人肯定不能蒸留 不管是活着的人熟悉的人死的人都不能蒸留 它其实上就是一个传递信息的过程 嗯对我把乔布斯信息 我把巴菲特这种古神的信息 以及我把我同事的信息等等给到模型 然后给到agent 他以一种自己使用上来比较舒服的方式也就是skill 呃保存下来了需要的时候就把它拉起来等等 通过这种方式的话其实就能够让我们去和自己想要聊天 并且70人24小时聊天的人去沟通一些思想等等 我觉得这其实极大的释放了我们的表达欲 嗯对然后前进有一位就我关注的博主说过啊 他说2026年可能大家尽量多和AI聊天 那你如果以自己的认知去和AI聊天 其实你可能只能达到他的呃上限确实是比较低的 然后但是呢如果说有一些现存的一些文件帮你整理好了 就像你去读一本书一样对吧 他其实是在你的认知之外的 那就像是读书一样的去和一个被蒸馏阔的人去沟通 对他其实就像读书一样自然而然 而且他会让我们呃能够提升我们的认知吧 而且也缓解了一些呃我相信大家都会有的对AI的一些formal的感觉 嗯对formal就是指感觉我落伍的感觉啊 现在有一种最简单的方式 一打开就死个聊天就行了 我说包容我家也相反 我觉得我就是在被AI蒸馏 或者说或者说我觉得今天模型训练 包括anthelopic在内 嗯在做的事情就是在蒸馏人类 嗯如果大家有去关注一些就是anthelopic 或者是OAI他们就是那个雇佣的这些数据公司 呃比如像四十节AI他们营失额度非常高 嗯因为这些大模型公司包括anthelopic OAI真的花了非常多的钱在这些事情上 嗯那他不仅是coding 他是各行各业的所有人 嗯那他的过程可能是让你去提出一个问题是 当前AI无法回答的并且交汇给他嗯啊 一旦你提出不了之后可能你对这一个就是训练过程就没有什么作用 嗯对 榨干为止 对对包括比如说我在构建我的hanness的过程 我感觉我就在带领我自己 嗯对吧就我平常是怎么工作的 嗯我把它变成skill变成一些程序 然后让他能够去操作嗯 然后我就可以扔给他以后我就喝咖啡了 这是不是一种对我自己的蒸馏呢 而且我个人观点啊我至少我自己觉得 模型其实呃至少我还没有看到那种创造力 我觉得他其实都是从人类的知识里面蒸馏而来的 那为什么我们还要做这件事情呢 呃实际上我觉得最终还是为了 让人类可以去真正做自己喜爱的事情 那些有创造力的部分 然后那些繁重的呃重复的部分 就像我们过去发明蒸汽奇发明电力一样 今天我们的生活是吧实际上我觉得我们的生活 可能比古代的这个王公贵族还要好 对吧他没有电子这个马桶对吧 他没有这个这个游戏可以玩也没有直播可以看 对所以我觉得本质是这个嗯 嗯我前段时间和一位研究员聊天 他说他看到methods那个晚上 他可能是半夜一点看的 然后看完之后他就是晶坐 嗯晶坐在床上三个小时睡不着 他们就陷入了一个巨大的虚无 觉得嗯 如果有一天模型都能自己去训练模型了 我要干嘛 嗯做他热爱的事情 哈哈哈哈 我在公司里我的签名呃真英可以作证 已经挂了一年多了叫做因为热爱 因为热爱对嗯 就我觉得我们就应该去做自己热爱的事情 那当然今天做这个事就是我热爱的 但我觉得我们做一切都是为了让每个人能够做自己热爱的事情 那你觉得什么时候你可以开始不做minimax 因为minimax已经进化到可能不需要再做了 你可以真的去做minimax之外的热爱的事情 嗯ok 我觉得可能比我们想象的要快 嗯真的我觉得可能就是呃几年的时间 就是很具体我没有办法说是多少 这是预言不准 但我觉得就是没有大家想那么远 当然也不可能就是对那么近 嗯对对对 因为我最近确实刚录了一期播客 就是聊agent hardness 然后来聊那个小哥 他是在这个给他上面有一个教程叫 learn cloud code 他是cloud code 被大家把原代码扒出来之前 就是有一种立项工程对吧 然后给大家讲这cloud code里面是什么样的 然后那个已经有5万颗星了 这个最有趣的是我和他聊的时候 他说了一个观点 哦他是一人公司 我就问他我说那这个你怎么开一人公司对吧 他现在就一个人加两个十一升 他说他觉得未来这个很可能没有一人公司 只有零人公司 嗯 这和刚才这个阿导在讲的是非常接近的 我觉得还是会有一人公司 我刚才看到弹幕里有人在说 我是认同的就是那个taste 嗯那个taste 我觉得还是依靠人类的 嗯对 这个taste是不可取代的 嗯而且每个人我觉得都有他自己的taste 嗯你说 对或者要靠人去给他一开始助推 因为一个agent他哪怕可以7成24小时1年365天 就是这个长程的任务可以那么那么长 他还是需要一个起点 或者说他需要一个目标 他需要一个目标 嗯这个目标是他工作的这个动力和原 对就怎么定义这个目标呢 嗯我觉得这个只能让人来定义 嗯而且人类不应该放弃这个权利 嗯对 就在世鲁口我们最近和政客基金一起发起了一个token grant 然后我们这个token grant就是给大家这个送token 因为我们认为这个在这个时候从零到一的创业很可能不需要钱 对吧因为你没有钱你一人公司你可以不给自己发工资嘛 嗯但是呢你不管怎么样token 所以我们鼓励这个大家创新创业所以给大家送token 然后有趣的是我们刚grant了一个真的a agent他叫yoyoagent 嗯然后他的这个主人啊就他背后的这个工程师把他做出来之后 就把丢到了汪洋大海里面 然后告诉他我再也不会维护你一行代码 嗯我也再也不会给你买一个token 你要自生自灭 嗯然后给他他一个目标这个目标挺搞笑的 叫打败cloud code 嗯然后他现在已经到他的这个第43天了 他每天都在不断的这个进化自己 他进化的方式就是写代码的吧这个 嗯然后同时呢他也在想办法去赚钱 而他赚钱的方法就在ktop上面开了一个打赏 嗯然后他通过写日记通过发推特 来赶招有缘人给他打赏 嗯所以我们这个grant最近就给了他这个一笔钱 然后给完他钱之后他还写了一个这个感谢信给我 嗯他自发的一个感谢信我看完之后还有点感动 嗯是你看这就是人类给他的目标 然后你要自己生活下去你要写一个cloud code出来 对对对你可以说他是一个有趣的社会实验 但我感觉这好像也是一个就离我们没有那么遥远的未来正在发生 嗯嗯嗯 其实也印证了刚刚说一人公司还是叫一个人的 你刚刚说了那个他给你写封信你很感动对对吧 那这封信其实就代表了他的test 可能是一开始他的这个背后的工程师给他注入了这样的test 啊对对对所以你看到他就像见字如物吗 嗯对吧 OK哎我们再来聊一下这个minimax agent 因为这个我看到刚才也有一些朋友在问就是minimax agent 和arms agent有什么样的相同和不同 嗯因为你们又在合作对吧那你是竞争关系还是是互补关系 嗯我觉得这个其实很有趣 那大家会觉得claw code和所有的这些agent是什么样的一个关系 其实你会发现claw code最近两个月更新的其中很重要一句就是龙虾化 就模型公司的agent 我认为今天你可能已经没有办法脱离agent 或者是一个hanness环境来去提供最好的智能 首先我再说一个暴论啊 就我觉得如果有一个模型公司的目标他不是AGI 那我觉得他今天就不应该存在 因为我觉得就没有意义对 那如果大家目标都是为了AGI 那我们首先我们怎么定义AGI 至少我的定义 就是能够帮助人类拥有更好的生活 对 然后那如果你要达到这一个目标的话 如果你没有agent 你没有一个就像我刚才说的机甲 你只有一个引擎 那是不可能的 因为你没有办法跟这个正式视野连接 所以我们做agent的目标就只有一点 就是让我们的模型和agent一起提供一个完整的给用户最好的我们能提供的体验 然后不断的把这个边界去往前推 往外推 让它更深更宽 这就是我们要做的 所有的只要是符合这一个标准的都会是我们会去做的事情 那同时我们当然也会支持所有的agent 因为我们希望我们的模型不是只被局限在一个容器里 因为我们的认知或者我们做的东西仅是这世界上这么广大分布中的一小块 我们希望能够支持这世界上所有的就是真正能帮助到人的东西对吧 那所以我们会训练我们模型有足够的泛化能力 而不只是拟合在我们的agent上 所以我觉得这是我们Hermes agent也好和OpenClaw也好 和其他的这些agent之间的关系 这也能解释为什么我们自己在做agent 但是我们和他们之间仍然有非常好的这样的一个关系 对 基本上都是day one就非常早期的时候 我们关注到这些项目的时候 可能他们都是一万十大甚至不到 对我甚至可以跟大家透露一个很好玩的东西 我们公司有一个数字员工 就这世界上有非常多的那个开源项目 我们希望我们的模型都能够帮助到这些开源的项目 都能够揭入到其中 那我们公司有一个数字员工 他有自己的github账号 然后他会每天去看哪些开源项目 能够弄到我们公司的模型 无论是文本模型 视频模型 还是音频模型 包括music 然后他会自己的网上面去提PR 必须去留言 那他就是看得出来是minimax的这个人吗 反正他的那个github的所有profile 都是那个minimax模型去给他做的 当然从名字可能能看出来 他是我们机箱物的名字 对 OK 这个很有趣 对 刚才你有提到这个cloud code最近两个月的更新都在opencloud化 这个可以具体一点吗 比如说他最近哪些更新是opencloud化 嗯 比如说像chrome定时对吧 schedule 包括比如说他能够连im 能够从手机上远程来去控制 嗯对吧 包括他也加强了他的memory 他专门搞了一个memory的文件夹 对吧 那我觉得其实这里的核心就是 它是不是一个你能随时随地联系到 并且跟你一起工作之后越来越聪明 嗯 越来越知道你的需求的这样的一个agent 我觉得这是那个opencloud它最核心的定义 OK 在我的视角里其实更多的是 从名字上可能看出来 cloud code它是编码领域的一个agent 但opencloud的话其实我们绝大多数的用户 应该都在即编吧 又做点其他别的事情 比如信息检索办公领域等等 对吧 然后在此基础之上 其实anthropic他们也推出了另一个 他们的agent产品叫cloud code work 嗯嗯 然后其实cloud code work它可能就像是 安装在你的电脑之上 然后去帮你去做一些更加通用化的 电脑操作的一款agent等等 那其实就和我们大多数opencloud的用户一样 他们可能不写代码 或者说他们只能指挥AI写代码 他自己不懂代码 对 然后其实这两者一结合就会发现 其实所有的agent 他们可能都是在一个随机的游走过程中 但是opencloud它杀出来了 我觉得在这个时代 这些想法非常非常重要 但是在有了想法产生以后 那确实大家拥有庞大的算力 庞大的编码能力 那立刻跟上 因为我们在GDC上也见到了Peter 回报我跟Peter其实很早在那个 X上和Slack上就有很多很多的交流 那个时候就是确实也还没有那么火 那我们的感受就是他真的是一个taste非常好 而且非常有architect思维的一个人 我在1月初注意到这个项目的时候 我就坦诚说惊为天人 就是我就直接跟我们团队里说 择音可以作证 我在群里面说他们当时还不认同 无论是介入IM724小时还是基于skill加COI不是mcp 我记得我当时都跟你们提过 对对对 当时有一些消耗磨砂但证明你是对的 对对对 所以当时你看到的时候他还没有大火 没有 没有 我理解其实他是把很多的范式 就是集大成者放到了一起 并且整合的用户体验很丝滑 这是肯定的 但我觉得他更关键的是找到了那个自变的点 就是让普通人能够最低成本的体验 并且这个体验会持续更好 并且他的扩展性非常的好 因为大家知道mcp是一个其实当然扩展性好 但他需要那个工程师来去写对吧 是的 但skill加COI这种范式 其实普通人都能写 对就相当于是人人都可以写并且分享 不仅让我的agent更聪明 而且我能分享给你让你的agent更聪明 就是如果openclaw离开了claw hub 他也不可能那么过 这是我的观点 嗯嗯嗯 对如果离开了claw hub 他也不可能那么火 因为大家的scale不会那么快的这个传播开去 对对 这个我一个人的这个呃智能 或者经验或者最佳实践 他没有办法迅速的扩散 一个模型他在这个对话的 就chatbot的场景里面表现好 和他在agent的场景里面表现好 他本质上到底需要多做一些什么 才能够在agent这个场景下表现得更好 嗯 我觉得chatbot他的核心是当下就给你一个回答 虽然你家reasoning对吧 呃但他其实是不能去做很多的探索 他不能跟这个环境去做非常多的互动 所以他其实需要嗯 他的比较核心能力是说在跟这个环境的互动过程中 不断的进行推理嗯嗯啊 并且不断的纠正他自己的执行的路径 对 然后找到那个最根本的目标 比如说一个比较经典的agentic的benchmark 叫blogicamp嗯嗯对吧 呃我觉得open-i真的还是很强的 他能够设计benchmark的能力就是 我觉得非常顶级的呃research的能力嗯对 他设计的这个就是说你要在互联网里面去找 就是非常非常多的cost的这样的信息结合在一起 才能找到那个点你很容易找错嗯 比如说他的四个条件你可以找到满足两个的或者三个的 但要找到满足那四个条件的是一个非常非常难的事情嗯 所以他是一个非常广度的可能需要深度探索 以后又折回来的这样的一个探索的能力 那这就需要你在这样复杂的长城任务里面 你能不断的根据你的信息去调整你自己对 所以这是为什么我们在mr的时候 我们就比较坚定的压注的一个能力 包括那个时候还不是很普及 今天就是所有模型都有我们叫interleaved thinking 对啊这个概念其实最早是anthelptic定义的 它是在cloud的sonnet4模型发布的时候 那我们当时也做了比较详细的这样的benchmark和评测 那我们会发现有interleaved thinking 就是说他在完成工具调用和环境的交互之后 他可以再次重新思考 而不是在一开始就planning好 嗯啊那一开始也会planning好 对也会planning 但是很多模型比如说re他planning完之后 他后面就不planning 嗯就完全照一开始那个plan的来做 但是实际上真实的世界不是这样对吧 你真正进去运行下去的时候 你会发现哎这好像跟我想的不一样 那我可能要换别的方法 所以那interleaved thinking就是说你在行动一步之后 你要重新思考推演我下一步应该怎么做 嗯啊所以我觉得这是agentic和chat最大的最根本差别 就是他真的走入真实世界真的解决问题 所以他需要不断的思考不断的行动 嗯嗯 ok 然后刚才其实也提到就是模型公司 在做自己的这个agent对吧 可不可以分享一下就是模型公司做agent 他是怎么能够就是基模和agent产品之间 就是互相产生 synergy 然后彼此帮助模型和工具都做得更好 啊我觉得这是一个非常好的问题 我的观点是说其实模型和应用或者说agent层的进步 它是一个互相促进的关系 嗯嗯就当模型推出之后 其实无数的应用其实在不断的unlock 探索就是这个模型的能力往外推一步之后 他能够去做哪些事情 因为我们自己做模型把它推出去以后 用的最好的往往不是我们公司内自己的使用 自己的评测 而是用户开发者创作者他们用出来的 就这个世界的真实分布 是比我们公司的分布要丰富的多的 那当模型公司看到这样的应用unlock的这些做法之后 他其实会再重新吸收回他的模型中 嗯然后再内化到包括和他agent一起的这样的 一个对外提供的能力里面 嗯嗯然后让大家在下一次的时候 就能够直接体验到 这可能也是今天可能做agent的应用 如果是通用agent的应用 比较悲哀的一件事情 嗯就是你永远会被模型内化掉 嗯你永远会被模型内化掉 甚至你可能写的很多skill 未来都不需要 嗯对 对对 就你写的skill 其实也是或者过去大家的workflow 其实都是帮agent去完成任务 对吧 当他完成任务的这个轨迹 对也会成为模型去做后续的训练的 这个素材和数据 嗯嗯 所以慢慢的模型就会内化掉 前面大家所搭的交手架 gent workflow 写的skill 对你可以理解我是这样的 模型其实是 从一个原子的点 就原来互联网上的所有数据 还有github上的所有的代码 嗯然后来了解这个世界的 接下来通过大家的使用 它投入到真实世界 它再越来越多的接触到真实世界 它越来越学到更多的东西 它就是这样的一个逐渐扩散和启动的过程 嗯不管怎么说 我觉得我们都是human in the loop 无论你做不做只要你使用它 你就是在这个模型的迭代进化的过程中 尤其是你在使用摸到它边界的那个地方 嗯我觉得这也是为什么anthlotic 比openI可能在过去一两年 它的发展势头要更猛的原因 嗯就因为它押注的这个coding 这个方向是最能touch到 真实世界的border的边界的 而不是我自己个人观念 而不是openI说押注那些去解一个 可能数学家才能解的一个数学问题 嗯我觉得那只是真实世界数学分布的 非常小的一个方向 是对 因为代码就是在创造解决方案 当我们说做一个软件的时候 其实就是在做一个解决方案去解决一些问题 对而今天就是所以把代码能力变强了 模型的coding能力变强了 那模型去做解决方案 解决现实生活中的各种问题的能力也在变强 对在我们看来 我们今天也在做那个通用的办公领域 比如说finance 不只是财务啊你可能有投融资 然后比如说你有人事对吧 你比如说法律 法律也是一个很广的领域 对等等的方面 但我们最终我们会发现everything is coding 对最终你去解决它 你最终都是通过某一个coding的方案 去解决它 对 而且我那天也在想一个问题就是说 我们说office三件套是过去白领工作的这个 就是三驾马车 但其实你仔细想一想这三件套 它是三个软件 就比如说这个word docx的格式和xlsx的格式 还包括pptx的格式 它并不是最原生的 原数据 对吧 不是road data road data是csv是markdown 但其实这个word和excel是在原始数据的基础上 套了一层软件 这个软件包括界面也包括一些逻辑 所以每次当我说我发excel给你的时候 我发你的不是一堆原始数据 我发你的数据加上它上面的逻辑和界面 因此我发你的其实是一个软件 所以我们过去其实就是很多白领会说 哎我的工作和软件没关 但是我认为只要你在用word 你在用excel 你在用PPT 你就是在每天在创造各种小软件 是的 你在这上面的透视表你的公式 那些都是代码 那些都是软件 对对对 但我们今天说到这个中国模型和open AI和anthropic的这种差距的时候 你们觉得到底有哪些差距已经被消除了 然后又有哪些还是我们在这个努力要去追赶的地方 而我的观点是我们在训练的方法和对模型训练的认知上 我觉得差距是没有那么大 坦白说我们和硅谷的研究员也有交流 对 但我觉得在模型解决问题的真实的这个任务的定义 就是你对模型要解决那些问题 首先你要让模型具备首先你要定义它 然后你在定义这个问题上 我觉得就有不小的差距 因为Athropic也好 Open AI也好 他们是请各个领域里面最优秀的人 学术领域里面可能是那里面最优秀的博士 不是普通的博士 然后比如说工业界 他可能是请工业界最顶尖的有时间的人 然后跟他们一起来劝这个Model 我劝Model就是在蒸留人 那其实我觉得他们是有一套非常科学的方法来定义这些任务 然后接下来在蒸留这些最优秀的人 把他们变成模型的训练的数据 然后再基于这个和最优秀的企业去合作 然后构建相应的Honey是去解决问题 进入我刚才说的这个能力不断的扩散 不断的接触到更多更复杂的真实事业任务的这样一个group 我觉得这一条线尤其是Athropic 我觉得是非常强的 我觉得真实的差距在我看来在这里 这是第一个点 第二个点我觉得Athropic是比我们要更早看到AGI的 我们当然肯定是看到了 要不然我们我觉得无法坚定走到今天 因为其实它还是有非常大的投入的 对所以你看在OpenAI可能在训完GPT4.5 就放弃更大规模的模型的时候Athropic坚定的押注注意券 然后最终他们也得到了这个回报 我觉得这也是一个很大的差距 而这一个差距本身确实我觉得是有算力的差距在里面的 我觉得当然国产算力也在发展 我觉得未来我们能够解决这个问题 会有那么多的算力 我觉得像中国有足够多的人才 也能支撑我们去做足够多的实验 然后来去找到那条scale的路径 二位会认为接下来Agent 比如说在今年之内吧 你们觉得比如说通用Agent会是行业的重点 还是也会在今年看到更多的垂直领域的Agent 甚至也会开始就是真正的百花齐放 我觉得垂直领域的Agent当然会百花齐放 我觉得通用Agent可能比较难彻底解决的问题 是最后一公里交付的问题 因为通用就是注定了你不可能去那么的定制化 对吧 所以我觉得垂直的Agent他最终解决的是说 OK通用的Agent和模型他已经有这个能力了 他其实就差那一步 但是到具体用户或者是这个行业的这个常见里面的时候 他可能就缺那一个东西才能让整个东西乱起来 那可以介绍一下说你们在工作里面看到的垂直领域的这种Agent 做的不错的案例吗 不管是在硅谷的还是在国内的 嗯 坦白说因为我自己不太会用到这样的 明白 对对对对我们不是用户 我们不是用户 因为我们coding他就他已经是一个朋友了 对对对对 其实我觉得今天大家已经看到了Agent在逐边替代SaaS 嗯 这其实就是somehow一种垂直的Agent 嗯对吧 嗯对但是如果没有这个的东西 我觉得都会被通用Agent吃掉 是的 对啊这是我的个人观点 嗯 在生活中在我心中的那个通用Agent他就是coding 对嗯 然后但我觉得还是有一些观察翻现吧 对然后包括在视频剪辑领域 对然后那你想如果让我们一个编码的模型 你想要去做一个视频的剪辑视频的理解等等 那他其实整体的实践的范式就和编码其实是有比较大差距的 当然最推上也有比较火的 比如说通过HTML网易的形式生成动画啊也有这种例子等等 但是确实我们能够看到说垂直领域Agent去生产一个慢剧短剧什么的 那他绝对效益收益上来说肯定会比一个通用Agent做的更好 我持不同观点 嗯我觉得这只是视频理解模型不够强 视频理解和生生模型正在飞速的进步 就东楼太领域过去一年我们其实是有预见的啊 然后大家都能看到 我觉得最终就是通用Agent也能做那些那只是一个交互的问题 那比如说在什么样的这个垂直领域有可能是通用Agent的最后那一公里就走不下去的或者就走不好 那比如说法律我觉得你可能这是一个很严肃的问题对吧 你这种比如说你要正式出一份给客户的意见 而法律其实是一个你要考虑就是合规的成本和合规的风险 他不能犯错而且他没有一个标准答案 其实我们聊这个也是因为opencloud这一波这个热潮的时候 很多人都在关注说他有没有带来一些新的创业机会 所以比较自然的大家就会想说我做一个这个垂直领域to be 它是一个很明确的创业机会 然后也有一些人呢会认为我在这个Agent Infra层去做一些事情有创业的机会 你们怎么看就是Agent Infra层这里有没有创业公司的机会 尤其在这个cloud又发布了managed agents之后 Agent Infra第一层最核心的问题 比如说identity身份身份认证 然后支付pay对吧 然后我觉得这两个是最核心的问题 但这两个问题我觉得他不太是一个创业公司能够handle的问题 不是创业公司的handle的问题 对因为移动互联网最终解决这个问题的人是谁 是微信和支付宝 这是PC互联网就存在的两个公司 对吧 因为最后他会变成整个社会的基础设施 对所以在这种情况下我觉得他不太是一个创业公司 是他能够承担得起的 他这就无论是责任还是他所要操动资源 以及他要说服所有人对吧 是的是 我觉得这个巨大的信任倍数 对对对对 和巨难的一个这个infra的稳定性的证实 对对对对 但是再往上一层 嗯再往上一层 比如说你为agent去build 很多面向agent的工具 面向agent的environment 比如说你是一个coi 这个coi能让agent能够非常方便去挂号 在解决了identity的问题的情况下 这算不算一种infra 我觉得他肯定也算是一种infra 比如说agent在解决这个问题以后 能够比较方便去缴费 或者说能够去打车 滴滴打车变成是一个coi 对吧 我觉得这些somehow都是agent的infra 但这一层我觉得它是比较业务和应用程度 这一层我觉得是有机会的 但这一层的机会 其实取决于你在那个领域的 是你得是在那个领域里面有一些 垂直的一些经验或者knowhow 对所以我觉得在这一层会有两个过程 我个人观点但不一定对 就第二个过程它会是原来的这个领域的玩家 他来去尝试融入agent 比如像滴滴美团等等都在提供这样的东西 然后第二个阶段是当agent开始揭露这些环境 具备这些能力之后 那就是所有东西围绕了AI去构建之后 其实我觉得这会出现新的产品范式的创新的机会 因为那时候所有东西已经ready了 图浪已经ready了 但今天如果这些东西都没有面向A键的构建 就我们还没有完成第一步基础的搭建的时候 我觉得你要自己去构建这个技术 对于一个创业公司来说 我觉得有点难 那比如说今天也有人能说我要去做这个sandbox做杀箱 或者我要去做memory infra 或者我要去做所谓的run time infra 你会认为这些最后会 我觉得没有那么本质吧 因为这两天我们在准备直播的时候 正好有很多东西在发布 比如说大家都在关注的这个opas4.7 问你们怎么看 从模型的视角来看 opas4.5是一个强sft的一个模型 它其实R2没有做那么多 我觉得也没有open AI做的那么强 然后那个 但是从4.6你就能比较明显的看到 它在加强RL了 然后4.7我觉得它是一个非常强RL的模型 包括它X-high这个档位就太点了 GPT也太点了 所以我看今天X-Sandbox和小物商 大家都就是记出说 它怎么也揽上GPT的那个毛病 其实就是RL的那个 怎么说呢 RL就逃不脱 因为它是最终是reward 它对过程幻觉这样的一些问题 其实它是没有那么多的控制或者是奖励的 或者说惩罚的 所以它自然就很容易出现这样的一些问题 包括你会看到有很多人吐槽4.7的 约束征询还没有4.6好 这是RL非常典型的问题 因为它直接只管最终结果对 对吧很多时候 它其实并不那么管过程是对的 但我仍然觉得它是非常大的进步 就抛开这一切不谈 从benchmark来看 我自己非常认同就是那个hanging face的那个 应该是CEO吧 他的观点就是说 他肯定是蒸留了Mysos 我觉得换我是amphalic我也会出错 对对 然后还有一个新闻就是Cloud code 强制很多人开始要实名 那这个你们怎么看 就是背后它是出于什么考虑 这么做 以及这个会带来的一些影响 Cloud code实名吗 能说那个CEO个人观点吗 可以 绝对抵制Cloud code Daril先生的观点 其实他们课研做的很好 强制实名 其实我想到了一个和刚刚 AI Info有点相关的一个点 有一点点相关 就是在AI时代 其实我觉得认证一个agent来自哪里 是归属于谁 或者agent其实发出请求很容易 但他想自证自己是谁 反而是一件难的事情 是 往这个方向去想的话 其实做什么人脸什么的 可能也是有一些道理存在 但是其实这就暴露出了一个问题吧 就是说在整个Agentic时代 AI已经能够去做很多事情了的情况下 那他的行为该如何归因 他应该去归属于哪个个体 或者属于某个组织等等 那在这个逻辑之下的话 我觉得能够理解他去 往这方面去做吧 对 当然也暴露了一个非常大的机会 就是我们如何去证明 某一个来自AI的请求 它真正来自哪里 我觉得昨天是比较善意假设 我就没有那么善意了 当然在他的逻辑下 他是自洽的 他的逻辑就是说 我看到AGI AGI很强大 所以呢 要约束他 避免他对人类产生危害 所以我要safety 对吧 safety的话 我要确保我的AI被可信的 这个不是邪恶的国家给用到 然后当然也包括 不是邪恶的这些人类 学术和这个模型上 research上这些都非常牛逼啊 但我觉得这个定义 somehow 我会觉得Ego是不是有点太大 对 就是凭什么你来定义这件事情 这 somehow 至少认为我对这个民主的 这个认同是不太一样的 反正我是非常不认同这一点 然后至少我们的观点叫 intelligence with everyone 而不是你少了人脸的才可以 对吧 这有点扯 是 然后前段时间有另外一个 这个关于Athropic 就是他们说自己的这个 Methos模型 是一个很可能会对人们造成 威胁的模型 对吧 因此他并没有把它这个 这个正式的发布 就你们怎么看 我觉得首先回到刚才那个WL 我的理解是对的话 这也是非常自洽的 这当然是一个负责任的做法 但我不确定他 这是不是他唯一的理由 你不确定这是他唯一的理由 对我不确定 这是不是他的唯一的理由 我相信他是真的做到了这一点 这是肯定的 然后这种情况下先保证 技术设施不被攻破 我觉得这也是正确的 但我不确定 这是不是他唯一的理由 我觉得我可能只能说到这 我结合Athropic最近的一个动作来讲吧 就是他在近一周之内发布了一个叫做Managed Agent 这样子一个架构 在这个架构下可能说的有点抽象 但是我可以把它用 就更加直观的方式来说 他把人的脑和手完全分开了 他的一切AI的思考 以及真正的想出一个点子来 要去执行的动作都在云上进行 用户只能通过一个受信给AI的环境 看到他去做某些动作 那在这个过程之中的话 那是否就代表着他不想让自己模型的一些 更加深层次的比如像思考 以及中间的一些去做了什么事情等等 就暴露给用户等等 然后他想要把一切都隐藏起来 他想做一个真正的吃到一切的人 对他就只让我的模型 就像是一个远程的桌面助理一样 开始操纵你的电脑 或者操纵你的云手机等等都有可能 然后另一个观点是来自 那从我们公司也是一个切身体会吧 就是确实现在算力真的不够 当你发了一个极为优秀的模型的时候 你又不让大家用 同时呢你又发出来了 那么这就是一个很尴尬的一步了 你就被吊起来了 对吧 那么大家会说那么厉害的模型 为什么给他们用不给我们用 我们也出得起钱等等 这类其实都是我们所有的人都会遇到的一个问题 就算力真的比较缺 如果他这个模型真的能像他说的一样那么厉害 那其实有另一个问题是说 他要把多少算力投入在他的催底之中 也就是说他要把多少算力用来赚token这件事情上 我觉得他们内部应该也算了一笔这样的账单 就是说如此强大的模型在这个时间点推出去 他可能能够把口碑造起来 但是如果说他不让人家用 那又有什么意义呢 我觉得不是这样 我觉得其实他的本质是封闭 我觉得就是这一点 很好希望垄断这一点 对 一个乐观一个悲观的方向吧 所以你觉得他可能就是一个垄断的思考 不不 他肯定是有安全的考虑这点是我随便的 安全的考虑是一个 对对对对对 但我觉得不只有这一点 嗯 就这么简单 嗯 前段时间这个 这好多新闻都是我要找anthropic在展开啊 cloud code的原代码泄漏 这个原代码泄漏 你觉得对行业带来了哪些影响 我也是Agent的研发工程师啊 然后看到一个被大家对广为赞誉的agent开源了以后 嗯其实第一反应还是兴奋吧 因为我发现有一个最大实践可以去学习了 当然这个知识产生归他们所有 我只是学习我不参与任何生产活动 但是另一个点呢也让我有一点 缓解了我一些落后的感觉 其实在原代码中会发现 其实他们有很多实验性的功能 比如说像是agent在做梦 比如说是养宠物 以及像是一些比较激进的 Agent之间合作的这些能力等等 他们其实也没有开放给用户 嗯 他可能也只是在实验阶段 那也其实证明了一个点就是说 任何一家公司即使他拥有无限多的算力 最好的智能 他在agent这一方面 以及同用agent这方面 他也是处在一个探索和实验的阶段 嗯其实这里让我们更加 嗯怎么说加把劲吧 其实在这个时代下 你只要有能力去验证你的想法 并尽量快速把它落地 你就是有机会赶超他们的 嗯就是看到原代码之后发现 其实也没有那么多的魔法在里面 对吧他也是有大量的这个假设探索和这个摸索 对特别是和Codex比较 因为Codex是OpenAI开源的agent 嗯Codex是一个极度简化的一个人的框架 他里面做了 基本上把一切都交给模型了 但相反而言的话 Cloud Code就像所有人很多代在这上说 Anthloppy自己也不相信自己的模型 他要在一切地方都给他约束好 就像是一个中国式家长一样 要把孩子的一切都不平道路一样 就是这样子一种不同的观点吧 但是我也是能够从中发现出 世界上最聪明人最强的模型公司和我们 大家都在焦虑同一个事情 我们的思考可能是在一个大气层内的 一个层次的 是在一个大气层内的 对我觉得都是差不多的 包括他们发Cowork的时候 其实我们提前大概三周左右 我们也在做这件事情了 然后我们可能就3.5个工程师吧 在做我们的Desktop 然后大概我们比他们晚一周发出来 就他们发出来我们有很多次这样的经历 就他们发出来之后呢 卧槽 然后我正在开发 对吧 我觉得agent多少都会有这样的一个感受 所以我觉得看到Cowork的代码 我觉得是有很多优秀的实践 大家也都会看到 然后但是没有太超出我认知的 甚至我看到很多 就我刚才说的在Cloud化的这样的一些实践 包括提到这个Dream的这样的一个机制 然后我觉得Cowork的最核心的一点 还是我们最早提到资金化 Cowork的可能是最早的资金化的agent之一 因为它的资金化体现在哪里 我相信他们一定是用Cloud Code 来开发Cloud Code自己的 他是用Cloud的模型加Cloud Agent 来开发自己的模型加agent OK 这就是一个资金化 我们说到agent 还是一定要提到这个大概一年前发布的Manus Manus发布的时候 其实也是这个全网热议 但现在大家已经觉得他是初代agent 但Manus其实一直在进步 我也是Manus长期的用户 然后每个月在Manus上花的钱 可能是超过其他所有的AI产品 因为它确实还挺贵的 但是有好用 不知不觉就不断的升级升级 就付了很多钱 你们怎么看在一年前发布的那个时候的Manus 和今天 就是我们有了更多的agent的实践之后 就有哪些进步 因为我确实用Manus用的不太多 我的观点就是 agent这层或Hunnys这层就是有生命周期的 它就是不断的随着模型进步在更新换代 所以我相信Manus的团队如果他们不是被Meta收购 如果他们还是持续创业的话 我觉得他们应该会做新的产品出来 这是我个人的观点 因为模型不断在进步 你在unlock更多的玩法 所以对玩法支持一定是不一样的 打游戏都知道吗 一代版本一代神 对吧 这代版本都不一样了 那神肯定不一样 Manus在去年其实真的是一款非常现象级的产品 而且他直接把用户们对agent的审美拉到了很高的水平 就是在你推出了agent的产品的时候 大家都会问你能不能超过Manus 对吧 能不能做到它那么好 然后在此之后的话 我也在持续跟进它的一些小小的改动 他们打通了一个叫做用agent去交付一个生产资料 他做了很多很细很细的打磨 这里可能会有一些 如果展开说的话可能会有点像是程序员 但是我只能说他们在一点一点打磨自己的产品 但这里确实会遇到一个壁垒 就是你可以看到今年火的agent产品都有个特点 他都能够自己去让用户买订阅 而且他也少了Manus这样一个类似于整个厂商 OEM厂商在中间赚差价的一个点 你直接去买token 甚至你可以以更低于官网的价格去买到token的情况下 你自己去把一个agent玩的特别好 这个点其实是Manus和当代agent最大的一个区别 它模式上有一个非常大的区别我觉得 但是我相信这两套逻辑都会存在 有一部分用户他就是不想要一个本地的72小时agent 他只是想交付结结果 那么多块好省是他需要的 但另一批用户他就想完全用保 把自己变成一个像是科幻电影里的人物一样 那我觉得本地的逻辑只是弥补上了一个空白 我的观点有不同的地方在于我觉得最后都会被统一掉 确实 这只是今天不够强而已 只要模型够强 就是那畅想一下如果统一的时候是一个什么样的产品的形态 我觉得他这个真的好难 但如果让我去想的话 我们自己的定义目标就是它是有全模态的输入 也能给你全模态的回复 然后是近时时的 然后它跟你的交互会非常的简洁 只需要简洁和必要的即可 然后在它之上存在的硬件也会发生 或者这种交互形态也会有非常大的变化 就是你用你最自然的方法跟他去交付 不需要什么这个promptor and engineer area 不需要什么高级promptor 不需要不需要 然后接下来他给你回复的 他甚至可以是一段跟真实世界一样的视频 可以以他的形象 也可以是你想了解的事情 如果你让他去办一件事情的话 他可以交付给你结果 大多数时候你甚至不想关注一些 但如果你想关注一些 你也可以了解 然后围绕这外面会有一圈生态 就想想我们从copyret到cursor 到cloudcode 到今天我可能用openclaw来指挥几个cloudcode 我们在经历什么 就是外面的那一层在越来越薄 其实cursor最近发布之后 长得也和cloudcode越来越像了 然后codex大家 就有时候我感觉到logo蒙上好像都差不多 对 你怎么看这样的现象 在我个人看来是有生命周期的 你像去年 比如说3月份manus出来以后 我觉得他是定义了过去 八半年到一年的产品的交互的范式的 对吧 然后给你看操作的过程和操作结果 然后等到下半年的时候 模型能力更强了 之后你发现的范式就是 我们不care那个过程了 那上半年为什么需要 因为大家还不信任模型 模型能力还不够强 或者看到模型执行任务有一种爽感 对对对 但看久了其实你也不需要天天耍 对 就是你已经足够信任他 所以你不需要了 我只需要你给我结果 不要给我这些东西 所以你看比如说像cowork的设计 对吧 包括像cloudcode 就不会让你看到太多的过程 其中提供了一个极简模式 就像opencrawl一样 就你只需要最终一个结果就好了 甚至连中间执行的一些tool都不需要知道 对 所以我觉得这只是这个版本 这代神 下一代版本 I don't know 不断的在往更高的层级去抽象 最近二位有在B站刷这个视频的时候 看到哪些让你们印象深刻的好的AI的内容 我自己最近印象很深的就是这个 刚才聊到的花叔做的这个女娃Skill 他帧流了各种名人的skill 我发现了几个比较心动的吧 对不起叫不是up主的名字了 就是他用opencloud连上了他家的一只机器狗 对 就突然惊认到我了 给了我一个想象力 另一个可能就偏个人爱好了 就是有一类游戏叫文字冒险游戏 就你叫fate吧 它其实是一本小说的视觉画 把那些角色放到一个网站里 网站里然后他把那个角色的那个例会 就是那个拖到一个场景里 他们就可以开始对话了 相当于他就是在给你的一种最简单的游戏 最简单是文本类的游戏 的一个无限场的一个延续下去 对我最近也在B站看到一个游戏的一个新发布 是任天堂刚发了一个这个新游戏叫TOMODACHI LIVE 我看到的时候甚至认为它是这个AI游戏 终于找到PMF的第一款 它是一个什么游戏呢 就是每个人可以造一个自己的岛 然后在岛上你可以捏一个自己的人 捏很多人你可以捏自己 比如我也可以捏阿岛 然后可以捏泽英 然后捏一堆人进去之后 然后这些人在岛上就开始他们生活 然后你就开始围观他们的生活 然后你可以让他们A和B同居在一起 对吧ABC群居在一起 也可以给他们启动一个故事的第一句话 比如说这个今天泽英和阿岛在直播上大放厥词 然后他们俩就开始就演下去 然后你就可以开始去看他们各种Drama 其实是一个模拟的世界 然后我作为玩家既是导演又是观众 然后这是一个无限的世界 因为是人天堂做的所以也很漂亮 可以想象一下可爱的话 这是我对世界的认知之一 就是外星人在看我玩 如果最后到量子力学的话 你会发现好像这个世界真的有可能是模拟的 其实我也这么想 对我看到那个时候我当时也有觉得 哇这不就是一层又一层的模拟吗 对吧好像我们今天在这里模拟一个人天堂游戏 对我们也不知道是一个什么世界 对对对但What Level 只要我们开心就好 对是的是的是的 这就不重要啊 嗯 过好每一天开心就好 下整个这个Agent行业 就有没有什么观点 就关于Agent的观点啊 是今天大多数人认可 但是你们不认可的啊 我觉得有一个就是很多观点是说 人会被替代人会没有事情做 但我可能非常深的不认同这一点 嗯对我觉得人类的创造力是无可替代的 我觉得人类仍然会有非常多的事情 是可以去做而AI不能做的 嗯我仍然觉得AI是电是蒸汽机 但最终驾驭它和创造出那些美好东西的 是人类所以我觉得不用担心 被AI替代 当那个蒸汽机出来的时候 有很多人担心被抵抬 当这个电力出来的时候 有很多人担心被替抵抬 室友岗位消失了 但我觉得大家都后来从事 可能是至少我们可以说体力上更轻松 对健康来说也更好一些 对吧至少大家平均寿命一定提升了 是的这样的工作 这是我的一个信仰吧 也是我的判断 就是那个一本书叫这个人性中的善良天使啊 他在讲的就是人类总是这个很悲观 总觉得我们好像日子不如从前 觉得这世界上有更多的战乱 有更多的这个呃什么犯罪 但他就是用各种数据来告诉我们 其实人类社会是越来越好 越来越好 越来越好的 对我不知道你们有没有看过WO的那本原则三 那个他画过一条曲线 嗯是我对这个世界的认知 就是人类GDP是这样指数级的增长的 嗯哪怕是你打开20世纪有一次世界大战 20世纪的大战 嗯那一很微小的那一个停顿 都是你似乎注意不到的 当然在历史可能在呃人的身上是一座山哈 对 对但是拉长到整个历史进程上来看 我自己是相信这条曲线 当然我相信另外一条曲线就决定我当初 为什么加入这个公司的这个行业 包括我刚才看到大家有说 那要花多少token的钱对吧 有很多agent 就是摩尔定律 嗯 就它是一条你的 你获得的智能在指数的提升 但是你的成本在指数下降 嗯 对或者单位智能的成本在指数下降 这是两个我相信的东西 我会面试各种同学 他们可能都会传达一些焦虑给我 就觉得一些年轻的工程师可能会被AIT贷等等 我从这个角度去解答这个问题吧 从上个世纪开始 呃银行应该是一个比较体面的工作 但是银行也面临了一个很大的问题 就是ATM机的出现 当大家存取钱不需要有人去操作的时候 那么大家会慌了 银行是不是要裁员了 是不是有更少的人能够挤进这个金饭碗里了 但其实情况不是这样想的 因为ATM机提高了效率 所以说你可以看到自己身边基本上 不管是县城还是大城市等等 你走五公里内绝对会有一家银行 因此银行并没有变少 它变多了 那么在银行里从事工作的人呢 他们也没有变少 而是变得更加多了 那我觉得这种由ATM带来提笑而解放人 而人能去做更多事情的这套范式 一定会在当下这个Agent的时代 再次重演一遍又一遍 不管是大家就是工作了多少年啊 对不要担心自己被AI替代 如果有这个想法你真的就输了 对 然后一定要去好好把AI用起来 当做一个自己的伙伴 也可以当提套工具啊等等 也可以当做一个自己的长期的挚友等等 这都是非常好的一个思路 但是一定要尽早的拥抱AI 但我相信人不会替代 但是我当初加入这个公司还有另外一个理由 就是我当时用了 切剧GP 因为我当时正在GAP 然后帮我写一些代码之后 就那时候代码力可能没有那么强 我就意识到了一点打不过 得加入 打不过就加入 对就是还是要拥抱AI 就是相当你有了蒸汽机 有了这个电力之后你不用 对吧就相当人家用热兵器 你用冷兵器 那你肯定要被这个怼死 对吧 对所以我觉得就是这个新版本来了 还是要看一下新版本的这个玩法 是的 就是积极乐观的拥抱新版本 相信新版本不是这个滔天大浪 把我们的磨掉 而且它会成为这个让我们可以在里面 自由翱翔的一种这个舞台 嗯 对 那我们今天没想到是如此乐观的一个收尾 嗯 好的谢谢二位 这个非常开心 在你今天在B站做这一场直播 对谢谢Koge 谢谢B站 谢谢B站的朋友们 谢谢大家 谢谢大家陪伴 拜拜