E78｜In-depth dismantling of the Web3 relationship behind Hermes Agent and OpenRouter, let’s talk about whether Web3 will still have a role after the talents go to AI

Hello everyone, welcome to Web3 101, my name is Liu Feng. Hello everyone, my name is Jack. Our recording today is on the afternoon of Friday, April 17th, and the guest with us is our old friend Teacher Wang Chao. Teacher Wang Chao, would you like to say hello to everyone? Hello everyone, I am Wang Chao. Today we invite Teacher Wang Chao because we want to talk about the topic of AI again. We must state that Web3 101 is not AI 101, but we think there is one topic that is particularly worthwhile. Let’s start a discussion. Teacher Wang Chao is the best guest on this topic. Because when I hold topic selection meetings with the Silicon Valley 101 team recently, they often discuss some very hot new AI applications or products. For example, OpenRouter, and recently Hermes Agent of course. Whenever this kind of product is discussed, I can tell that both teams originally transferred from Web3. So everyone will be surprised every time. Everyone asks a question every time. Can you talk about why the team from Web3 can shine in the world of AI? These popular AI products are all from the Web3 team. Is this a common phenomenon? I think this topic is also very interesting, so I talked about Teacher Wang Chao. Because Mr. Wang Chao comes from the Crypto world, but has been watching AI for three years. So we can discuss this topic today. It must be said that at the beginning we made a very preset judgment. Many product founding teams come from Web3. Is this already a common phenomenon? Of course, we can make a more accurate judgment on this during the discussion. At the beginning, I thought we could start with some very popular products recently, but it is obvious that the team has Web3 and Crypto background. For example, OpenReuters and Hermers Agent. I know that Mr. Wang Chao has communicated with these two teams before, and he was very close to becoming their investor. Can you tell us the previous story and your understanding of the founding teams behind these two popular AI products? I have indeed been following these two teams for a long time. You may have seen, for example, that News Research’s Hermers Agent suddenly became popular, but this is not the case. He has been working in the AI ​​​​track for a long time, and he has achieved many achievements in the past that we have not paid attention to, but they are actually very big results. Including OpenRouter too. OpenRouter's story may not be that twisty, because he has been doing the same thing from day one to now. He did not say that I will release a product today and a product tomorrow. But he got here step by step, and the thoughts in the middle were actually quite interesting. I suggest that we can briefly give you some background. What are OpenRouter and Hermers Agent? The team behind them is related to Crypto and Web3. I think let’s start with the more popular one right now, Hermers? Hermers itself or New3 Search, which is the team behind it. Of course, his current popular product is an Agent product, which is very similar to the previous more popular product called OpenCloud. You can install it on your local machine, on various cloud servers, or even on your mobile phone, and connect it to large models to help you do various things. Of course, his popularity is partly unexpected, but in terms of my positioning of him, it is such a non-mainstream AI Lab team that is difficult to describe with one product. And this team has actually made a lot of unknown achievements in the past. New3 Search this community, it originated in 2022. At that time, some AI enthusiasts scattered around the world started chatting about AI and models on a Discord Server. Among them are veterans who have been playing with AI models since, for example, the GPTR era, but more of them are enthusiasts who will join the circle in the second half of 2022 with the release of Stability Fillion and HITDP. Of course, there must be many such small communities, and New3 may be just one of them. There is a buddy here named Technim. He is a pseudonym. No one knows what his real name is. He was very good at tinkering at that time, and he would work on it every time a new model was released. In addition to him, there are several people in the Discord Server, one is Karen, this guy studies religion and philosophy. He does machine learning research in a brain science laboratory at Stanford University during the day, and then thinks about philosophy at night. Another very famous one is that his current CEO is named Jeffrey. Jeffrey is also very complex. Before this, he was more famous or considered to be a native of encryption. He used to be the chief engineer of Eden Network. This was considered a top-level project when MEV was very popular. Eden is a very typical Crypto project, targeting Flashbots. It is a product of MEV. It was actually very valuable to Flashbots at the time, but Eden’s valuation was very low, that is, we always called it a low-end version of Flashbots. After a while, I lost a lot of money. So Jeffrey, I know that he has a strong Crypto background. Yes, this token has returned to zero. It is so miserable to return to zero. Really? He's at zero. After returning to zero, he later stopped tanking, and Eden stopped all its products. Because they also encountered many technical changes in the entire Ethereum, their set of things finally stopped working. But they were okay, they distributed the money in their treasury to the holders of Eden tokens. I have the impression that I also got some Ethereum back from a lot of Eden tokens. Strictly speaking, it is not a return to zero, but this product and Eden Network do no longer exist. So I still know Jeffrey. As soon as he saw me, he knew that the News Research team has a strong Crypto background. It's just such a group of people, of course there may be more than just a few of them, they are the most active ones here. This team is quite diverse, and I remember Karen as a person whose behavior can be described as chaotic or open. My impression is that she and Crypto's ideas are quite acceptable to her, and they were also in an environment like Discord at the beginning. So my impression is that they are a team that has absorbed a lot of the Karen community and Crypto’s culture, right? Yes, we were actually an interest group before. At that stage, it was actually difficult to call them a team. They have been playing with some small models of their own, and there is still a huge gap between them and the real business model. A truly landmark event is March 2023. First, in February 2023, Meta released its first Karen model Lama. A week after the release, the weights of Lama's model were accidentally leaked. Suddenly they had a situation that was really at least close to commercial use, and a model could start to be tossed around. A month later, Stanford released a model called Alpaca. Lama is the llama in South America, and Paka is the little alpaca in South America. How did Alpaca do it? They used GPT3.5 to generate more than 50,000 pieces of data. The so-called data was actually just questions and answers. They used these more than 50,000 pieces of data to fine-tune the Lama model. As a result, after fine-tuning the model, its performance was close to GPT3.5. This paper was actually very sensational at the time, not only because it produced a model, but more importantly, it proved a new paradigm, that is, I can use high-level models to steam and retain data, and then use these data to train small models. At that time, after TechNim read this paper, his first reaction was that I only spent 600 yuan on this thing at Stanford. He used GPT3.5. So if I use GPT4, what can I get out of it? So he started working directly and used GPT4 to generate a large amount of so-called instruction data. The quality must be significantly better than that of GPT3.5. By the end of April, when he finished the data set, he published it on Hugging Face and named it GP Teacher. Then on June 3, when he used GP Teacher to fine-tune a new model, he released it and named it Hermers, which is the first version of Hermers. This model was very easy to use at least at the time, so it was on the Hermers open source model rankings a few days after its release, and it became popular immediately. So at this time, Teknip alone was already too busy, so he brought in people who had been discussing these topics together to continue working on the next edition. At this time, for example, Karen also came over to do some data sorting and other things, including providing some community resource support to provide them with some computing power for missions. At that time, News was not actually a team, and it did not have any income. In fact, it was still very much a community collaboration. At this time, they were actually seen by HLZ, and HLZ gave them a Grant. The amount was not large, but this amount was enough for them to stop their own work and start focusing on News. On August 30, 2023, they officially registered a company, News Research. The co-founders are the people we mentioned just now. They either have a heavy encryption background, or they also have some encryption-related genes here. It’s very interesting. One day after the company registered, they published a paper called YARN, Y-A-R-N. Because they were fine-tuning LAMA themselves, they found that the published LAMA actually only had 4,000 Token contexts. We are eager to have a Million context now, but we felt it was not long enough. At that time, there were only 4,000, which was not enough. At that time, Jeffrey was studying this thing, how to extend the window, so based on some existing solutions, they made a lot of mathematical innovations, and finally succeeded in expanding the context window to 32 times the original size. They published this set of code, including methodological methods, on the Internet, and then named it very casually, YARN's Y-A, Y-A is YET, and A is ANOTHER, so when translated, it probably means that another wrong diagnosis was made. It was a very random naming. Of course they were very proud of this thing at the time, but I think they were far from expecting that this thing would become so important. Later, YARN became almost a standard in the open source model industry. Until now three years later, if you go to Deep Seek, you go to Qianwen, you go to Lama, you go to almost all open source models, you will find that the things they cited in the long context may be running YARN. Or a certain variant of YARN to support their extended context, which means that this thing has become almost a standard in the open source world. We can also read some information from this, that is, although today we feel that Hermers Agent has exposed the News Research team, but in fact this team does not have nothing. In the development process of the earliest open source models of AI, they have the position they should have. Yes, of course they were the first to become popular. It’s not because of this, it’s because the Hermers model is easy to use, and the Hermers model actually has values. This value is, you shouldn’t censor me, because you know these big model companies, for the sake of safety and so-called alignment, actually designed various things. When you ask this, they don’t answer, and they don’t answer that. So when they were doing fine-tuning training, they were actually aware of it, and used it in technology, and dismantled this thing. Of course, it is difficult to say that the demolition is completely clean, but at least the Hermers model is actually a very open model that supports users to do whatever they want, ask questions according to their own needs, and get results. Then, because of its achievements, whether it is his open source model or the expanded methodologies he provides, it has attracted the attention of capital at this time. In January of 2024, they received a seed round. This seed round was Distributed Global and OSS, both together, had more than 5 million US dollars, and of course there were some angel investors, including Alex from OpenRouter. But at that time, it was actually a very, very small company. Everyone was working remotely and still relied on Discord for collaboration. But as I mentioned earlier, Karen’s background is in religion and philosophy. As you mentioned just now, Mr. Liu, he has a lot of thoughts other than technology. In March 2024, they released another product called WordTheme. Its interface is a command-line web interface, and the bottom layer runs the Athorpic model. This thing is actually the product of a very exquisite prompt sub-project, that is, users can give it instructions using natural instructions. For example, if you simulate what it would be like in 2000 when the industrial revolution never happened, or simulate what the earth looks like now when dinosaurs are not extinct, Cloud will expand these scenarios into a stable world based on such an assumption. You can talk and move around in this world. You can even fork it if it fails, but a very interesting thing, WordTheme, exploded, so it was actually a very good star product in 2023, and everyone started to have a great time, but this usage triggered Athorpic's content alert, because some simulation scenarios may exceed the boundaries that Athorpic can accept, so one month after its release in April, Athorpic cut off all Cloud API access to News, and he was forced to go offline. In the end, it was deleted and relaunched, but in fact, its appeal dropped significantly at this time, so in the end, the product achieved short-term success, but in the end it was abandoned. This team originally had the ideological genes of decentralization. After being done like this by Athorpic, I think they may feel this more strongly. This product actually lived for a month and then disappeared, but in fact, a lot of things happened to it unexpectedly, and it has nothing to do with News. This thing in itself is a very sophisticated methodology. This methodology was later watched by a buddy named Andy. He focked this thing and changed it. After the change, he used his set of prompt words to let the two clouds chat with each other 9,000 times. In the process of chatting, a virtual religion was invented. This virtual religion eventually developed into a tiny robot called Terminal Truth. Terminal This is how the truth came about, yes, this is how it came about. This robot inadvertently defrauded the founder of HLV of 50,000 US dollars, which is equivalent to a grant in Bitcoin. In the end, this matter was amplified by the meme community, and finally turned into a gold meme coin, which may reach 1 billion US dollars when it reaches its highest level. Such a thing, but these subsequent stories have nothing to do with news. But the source is indeed their word theme, which is a very interesting thing. Yes, this actually shows that they are very good at making such popular products, although it may not be popular for a long time. In addition, their popular products can also inspire more people to recreate in an open source way. Many things are actually the source of inspiration for them. Yes, so they are already well-known in the entire circle at this time. We talked to him. In fact, it was also at this time that we discovered that the word theme was very good. I checked later and found out that this team had more unknown achievements here, so I quickly asked them to chat. It was May 2024 when we made an appointment with them. In fact, it was a bit late at that time. Their round was close to close, but they were still very encryption-friendly, so knowing that we were actually people with a strong encryption background, we still had a good chat, but it was too late. After the chat, because he was already very popular at that time. So its valuation was quite expensive, so we hesitated for a while, but finally decided to invest, but after hesitating, it was so full that we couldn't get in, so we missed it. The last announcement was in June 2024. They completed those companies from Delphi and took an additional round of 15 million US dollars. After that, they had more money, so they continued to do it. Modeling was their root level, so they still followed Lama step by step, because Meta was still OK at that time. Still an eye-opener, I kept updating my Lama model, from 3 to 3.1 to Lama4. The last crash was Lama4, so Nius also kept updating with Lama. But during Lama3, there was actually a pretty big thing. Lama3 actually released three versions, namely the 8B version, the 70B version and the 405B version. At that time, the 405B version, In fact, the community has some different reactions. The first is that the 405B version is very powerful. At least in Banchmark at the time, it is actually very close to the B-yuan model. But from another perspective, for a model of the size of 405B, in fact, for an open source team, or for the community, you have no way to fine-tune it. You use traditional fine-tuning methods to adjust it, because it is too big. If you try to adjust it, you will find that it has no effect. But if you are going to carry out a large-scale full-parameter fine-tuning, the resources required at this time are also very high, and the ordinary community has no way to do it. So Nius made a big announcement at that time, saying that we were going to carry out full-scale fine-tuning of Lama3.1, the 405B version. Of course, everyone did not believe it and thought you should stop talking nonsense. As a community team, I think it is impossible for you to accomplish this, both in terms of engineering and resources. But those who questioned were quickly slapped in the face. It didn’t take long for them to come up with the fine-tuning model. I think the model itself was a matter of political significance or ideology. I think this model may not have gotten better after fine-tuning, but at least it proved that our community, as a force, is fully capable of doing this kind of thing. You professionals look down on us and think we can’t do things, so it should be said to be the 405B model at that time. The whole parameter fine-tuning itself was a bit shocking to the industry. At that time, I thought they actually had some thoughts and routes of their own, but this may be left to us outsiders to guess. Not long after they completed the fine-tuning, they began to put some of their energy into a new thing and released a report called Distributed Training Over the Internet, which is distributed training across the Internet scale. The core point of this report is that they found a method. Because we know that when training large models, the requirements for bandwidth are actually very high, because every time you update the training parameters, you have to do backpropagation. These things actually require a huge amount of communication. So we know that in those data centers, NVIDIA's Infinity Band is used, which has a communication volume of hundreds of quarters per second, or even higher. This means that even if you have enough GPUs, if they are not connected at high speed, you will not be able to train. But what did Nius do? They found a way to reduce the amount of data that needs to be synchronized when training a large model by nearly 1,000 times, and its training quality is almost unaffected. This means that when you train cutting-edge models, you no longer need this large-scale, high-speed bandwidth network of tens of thousands of GPU clusters. You may have GPUs scattered around the world, and ordinary home bandwidth can support them to be connected for training. This is very crypto, right? Yes, yes, I will immediately think of it. The story of DPing will appear immediately. Yes, the story of DPing is actually closely connected with this. If this is not true, DPing can only do some things such as reasoning. If this path is proven to be effective, what you can do with DPing will definitely not only be reasoning, but also overcome the final training step. So Distro was built with just such a thing. A few months later, News wrote the core algorithm of Distro into a formal paper. This paper was posted to archive, a preprint website of the paper. After the paper was posted, everyone was stunned when they saw it, because in addition to a few core members of News Research, the list of authors of the paper includes a buddy named Dark Kimmer. He is the inventor of the Adam optimizer. Nowadays, when training any neural network in the world, the Adam optimizer is almost always used, so he is at Google. The total number of Scholars citations is hundreds of thousands, and he was one of the 11 co-founders of OpenAI when it was founded in 2015. He is such a prestigious figure. He recognized the importance of News Research’s work and the advancement of technology, so he may also be interested in this development. We don’t know the specific process, but in short, he chose to participate in this project at such a stage. And the last paper also has one of his signatures, so from this perspective, we can definitely see News Research. In fact, some hard-core work is very valuable. From the YAR mentioned just now, to their distributed training framework now, this is a very hard-core, very hard-core team. Let us continue to sort it out. They continue based on these things. The first is a lot of their own training. Now they can rely on this framework to do it, so the following should start from Hermers 3. At the beginning, their models should have started to use their own training framework and some distributed GPUs, and these were getting better and better with the evolution of open source models. By the end of 2025, a relatively new model was a 36B model. This model used a base model of ByteDance’s open source. This base model actually has a milestone significance, that is, it serves as a reference. The performance of the version of the model they trained using the neutral network exceeded the control version using centralized training. This is very powerful. In other words, not only can this methodology be used, but at least in certain scenarios, the things it trains are better than centralized training. This is a very cutting-edge and very powerful proof. It proves that if one day, we may have to break the monopoly of the giants' computing power, then the distributed GPU network can also train competitive cutting-edge models. Next is the final story. Now everyone knows that Hermès became popular. It was in early 2026, so it released such an AI agent with a self-learning function. After lobster became popular first, everyone later discovered that lobster has a lot of things that may not be useful. They found that this thing is very close to lobster and is very easy to use, so it has become a popular frying machine in the past few weeks. This is also a real consumer product, so maybe we have this discussion today, because on the C side, It actually has a really strong flagship product, which is very clear. I only knew briefly that they have their own open source model before, but I couldn't understand some of the remarkable contributions they have made in the AI open source community. This team is actually a very comprehensive team. On the product line, they have open source models, such as Hermers, and they also made their own open source training framework, which is Distro. In the end, what is popular now is the open source intelligent agent framework for C-end consumers. The Hermers Agent is quite comprehensive. Yes, it is very comprehensive on the one hand, but it sounds a bit complicated. The model is also made, the Agent is also made, and the training framework is also made. But in fact, if you look at the entire line from end to end, you will find that the ideological aspects here are actually the same thing. For example, why Hermers became popular the first time, it is not just because his model is very effective. But this model has chosen a different path from the big field, called neutral alignment. This model does not presuppose its own position and does not review. It is loyal to the user's intention, rather than the ethical principles of a certain commercial company. There is actually a lot of philosophical thinking here, which is one of the reasons why it was so popular in the community in the first place, because they believe that the sovereignty of users is higher than the ethical rules set by the company itself. In my opinion, this is actually a very strong ideology, and it is the same thing as many things that the crypto community believes in. What's next, we just talked about it for a long time, why he did full parameter fine-tuning based on 405B, he wanted to prove that the power of our community is actually enough, not what you think. In the end, he did this decentralized training, which is actually very strong. There is a hidden danger in his whole process, for example, when his Hermers went from 1 to 2 to 3, he actually used the open source Rama model before. There is actually a problem, that is, if one day Meta decides to avoid it. What should I do? I don't have a good base model to build things based on. At that time, China's open source power had not yet risen. At that time, basically the whole world was relying on Rama, and it was iterating board by board. So at this time, I think he had this ideological thinking from beginning to end, and he was also crisis-conscious. Then one day Xiaojia said that we must go away, and he must go away now, right, the latest version, or one day he changed the license, so what should we do? What should we play? We cannot have tens of thousands of GPU clusters to maintain it like those big companies, so the only thing I can do is to pool resources from all over the world so that our community can also do this. So this is a deep reason why they decided to use the Yinghua training framework. In the end, Hermers Agent seems to be a C-end product. In fact, Hermers Agent hides something very, very deep. As mentioned earlier, it has an infrastructure that can train models without relying on giants. But from the end of 2024 to 2025, when large models become important, pre-training is of course very important, and later RH training is actually also very important. RH training requires a lot of behavioral data, so these large model manufacturers can spend a lot of money and do a lot of formula-based things, whether it is the generation of real data or the high quality of synthetic data, but the community actually does not have this ability, so the Hermers Agent actually has to take on something that no one has noticed. It actually has a deep RH training framework built into it, that is, all of our Hermers users, all your behavioral data can be directly exported into training data organized in a professional format for large models and RH training. Of course, because it is open source, it is also purely ontologically deployed. It does not collect it, and it cannot collect it even if it wants to. But at least he put this technical vehicle here first. When I need to collect it one day, will I be able to collect it in some way? To motivate and mobilize some community members to be willing to bring my own data, even if 95% are not willing, maybe 5% is still a large data set, and it is dispersed enough, with enough various possible weird scenarios. This implies a style of distribution, which can be seen in the future market. This is very crypto. Yes, I am also using hermothagen now. In fact, each of us has the potential to become a contributor to his training in the future. It is conceivable that if he introduces an incentive layer in the future, then this thing can form a data market for reinforcement learning. Yes, he has not done so now. Yes, your review is very exciting, because I have seen this news research before, but my attention to it is purely because they raised a large round of almost 50 million US dollars last year, and it was invested by paradigm. And it was clear at that time that they would issue coins, and one of their highlights at that time was that Distroy, the decentralized computing power and training system you just introduced, they actually wanted to build a network on this, I remember it was a network called Psyche, and put it on the Solana blockchain, right? Yes, they have actually made this, right, because this is equivalent to a crypto product, and I have the impression that this thing is going well. Because he is and is considered by everyone to be a killer of Beaten Sum, a core competitor, so when it comes to news research, my first reaction is that it is a team in the currency circle. Yes, I think your summary is correct. Maybe our topic today is how can the encryption team enter the AI field and achieve good results? In fact, news research, he does not go into the AI field, he is still a cryptography team that is aiming at the fundamentals, but he separates the two sides very clearly, I am consumer-oriented, I am oriented in the AI field, I will not mention tokenomics, I will not mention blockchain, I will give you a very good product, I will give you a very good algorithm improvement, I will prove myself, if one day I need encryption, no matter I have needed funds here, I still need some initial market support here. This is a matter here. I will tell this group of people, but I will not tell you. He is very divided between the two sides, yes, so my impression is that wherever his coins can be issued, Sychic must be a way to issue coins. Recently, after Hermas Agent became popular, I heard many crypto friends say that after this team and this product become popular, they may not issue the previous coins. But listening to what you said, I think actually this team, From its fundamentals, it is actually a team that believes in crypto, decentralization, open source and community contributions. On the contrary, it is conceivable that in the future, it will have many methods that may surprise everyone, and it has buried many interesting clues. This team is completely beyond my expectations. It can clearly and beautifully demonstrate its advantages in both the AI field and the crypto field, which is great, but when it comes to issuing coins, I am really not sure whether it will. I know that there are different opinions on this within his team. If he is so successful now, there is no doubt that he may raise more money in the future, whether it is capital from the encryption circle or from the mainstream AI circle. If you see that he can be divided so lightly now, it is because he clearly knows what is right and what is wrong. If I make something, if I integrate encryption into it, I may lose some more mainstream users. I definitely don’t know about the issuance of coins, but if they haven’t thought about it clearly, I don’t think they will do it rashly. Of course, today I feel that the world of crypto web3 is a bleak environment, but AI has a prosperous heart. Discussing their coin issuance today is indeed a matter that goes against Tianzang. Who knows that the world will still move forward. At least when I saw them discussing financing last year, it was very clear. Their valuation is based on tokens. The good thing is that this team is really delivering. Another good thing is that they actually have very strong ideas. They actually believe in the community, open source, and decentralized culture. And they know that they have a very beautiful way of playing. In fact, there are great possibilities. In terms of this technology, we can actually take a quick look at openreut. Compared with openreut, their background and their development, It should be more simple and clear, because his founder, in my impression, was in 18 years. As an important member of the founding team, he joined openc as CTO. In 18 years, openc was unknown. Later, in the NFT cycle, openc shined. However, in my impression, Alex, he should be in the middle of 22 years, which is the time point when NFT is the most popular and going down. He has left, and he founded openreut 23 years ago, because Mr. Wang Shang should have communicated with them. In fact, I am very concerned about when did you pay attention to openreut, and why do you think this thing is worth paying attention to, because the phenomenon I have seen is that because of the popularity of lobster, everyone suddenly began to pay attention to large models, especially the business of large model APIs. At this time, openreut suddenly appeared in everyone's career. But I believe that it has actually become a very important infrastructure very early, because I have seen A16Z talk about it several times before. Through business and data analysis like openreut, we can observe large models and large model companies, their delivery capabilities and the attractiveness of their users. Yes, I noticed openreut very early, because when large models first became popular in 2023, there were many APIs that were very troublesome to remove. For example, if you wanted to remove the XGBT API, If you want to solve the payment problem, they will give you a layer of skin. Of course, there are many domestic exhibition sites. In fact, you find that the experience is not good enough, and there are even some fake models in it. Therefore, Openreut was a company that provided a neutral third party at that time, which was already very useful. So I started paying attention to it probably at the end of 23, and I also looked at its volume, because it has a ranking list, and its volume is always announced. I can also see that its volume has been rising, so I also learned that Alex is behind openreut. Isn't this the guy from OBC back then? So I thought, hey, should I ask him to chat, so I found a lot of people, and finally I made an appointment. But after the chat, people asked me, but they may not be too interested. At that time, it was not raising funds, so it was left alone. Later, it actually became a very important ecological niche. Because this essentially provides a universal model calling layer, its position in the circle is actually a very important one. You can see that for example, XI GPT and GPT 4.1 were first released on openreut with a fake name. , let everyone guess, and finally guessed that it was the latest model of XI GPT, including this year's Xiaomi model. Two models were also put up for everyone to guess. Everyone knew it was a Chinese model at the time, but no one guessed it was Xiaomi. That is to say, it is actually very important in the ecosystem now. It is an important distribution layer. Yes, yes, the distribution layer, for developers, It is also a very important and very stable universal model capability layer. You don’t need to worry about the various frictions and various messy things in different models. When you go to it, there is a set of APIs, just one interface, just one key. You can accept whoever you want, and it will never block you. I don’t think there are too many twists and turns in the story itself. It is that it has been doing this since day one, and it has been doing this since childhood, and it has become an industry leader. I think it’s okay to say that it’s a giant, so the story itself is very bland, but if we go one step further and look at Alex, I think this is quite interesting. He recorded a podcast with Bankless. I think the reason why he was able to turn around like this is that he jumped from web3 to AI in an instant, and he did it so well. The first thing was that when he was doing encryption, He has no ideology, unlike news-ray-search, for example, he has a very strong ideology, and these things are very strong here. In his opinion, what do I do in open-sea? I am not doing NMT, I am not doing these things that change the world, I just sort out these messy things in the NMT field, which are complex, different formats, and represent different meanings. In his opinion, these things are not NMT. These are just heterogeneous inventories. I made an aggregator out of these heterogeneous inventories. NMT is just something he happens to aggregate. Yes, NMT is just a presentation form, just something that everyone feels. But in fact, what he processes in open-sea is a large amount of data and the processing behind NMT. The processing of dataset, yes, yes, so in his self-perception, this is also his own words. He said, I am actually just doing the aggregation of heterogeneous data. NMT is heterogeneous data, so now there are various different models, each manufacturer's own model, these very specific parameters, and then the calling method, API format, and billing method are all different. Isn't this the same thing, right? Let me do it here. Yes, this really clarified the core points of his two jobs. In fact, it is the same thing. Yes, it is very interesting, and the reason why he made up his mind to do this, From the source, the impact is the same as news-resource. News-resource is that the guy from TechName saw Lama and came out with a fine-tuned version. He said, wow, if he can do it, I can do it, so he did it. After Alex saw it, his reaction was different. He said, wow, if you can build a fine-tuned version for 600 US dollars, then there will be thousands of models in the future. Then this is my business. I can aggregate these thousands of models through a set of things. So for the same thing, different people judged different things, did different things, and finally achieved great success. But he has a very interesting habit, that is, why he can capture these things. He actually also has a strong community-oriented thinking. His community-oriented thinking is to identify why these strange communities, such as OpenSea at that time, were able to be created. Because they discovered that at the end of 2017, why suddenly there was a group of strange people. Spending so much money to buy a picture of a cat, what is this, why are they doing this, so he went to this community to study this thing, and after researching, he found that if this thing can grow, then they must have a layer of things on top to serve them. This is actually the model of OpenSea, but now he found, wow, there are so many weird communities working on various models, including many large model companies, what can I do to serve them? This is another one of his, which can be regarded as his own methodology. He first identifies this weird new thing, sinks in, sees what everyone is talking about, sees what everyone is doing, and then thinks about a question, that is, if they can make this thing bigger, what do they need? Finally, based on this methodology, two products were made. These are some connections between OpenReuters, Crypto, and Web3. I don’t know if Jack has anything to add, or Jack, do you have any questions to ask? Yes, When I read his story on OpenReuters before, I thought that maybe he was still in the encryption circle and was not very representative of the Web3 circle. Then he came to the AI world and made a big splash. This is a typical example of a technical man. He understands everything, whether it is Web3 or Web set, as solving a particularly deep and practical technical problem and seeing a business prospect from this problem. Yes, and then AI gave him a larger platform. This reminds me of a video I saw of CZ before. It may not have anything to do with AI. He said on a forum why he wanted to do Binance. He said he had no choice because he had been working on a trading platform before this. He felt that he was a hunter who had been trained many times. Then he felt that when this target appeared, he must cancel the flight. I think there is a similar logic in it. For me, it is a realization, that is, we should not be so licking our faces. If someone succeeds, they will immediately jump out and say, you see, they are from the currency circle, from web3, and people from web3 will have an advantage in doing AI. This judgment is actually too one-sided. It is very arbitrary and unscientific. We should blush. I think what Teacher Liu said before we started recording is indeed quite right. Maybe many crypto practitioners, or people in the currency circle, will be more sensitive to using AI or discovering opportunities in AI. This reminds me of the extraordinary cloud code leaks discovered two days ago. The founder of Fuzzland actually has a crypto background, and Fuzzland itself is also doing security in the crypto field, so I think there is still something there. Chaofan is also a very cool geek. Why do I want to give a question that has been asked before? I think it is difficult for us to draw conclusions because some popular AI products are made by founders or teams who were previously related to the currency circle. To judge that web3 has an impact on AI, or that web3 people actually have certain advantages, but I think there are some facts that we can put forward. After all, in the last cycle, especially the cycle around 22 years, Before the explosion of AI came, crypto and web3 attracted the best young people. When these young people were at that time, they bravely jumped into the tool hole of crypto to try, and then they saw some more interesting things, and they switched to AI. This is a development process that is easy to understand, so you may indeed see that many people who are currently in the AI ​​field and constantly trying new things have switched from the previous round of crypto world. Another phenomenon is that you will find that because crypto is not a field that is easily accepted by the mainstream for a long time, it will attract many people with very open ideas, and may even be able to oppose this tradition, and they will bravely try it. Therefore, many users and developers in the crypto world have a strong sense of adventure and open thinking. So you will see that in the application of many new AI products, friends in the crypto world are using them very well. I will find that some friends in the web3 world will quickly embrace OpenCloud and Hermers Agent, and have already started trying them before they become popular. I think this is all due to some endogenous reasons. But I also think that we should not jump to the conclusion so quickly, that is, people from web3 can still shine in the AI ​​world. I think this judgment is indeed premature. But I think we can also take a look at other teams or entrepreneurs who have a background in web3 and are currently doing well in the AI ​​world. My impression is that the founder of MultiBook, a product that allows agents to socialize, must have been an active member of the Bitcoin community for a long time. He has also made things like MemeCoin and was quite active in the crypto world before. Another point that I can think of is Amert, the founder of stability AI. He seems to have been in the field of crypto before, and he has been involved in it very early. Amert is actually quite involved in the field of encryption, but he is not very well-known, and the earliest stability AI almost became a Tao, but fortunately he did not become a Tao, and if it did, it would not be so successful. Stability AI, what kind of influence does it have in the AI ​​field? In addition, I think you can talk about the background of its earliest founder, Amert, in crypto, and its impact on this company. Of course, Amert should have been kicked out today, and he is now back to web3. It's just that Amert himself is very controversial. For example, the founder's disputes, for example, he talks openly and often brags. Yes, dishonest. Yes, these are the problems, but in terms of past stability, I think his historical contribution in the field of AI, especially in the field of AI for practitioners, is very great, and it is not an exaggeration to say that he is iconic. Then he went over there. In fact, before he entered the AI ​​field, he was working in finance. I don’t know much about that history. He also tried to do an encryption project related to public utilities, but that project used encryption technology, but no coins were issued and it was not completed. But in fact, he is very familiar with encryption technology and the ideologies of encryption. So when Stability was founded later, because Stability First was also a community-based startup, it was also a thing spread by contributors from many places around the world. So at that time, he thought that this should actually be a Tao, or even a Tao of Tao, that is, Stability will support various open source AI communities and do many things. He has told this story more than once. In the end, he said why he didn't do it because he felt that the recording facilities were immature. I think his observation was correct. Fortunately, he didn't do it, otherwise there might be no stability. As for Stability, I think his most important contribution is that he open sourced his fist model in the summer of 2022. Now we think that open source models are something that should be taken for granted. There is a school doing this. In fact, it was not the case before. There is a huge debate in the industry that we should open source models. What are the legal issues if I do it? There are all kinds of things. Anyway, Stability is not afraid of wearing shoes when he is barefoot. Anyway, at that time, of course it may be a business strategy of his own, and it may also have some will-shaped things, so he open sourced this thing. In fact, the wave of AI was earlier than Chai DPT. There were many geeks before Chai DPT, and many people who were interested in this kind of image generation were already having fun playing it. I think it all started with his business decision to open source this model. After real AI came to life in 2023, he was also on the main table at that time. I remember that there was a president in the United States, or a dinner in the White House, or in Congress, and maybe about 10 people were invited, including Elon Musk, Sam Altman, and among them was Ema. So this is a man who has been on the main table seriously, but this latter thing is not sustainable enough, and there may be various problems that we don’t know about. Slowly he left the poker table and came back again, so he was not actually a person who ran from the AI circle to the encryption circle to work on a project. He ran from the encryption circle to the AI circle and made some achievements. However, in the AI circle, there was no Great Wall General, and he couldn't do it anymore. He came back to think about other things. What is he doing now? I remember that what he is doing now is related to the technical facilities of decentralized AI. The last time I paid attention to him was about two years ago, because at that time he claimed to have resigned from Stability, but in fact he might have been kicked out by shareholders. He just didn't admit it. We made an appointment with him immediately after he was kicked out, because his earliest investor had a familiar friend who was very familiar with him, so we made an appointment to chat with him. At that time, he was bragging a lot. Just tell us about sovereign AI, not your mind, not your model and so on. The general meaning is that I want to build a set of technical facilities. I support any country, or any organization, or any group in the world to believe in its own model. He told such a story. After he started his project, he was still in the same direction of the story. The details may be slightly different. But later on, he seemed to have really not done anything, and the feeling disappeared. I can remember that it was him who wrote the book. It’s called The Last Economy. It talks about the scarcity that AI breaks and makes everyone’s life rich enough. Please give me a sense of the future. But I haven’t read the book, I only read some book reviews. In addition, I still have the impression that he has been very close to the Render project. Render is the team that does decentralization and GPU rendering. He has been a consultant in this team. In addition, I can see that he often talks about open source and decentralization. I haven’t seen the product for these discussions about giving sovereignty to everyone, but I can be sure that it is actually in the world of Web3, making some attempts, so Stability AI is indeed somewhat related to Crypto. We just talked about the connection between Hermers Agent and Web3, and the connection between OpenReuters, briefly mentioned MountBook, and then also talked about Stability. In addition, I can see that what everyone often talks about is actually FTX, the FTX launched at that time. Future Fund hired some very cool young people, who are now shining brightly in the AI world, but I personally think that these people are not closely related to Crypto. This is because in 2022, because FTX had money, he could spend money to find these cutting-edge smart people. Whether it was the senior managers he hired or the young people he recruited from Future Fund, they were all on their own. What is more famous today is that he later went to OpenAI. Later, I wrote a long article, which was so insightful. Because I wrote this long article, I got a lot of money from big names. The young man who built a situational awareness hedge fund has a very long name, Leopold, and Ashbrenner. He also made a lot of noise in the media some time ago. This guy came from FTX Future Fund, and later went to OpenAI, and later started a hedge fund, which was very popular. Another one is Leopold’s girlfriend, Avital. He is now the Chief of Staff at Authropic, and he was also at FTX Future Fund when they were together. People often regard these two as the craze of Crypto to AI, but I don’t think this is particularly accurate, because at that time they were smart young people, and because FTX gave them money, they worked for them to invest in the future. And Narrative’s story told also very attracted some young people with ideas, and FTX has been telling it. That is to say, I want to be altruistic, advocate open charity, make long-term investments, and invest in the future of mankind without any return, which will definitely attract a group of young people with very avant-garde ideas, so I don’t think they are real geniuses from Crypto. In fact, I think there are still some companies that have come out of computing power to AI, like CoreV, that’s another story, that’s right, if we don’t talk about Web3 but encryption, in fact, this is an elephant in the room. We shouldn't ignore it. It's due to fate. They actually have a structural impact on the current AI landscape in the United States. If this group of people hadn't been here to get the approval and some preliminary construction of these electricity in advance, the electricity shortage in the United States would be much more tense than it is now. You may not be able to build anything before 2030 or even 2032. Maybe a month or two ago, everyone's attitude towards these computing companies switching to AI was still, You see, these companies tend to follow hot trends and play with hot topics, but there are still some that have emerged. They have made great contributions to the AI infrastructure in the United States. I think CoreV is really shining now. When we were sorting out related daily hot topics, it often appeared. Before, either Meta wanted to acquire it, or Jane Street seemed to have invested 6 billion US dollars in it two days ago, and spent another 1 billion US dollars to work on it. I just want to use this CoreV to do some quantitative training or business. I think this is indeed a part that is closely related to the fields of encryption and AI. You don’t care whether they are chasing hot spots or not, but in fact everyone has completely underestimated their importance. They are definitely not what everyone imagined. A group of miners are now popular in AI, and they come to transfer AI to transfer AI money. I think they occupy a very important ecological niche in the entire, or at least in the field of AI computing power in North America, and they occupy this ecological niche not because of them, because now that AI is popular, I transferred. There is of course a historical coincidence here. For example, in 2022, Ethereum Morge was converted to POS, which immediately freed up a lot of resources. Then two months later, GPT was released and connected, as well as large-scale power construction, approval of long-term power supply agreement, coordination of power grids, including the construction and rapid deployment of data centers, and the management of large-scale GPU clusters. This is actually a very, very professional skill, which means that these companies have produced many, many professional talents here in batches, and they only need a little training. Some specific things can be used by converting them into AI methods. At the same time, some of the electricity they negotiated during that historical period, such as construction approvals, is very valuable. I remember Tiangui Valley 101 said before that some estimates are approximate. They have about 6 to 10G watts, but there are more optimistic estimates, saying that it is about 20 to 25G watts. One gigawatt is almost a large nuclear power plant, which is about the same size. This is already a huge amount of power capacity. If there is no such power capacity here, it is almost ready to transfer to AI. Now we may use Cloud quota every day, we use CHATGPT, and we use GemNet. It may not take three minutes to tell you, I’m sorry, what is the quota? In fact, its strategic significance is very, very high. We don’t care whether it is for commercial reasons or for whatever reason. They just want to make money. But they actually played a very important role in history. That is to say, whether they were chasing the craze at the time, or when history was turning, whether it was Ethereum's POW to POS, or the Crypto cycle, which forced them to switch to AI, their role is obvious, especially today. In fact, the construction of AI computing power centers in the United States is not as simple as everyone thinks. You can just start building it with money. No matter the power supply problem in various places, various local community politicians have different perceptions and existing games on this issue, in fact, these computing centers from Crypto and mining have also supported some AI computing power, so this is a positive impact, yes, yes, not only a positive impact, but also if we are promoted, because these skills are really useful, unlike us in the Web3 project, we thought about it, After thinking about it for a long time, I can’t seem to think of anything that has advantages in the past. This can really be used. This is the ability to actually deliver. In short, we have sorted out what we can think of, the talents, companies, technologies from the Crypto world, and even the contributions of electric energy to the AI field. Then we can return to some relatively larger discussions. I really want to ask Teacher Wang Chao, because behind what I just mentioned, there is actually a phenomenon that cannot be ignored. In fact, Web3 talents are flowing to AI, whether they are seeking profits or chasing hot spots. The more important thing is that Web3 or Crypto used to be Olympiad gold medalists. Researchers from the computer science department and the cryptography department of the top universities, these smartest young people, Crypto is one of the best choices. However, when we are doing this program today, it is obviously not the case. The best choice is AI. How do we see such changes? It is not heavy. It seems to be a very natural thing, because the encryption world has now developed to such a scale, how many people here do you think are really coming to this field because of their recognition of cypherpunks and decentralization? Everyone comes here, most people come here for profit, and of course there are several benefits here. Some people think that it is actually very easy to start a project here. If I think of a good thing here, I will make it, and I may get rewards very quickly. Then more people actually come to participate in this speculation, or there are many people who come here to cheat. No matter they are outstanding talents who do not have such underlying values, of course we do not mean to deceive here. If you think encryption is good, you are a talent. You come here, now the encryption market has reached a bear market, AI is more fun, you go to AI, this is a natural flow, so there is actually not much depth to explain things, because I think in the past, everyone thought that the fast pace of encryption and the fast pace of AI, these two things are not the same fast. In fact, encryption is a very slow thing, because what you fight against in encryption is the system, and if you want to make something in AI, what you fight against is only engineering, and what you fight against is only some technical things and some product implementations. So from this point of view, encryption is extremely slow, but AI is extremely fast. The reason why encryption makes people feel fast is because it is too easy to make money in this circle, it is too easy to cheat money, and there is too much money. The money-making cycle is fast, but the delivery cycle is very slow. Popular products are very fast, but they are eliminated very quickly. I really want to make a good Crypto or Web3 product, not just a three-minute hot product. As for Crypto, a product that can issue coins and make money quickly, in fact, this cycle will be much slower, so first of all, I think, why should we discuss it? This is a fact, and several old friends have written long articles in various places, calling for Web3 to retain talents. I all think, hey, I understand this idea, but I think the flow of talents should be free. When the AI world is really delivering, it is really constantly and rapidly iterating technology. A large number of new applications and new ideas are emerging, and users can really feel the real impact on our lives, our work, and many societies. I think it is really good for talents to stay there. On the contrary, in fact, in the world of Crypto today, you will see that it is more emerging in the combination of traditional finance and Crypto, which has its own charm. I think what Teacher Wang Chao just said really touched my heart. We may feel a little sad when discussing this issue today, but the good thing is that I still believe that because Crypto itself, especially Crypto and Native, are native to these ideas, I believe that it has its very strong and solid foundation of Solid. These things, at a certain time, in the appropriate historical cycle, it will definitely bring huge impetus. I believe that by then, the talents will be retained. Of course, the premise is that we can really see the progress of breakthrough technologies and the huge conflicts in the external world, which may make the entire society around us re-recognize some social and cultural changes. I think that at that time, maybe we will look at the choices made by talents, which may be more meaningful. And now, we should be for these, from Web3 to the AI world. I applaud these talented people who are generous and brilliant. I think what Teacher Liu said is very good. Let’s not talk about the will form and so on. The successful people I see are actually a group of people with strong curiosity and very good construction ability. They move between different technology waves. One day the market may pick up, and one day we may find something more interesting. I believe that many talents will come back and build something more interesting that we can’t imagine now. Thank you very much to Mr. Wang Chao for helping us sort out the two popular teams, Hermeless Agent and Open Reuters, their development history, their impact on the AI world, and their connection with the Web3 world. Of course, we also took a quick look at the talents that emerged from Web3 and some of their impacts on the AI field. Going back to the original question, in fact, we should not be so optimistic, proud, or even a little one-sided in judging that Web3 can bring such a popular product of AI. In fact, these are all examples, but I am very happy that we can finally discuss some reflections that the current wave of AI has brought to us. I feel that I have gained a lot from this program. Thank you, Teacher Wang Chao, thank you, Teacher Liu, thank you, Teacher Jiake, bye, bye, this is the program. Everyone is welcome to subscribe to the Web3101 program through Spotify, Apple Blog, Little Universe and other channels. If you are used to listening to podcasts on video platforms, You can also search for the Web3101 podcast on YouTube to follow us. We also look forward to everyone giving us suggestions by email or leaving comments. My name is Liu Fen. Thank you for listening.

大家好,欢迎收听Web3 101,我是刘峰。哈喽大家好,我是Jack。今天我们录制的时间是4月17日星期五的下午,和我们在一起的嘉宾是老朋友王超老师。王超老师跟大家打个招呼好吗?大家好,我是王超。今天请到王超老师是因为我们又想聊一聊AI的话题,必须声明Web3 101不是AI 101,可是我们觉得有一个话题特别值得。 展开讨论一下,王超老师就是这个话题的最好的一个嘉宾。因为在最近和硅谷101的团队开选题会的时候,他们每每会讨论起来一些非常火爆的新的AI的应用或者产品。比如说OpenRouter,最近当然是Hermes Agent。每当讨论这种产品的时候,我都可以讲一下这两个团队的最初其实都是从Web3 转过来的。所以每每大家都会很吃惊。大家每次都问出来一个问题,你们是不是可以聊一下,为什么从Web3出来的团队能够在AI的世界里面呢?大放异彩。 这些热门的AI产品,背后都是来自Web3的团队。这是这种普遍现象吗?我觉得这个话题也很有意思,所以就讲到了王超老师。因为王超老师来自Crypto世界,但是看AI看了有三年的时间了。 所以我们今天可以就这个话题展开讨论。必须说明,刚开始的时候我们是一个非常预设的判断。很多产品创始团队来自Web3,是不是它已经可以是一个普遍现象? 当然我们在讨论中可以对这个进行一个再更准确的判断。 刚开始的时候,我觉得我们可以先从最近一些非常热门,但是明显团队是有Web3,有Crypto背景的产品展开。比如说OpenReuters和Hermers Agent。 我知道这两个团队之前王超老师都和他们有过交流,并且还差一点点就成为他们的投资人。 能不能给我们讲讲之前的故事和您对这两个热门的AI产品背后的创始团队的了解? 这两个团队我确实都关注了蛮长时间了。大家看到的比如说News Research现在他的Hermers Agent突然火了,但其实不是。 他已经在AI赛道里面耕耘了很长很长时间,并且他过去获得过很多我们并没有关注到的,但其实是非常大的成绩。包括OpenRouter也是。OpenRouter可能他的故事没那么曲折,因为他从第一天到现在, 他一直在做同一个事情。他并没有说我今天发一个产品,明天发一个产品。但是他一步一步走到这儿,中间的这些思考其实也是蛮有意思的。 我建议是不是我们可以先简单给大家一些背景。OpenRouter和Hermers Agent他们是什么?他们背后团队和Crypto和Web3的关联。我觉得我们先从目前更热门的Hermers开始好吗? Hermers本身或者说New3 Search,这是他背后的团队。他目前火起来的当然是一个Agent产品,跟之前更火的一个叫OpenCloud是非常相似的。你可以把它装在你的本机上,装在各种各样的云服务器上,甚至手机上,连上大模型帮你做各种各样的事情。 他这个火当然是有一定的意外成分,但是在我对他的定位上面,它是一个很难用一个产品来形容的这么一个非主流的AI Lab团队。 并且这个团队在过去实际上做出了很多不为人知的成绩。New3 Search这个社区,它是起源于2022年。 当时实际上是散落在全球各地的一些AI爱好者,开始在一个Discord Server上去聊AI,聊模型。 他们里面既有从比如说GPTR时代就开始在玩AI模型的老炮,但是更多的是2022年下半年,随着Stability Fillion,随着HITDP发布而去入圈的爱好者。 当然这样的小社区肯定很多,New3可能只是其中的一个。 这里边有一个哥们叫Technim,他是假名,没有人知道他的真名是什么。 他当时就特别能折腾,每次有新模型发布他都会去搞一搞。 除了他在那个Discord Server里面还有几个人,一个是Karen,这哥们儿学的是宗教和哲学。 他白天是在斯坦福大学的一个脑科学实验室做机器学习的研究,然后晚上就去思考哲学这些东西。 另外很有名的一个就是他现在的CEO叫Jeffrey。 Jeffrey也是很杂的,他在这之前更有名或者说更被认为是加密原生的身份, 是他曾经是担任过Eden Network的首席工程师。 这是在当年MEV很火的时候也算是一个很顶级的项目。 Eden是一个非常典型的Crypto项目了,对标的是Flashbots。 它是做几菜访是MEV的产品。 它其实对于Flashbots当时来讲Flashbots很值钱, 但是Eden的估值很低,就是我们一直称它是一个低配版的Flashbots。 轰过一段时间也让我亏了不少钱。 所以Jeffrey我是知道他是很强的这种Crypto背景。 对,这个代币归零了,归零了好惨。 是吗,他归零了。 归零了,他后来他是停了坦,就是Eden停了自己的所有的产品, 因为中间也遇到了整个的这个以太坊的很多技术的变化, 所以他们那套东西都不work了最后。 但是他们也还好,他们把自己财库里的钱又分给了Eden代币的持有人。 我印象中我一大堆的Eden代币还领回来了一些以太坊。 严格意义上不算归零,但是这个产品和Eden Network确实是不复存在了。 所以Jeffrey我还是知道,他一看到我就知道News Research这个团队是很强的Crypto背景的。 就是这么一帮人,当然可能不止他们几个,他们几个是这里边最活跃的。 这个团队还是挺多元化,并且因为我记得Karen她是一个行为上面可以用混乱, 也可以用开放来形容的一个人。 我印象中她跟Crypto的思想她也挺能够接受的, 并且他们一开始也是在Discord这种环境下。 所以给我的印象,他们这个是一个跟Karen社区, 跟Crypto的文化真的是吸入挺多的这么一个团队,对吧? 对,之前大家其实是个兴趣小组, 在那个阶段其实还很难叫他们一个团队, 他们一直在玩自己一些小模型, 一些跟真正的商业模型比, 还是有巨大的差距的这些东西在玩。 真正一个标志性的事件是2023年的3月, 首先2023年的2月,Meta发布了它的第一个Karen模型Lama, 发布了一周之后,Lama的模型的权重, 这些也意外泄露了, 突然他们就有了一个真正至少接近商用的这样一个情况, 一座模型它开始可以折腾了。 那一个月之后,斯坦福发布了一个叫Alpaca的一个模型, Lama就是南美的大羊駝, Paka就是南美的小羊駝, Alpaca是咋做的呢? 他们是用GPT3.5去生成了5万多条的数据, 所谓的数据其实就是你问我答而已, 他们拿这5万多条数据去微调了Lama模型, 结果微调出来这个模型, 它的表现就接近了GPT3.5。 这个论文其实是非常轰动的当时, 不仅是说它搞出来一个模型, 更重要的是它证明了一个新的范式, 就是我可以用高等级的模型去蒸留出数据, 再用这些数据去训练小模型。 当时呢, TechNim看完这个论文, 它的第一反应就是, 我斯坦福搞这个东西只花了600块钱, 他用的是GPT3.5, 那如果我用GPT4呢, 我能蒸留出来一个什么东西? 所以他就直接开始干, 就用GPT4生成了一大批的所谓的指令数据, 质量肯定是比GPT3.5的明显更好嘛。 到了4月底, 他这个数据集做完, 他就发布到了Hugging Face上, 起名叫GP Teacher这么一个东西, 紧接着到6月3号, 那个时候用GP Teacher去微调出来的一个新的模型, 就发布了给他的命名就是Hermers, 也就是Hermers的第一个版本。 这个模型至少在当时还是非常好用的, 所以他发布几天之后就登上了Hermers开源模型的排行榜, 就一下就火了吧。 所以这时候Teknip一个人就已经忙不过来了, 所以他就把原来大家本来就在一块讨论这些话题的人拉进来, 继续做下一版。 这时候比如Karen也过来了, 他来做一些数据整理啊什么的, 包括还有一些社区的资源支持给他们提供出使的一些算力。 那那个时候其实News还并不是一个团队, 他也没有任何收入, 其实还是一个非常社区协作的这么一个东西。 这时候他们其实是也被HLZ看到了, HLZ就有给了他们一笔Grant, 数目也不大, 但是这个数目已经足够他们停下自己的本职工作, 开始专注在News上面了。 到了23年的8月30号, 他们就正式注册了一个公司, 就是News Research。 联合创始人就是刚才我们提到的那几个人, 或者是重加密背景, 或者是说也有一些加密上面的基因在这边。 非常有意思啊, 就是公司注册了一天之后, 他们就发了一篇论文, 叫YARN, Y-A-R-N, 因为他们自己在做LAMA的微调嘛, 他们就发现出版的LAMA, 其实只有4000个Token的上下文啊, 就是我们现在恨不得一个Million上下文, 我们觉得还不够长, 当时只有4000个, 这个根本不够用。 当时Jeffrey就在研究这个东西, 怎么能够把窗口拉长, 所以他们基于一些现有的方案, 搞出了很多的数学上面的改造创新, 最后就成功地把上下文窗口, 扩展到了原来的32倍, 他们就把这套包括代码, 包括方法论方法, 就发布到了网上, 然后名字起得非常随意, YARN的Y-A, Y-A是YET, A是ANOTHER, 所以翻译过来大概就是说, 又搞出来了一个错诊, 非常随意的一个命名, 当时他们当然对这个东西很骄傲, 但是我觉得他们远远没有预料到, 这个东西会变得如此重要, 后面YARN这个东西, 几乎变成了算是开源模型界的标配, 一直到三年后的现在, 你去看Deep Seek, 你去看千问, 你去看Lama, 你去看几乎所有的开源模型, 你会发现, 他们引用的这些长上下文的这些东西, 背后都是可能跑着YARN, 或者YARN的某一个变种, 来去支持他们扩展上下文, 也就是说这个东西变成了, 近乎开源界的一个标配, 从这我们也可以读出来一些信息, 就是尽管今天我们觉得, Hermers Agent让News Research这个团队, 好像大露了头脚, 但事实上人家这个团队, 并不是一无所有的, 在最早期的AI的开源模型的发展历程中, 他们有他们应该有的位置, 对的, 当然他们最先火, 也不是因为这个, 最先火是因为Hermers模型很好用, 并且Hermers模型其实是有价值观的, 这个价值观就是, 你不应该审查我, 因为你知道这些大模型公司, 为了安全, 为了所谓的对齐, 其实是设计了各种各样的东西, 当你问到这个, 他就不回答, 问到那个不回答, 所以当时他们其实是做微调训练的时候, 其实就有意识的, 并且用在技术上面, 把这个东西拆掉了, 当然很难说拆的完全干净, 但至少Hermers模型, 其实是一个非常开放的, 支持用户去随所欲的, 按自己的需求, 进行提问, 去拿到结果的这么一个模型, 然后再往下, 因为做出成绩了嘛, 无论是他的开源模型, 还是说他提供的这些拓展的方法论, 这时候他就吸引了资本的注意, 到了24年的1月, 他们就拿到了一轮种子轮, 这个种子轮是Distributed Global和OSS, 两家一块头的, 500多万美元, 当然还有一些天使投资人, 这里面也包括OpenRouter的Alex, 但当时其实他还是一个非常非常小的公司, 所有人都在远程办公, 还是依赖Discord在协作, 但是刚才讲到Karen这哥们, 他的背景是宗教和哲学, 刚才刘老师你也提到, 他有很多技术之外的思考, 到了24年的3月, 他们就发布了另外一个产品, 叫WordTheme, 他的界面就是一个命令行的网页界面, 底层跑的是Athorpic的模型, 这个东西其实是一个非常精妙的提示次工程的一个产物, 就是用户可以用自然指令给他下指令, 比如说你去模拟一个工业革命从未发生的2000年是什么样, 或者说模拟一个恐龙没有灭绝的地球现在是什么样, Cloud就会把这些场景基于这样的一个假设去拓展成一个稳稳世界, 你可以在这个世界里面去对话, 走动, 甚至是你走崩了还可以分叉什么的, 但是一个非常好玩的东西, WordTheme就爆了, 所以他其实23年的时候等于又展出了一个很好的明星产品, 大家开始玩得不亦乐乎, 但是这个用法就触发了Athorpic的内容警报, 因为有一些模拟场景可能超出了Athorpic能够接受的边界, 所以到了4月发布一个月之后, Athorpic就切断了News的所有的Cloud API访问, 他就被迫下线, 最后删除重新上线, 但是实际上这个时候它的吸引力就大幅下降了, 所以最后这个产品获得了短期的成功, 但是最终的不要了之了, 这个团队本来就是有这种去中心化的这些意识形态的基因的嘛, 被Athorpic这么搞了一次, 我觉得可能他们会更加强烈的感受到这一点, 这个产品其实就是活了一个月就没有了, 但是它其实很意外的后面发生了很多事情, 就跟News没关系了, 这本身这玩意儿是一个很精妙的一套方法论嘛, 这个方法论呢, 后来被一个叫Andy的一个哥们看着了, 他就fock了这个东西, 去改了一改, 改完以后呢, 他用他的那套提示词, 让两个cloud互相聊了9000次, 就在聊一聊的过程中发明了一个虚拟宗教, 这个虚拟宗教发展最后, 就变成了一个叫Terminal Truth的一个微小机器人, Terminal Truth是这么来的, 对是这么来的, 这个机器人就无意中忽悠到了HLV创始人的5万美元, 等值比特币的一个grant嘛, 最后这个事就被meme社区搞来放大, 最后就变成了一个gold的meme币, 最终涨到最高的时候可能到10亿美元, 这样一个, 但是后面的这些故事都跟news没关系, 但是源头确实就是他们的word theme, 很有意思的一个事情, 对, 这其实说明他们第一是很能做这种爆款产品, 尽管可能就是不长时间的热度, 另外点, 他们这种爆款产品也能够激发更多的人, 用开源的方式再创作, 很多东西其实他们是那种灵感来源的, 是的, 所以他们这个时候在整个圈子里面就已经很有名了, 我们跟他聊过嘛, 其实也是在这时候, 当时发现word theme这个东西搞得非常棒, 再往后去查, 发现这个团队原来他有更多的一些不为人知的成就在这边, 所以就赶快约他们去聊, 当时我们约他们是24年的5月嘛, 其实那时候已经有点晚了, 他们那轮已经接近close, 但是他们还是很加密友好的, 所以知道我们其实也是很重加密背景的这些人, 我们就还是聊得很愉快, 但是就是太晚了, 聊完以后, 因为他那时候已经很火了, 所以他估值还蛮贵的, 我们就犹豫了一下, 但是最后还是决定投, 但是就犹豫的这一下就满到已经进不去了, 就错过了, 最后公布的是24年6月, 他们从Delphi完成那几家, 拿了一个1500万美元的追加轮, 在这之后钱多了, 就继续往下做, 模型这个是他们的根级, 所以他们还是一板一板的随着Lama在, 因为Meta那时候还行吧, 还是开眼界的扛把子, 还是在一直更新自己的Lama模型, 从3到3.1到Lama4, 最后崩盘是Lama4崩的, 所以Nius也跟着Lama一直在更新, 但是Lama3的时候, 其实有一个还蛮大的一个事情, 就是Lama3其实发布了三个版本, 分别是8B的版本, 70B的版本和405B的版本, 当时405B的版本, 实际上社区其实是有些不同反应的, 第一是这个405B的版本非常厉害, 至少在当时的Banchmark里面, 其实就已经很接近B元模型了, 但是另外一个角度呢, 405B这个规模的模型, 其实对于一个开源团队, 或者说对于社区来说, 你已经没有办法微调它了, 你用传统那些微调的方式呢, 去调, 因为它太大了, 你这么去调呢, 发现没效果, 但是如果说你是去进行一个大规模的全参数的微调, 这个时候需要的资源也是非常高的, 普通社区根本就没有办法去做去, 所以当时那个Nius就放了个炮, 说我们要进行Lama3.1的全尺寸微调, 405B版本, 当然大家就不信, 觉得你别胡扯了, 你这么一个社区团队, 无论是从工程上面, 还是从资源上面, 我觉得你根本不可能搞成这个事情, 但是就很快质疑的人都被打脸了, 没有多长时间, 他们就把微调模型搞了出来, 搞出来这个模型本身, 我觉得是一个可能偏政治意义, 或者说偏是意识形态的一个事情, 这个模型我觉得它微调完, 可能也没有变得更好, 但是至少它证明了说, 我们社区作为一股力量, 我们是完全有能力去做这种, 你专业人士看不上我们, 觉得我们做不到的一些事情了, 所以当时应该说是405B模型的, 全参数的微调这个事情本身, 还是对业界有些震动的, 当时我觉得他们其实还是, 有自己的一些思考和路线吧, 但是这个可能就是, 都是我们外界来猜测了吧, 在他们完成这个微调以后, 没有多久, 他们就开始把一部分精力, 放到了一个新的东西下, 发布了一个报告, 叫Distributed Training Over the Internet, 就是跨互联网规模的分布式训练, 这个报告的核心的点, 就是说他们找到了一种方法, 因为我们知道大模型训练的时候, 实际上是对带宽的要求非常高, 因为你每一步训练参数的更新, 都要去做反向传播呀, 这些东西, 实际上是要有巨大的通信量的, 所以我们知道像那些数据中心里面, 都是用的英伟达的什么Infinity Band, 都是每秒几百季, 甚至更高的一个通信量, 这也就代表着说, 即使你有足够多的GPU, 如果他们没有被高速连起来, 其实你也玩不成训练, 但是Nius做了一个什么事呢, 他们找到了一种方法, 把训练大模型的时候, 需要同步的数据量, 降低了将近1000倍, 而且它的训练质量几乎不受影响, 那这就意味着, 其实你训练前沿模型, 你不再需要这种大规模的, 几万个GPU集群的高速带宽网络了, 你可能你的GPU分散在全世界各地, 普通家庭带宽就能支撑他们连起来做训练, 这就很外部三很crypto了, 对吧? 对对对, 我马上就会想到, DPing的故事马上就冒出来了, 对, DPing的故事其实跟这个是紧笔相连的, 如果说这个不成立的话, DPing就只能做一些比如推理, 如果这个路径被证明这个技术是有效的, DPing你就能做的就肯定不只是推理, 而是把最终的训练这步也攻克了, 所以Distro就是这么一个东西就做出来了, 几个月之后呢, News又把Distro的核心的算法写成了一篇正式的论文, 这个论文就发到了archive上面, 一个论文的预印版网站, 这个论文发出来了以后, 所有人都看到以后就愣了一下, 因为这个论文里边除了News Research的几个核心成员, 论文的作者名单里面有一个哥们叫Dark Kimmer, 他是Adam优化器的发明人, 现在全世界训练任何神经网络, 几乎都会用到Adam优化器, 所以他在Google Scholars总引用数是几十万次的, 并且他是OpenAI2015年创立的时候的11位联合创始人之一, 也就是这样的一个算是殿堂级的人物嘛, 他识别到了News Research他们这个工作的重要性, 以及技术的先进性, 所以他可能也对这个发展有兴趣, 具体这个过程我们就不知道了, 但总之是他选择在这么一个阶段去参与到的这个项目里边, 并且最后论文也是有他的一个署名的嘛, 所以从这个角度来说, 我们也一定能够能看到News Research, 其实一些硬核工作的含金量是非常高的, 从刚才说的那个YAR, 到现在他们这样的一个分布式训练的框架, 这是一个很硬核很硬核的团队, 那接着我们继续往下梳理啊, 他们基于这些东西就继续, 第一是他们自己的很多训练, 现在就可以依赖这个框架去做了嘛, 所以后面应该从Hermers 3开始, 他们的这些模型应该都是开始用他们自己的训练框架呀, 去用一些分布式的GPU再搞了, 并且这些也是随着开源模型的进化再越来越好, 一直到了25年底, 比较新的一个模型, 是一个36B的模型, 这个模型呢, 它是用的字节跳动开源的一个基座模型啊, 这个基座模型, 它其实是有一个里程碑意义的, 就是它做了一个参照, 他们用去中性网络训练的这个模型的版本的性能, 超过了用集中式训练的对照版本, 这个是很厉害的, 也就是说我不仅这个方法论是能用的, 而且我至少在一定场景下, 它训练出来的东西比中心化训练还要好, 这是一个非常前沿, 并且也是非常厉害的证明, 它证明了如果说有一天, 我们可能要打破巨头的算力的垄断, 那么分布式的GPU网络同样可以训练出有竞争力的前沿模型, 接下来就是最后的一个故事, 现在大家都知道了, 它的爱马仕火了, 那是26年初, 所以它发布了这么一个带有自我学习功能的AI agent, 龙虾先火了以后, 大家后来发现龙虾有这儿, 一堆的可能不好用的地方, 发现这个东西跟龙虾很接近, 很好用, 所以最近几周一下就变成了一个当红炸子机, 这个也是一个真正意义上的消费仪产品, 所以也是为什么今天可能我们有这个讨论, 是因为在C端, 它其实有了真正自己特别强的一个拳头产品了, 非常的清晰, 我之前只简单的知道他们有自己的开源模型, 并不能理解他们在AI开源这个社区中的所做的一些可圈可点的贡献, 这个团队他们其实是一个非常全方位的团队, 在产品线上面, 既有开源模型, 比如说Hermers, 另外一点他们自己也做了开源训练的框架, 就是Distro, 到最后现在火的是面向C端消费者的开源的智能体的框架, 这个Hermers Agent, 人家做的东西是挺全面的, 对, 它一方面是很全面, 但是听起来好像是有点杂, 模型也做, Agent也做, 什么训练框架也做, 但是其实如果说把整个这条线通通到尾去看的话, 你会发现这里边的意识形态层面的东西, 实际上是始终是同一个东西, 比如说他为什么第一次Hermers能火, 不仅仅是因为他这个模型效果很好, 而是这个模型选择了一个跟大场不一样的路径, 叫中性对齐, 这个模型他不预设自己的立场, 不做审查, 他去忠于用户的意图, 而不是说某一个商业公司的伦理准则, 这里边其实有很多哲学思考, 这也是他最初在社区那么受欢迎的原因之一, 因为他们相信用户的主权, 要高于公司自己制定的这些伦理规则, 这在我看来其实也是一个很强烈的意识形态, 以及还是跟加密圈所相信的很多东西是一回事, 那后面呢, 刚才我们重点讲了半天, 他为什么要基于405B做全参数微调, 他要证明我们社区的力量, 其实是够的, 不是你们想象的那样, 到最后他做这个去中心化训练, 这个其实是很强烈的有一个隐患在他整个过程中, 比如说他Hermers从1到2到3, 他其实之前一直用的开源的这个拉玛模型, 那其实有一个问题, 就是如果有一天Meta决定避原了, 我怎么办? 我没有一个很好的机座模型可以基于他去做东西来, 那个时候中国的开源力量还没有起来, 那时候基本上全世界都在依赖拉玛, 在一板一板在迭代, 所以这时候他既是我觉得意识形态上从始至终都有这个思维, 同时也是有危机意识的, 那有一天小家说我们必远, 他现在确实必远了, 对吧, 最新的版本, 或者说有一天他改了许可的这些license, 那怎么办呢? 我们玩什么呢? 我们不可能像那些大公司一样, 拥有几万个GPU集群在那边去维护他, 那我唯一能够的事是, 让全世界的资源汇集起来, 我们社区也能做这个事情, 所以这个就是他们决定做去用英华训练框架的一个深层的原因, 那到最后Hermers Agent看起来是一个C端产品, 其实Hermers Agent隐藏了非常非常深的一个东西, 前面讲了它搭了不依赖巨头就能训练模型的基础设施嘛, 但是从2024年底2025年开始, 这个时候大模型变成了说预训练当然很重要, 那后期的RH训练其实也很重要, RH训练就需要有很多这种行为数据, 所以这些大模型厂商他自己可以花很多的钱, 做很多的公式上的东西去做, 无论是真实的这些数据的产生, 还是说合成数据的这种高质量, 但是社区其实是没有这个能力的, 那Hermers Agent实际上他要承担着一个, 大家没注意到的一个东西, 就是他其实里面内置了一个深层的RH训练框架, 也就是我们所有的Hermers的用户, 你的所有的行为数据, 都是可以直接导出成一个直接用于大模型, RH训练的以专业格式组织的训练数据的, 当然这个东西因为它是开源的, 它也是纯粹本体化部署, 它并没有收集, 它想收集也收集不到, 但是至少他先把这个技術车是放在这儿了, 当有一天我需要收集的时候, 我是不是能够通过一个什么样的方式, 去激励去发动一些社区成员, 愿意把我自己的数据拿过来, 哪怕95%不愿意, 可能5%这也是一个很大的数据集了, 而且它是足够的分散, 有足够各种各样可能奇奇怪怪的场景, 这个隐含的一个风布式的, 未来的marketplace都可以看到了, 这个非常的crypto了, 对, 我现在也在用hermothagen, 其实我们每一个人将来都有潜力, 成为他这套训练里面的一个贡献者, 可以想象他将来如果引入一层激励层, 那这个东西就可以形成一个强化学习的数据市场, 是的, 他只是现在没有这么做, 对, 您这个梳理非常的精彩, 因为我之前也看过这个news research, 但是我对他的关注纯粹是因为他们在去年, 因为融了一轮大的差不多5000万美元, 并且是paradigm投的, 那一轮他们也很值钱了, 去年4月份已经值, 我印象中估值已经是在11美元了, 并且当时是很明确他们是会发币的, 并且那个时候他们还有一个亮点就是在于, 刚才就是您介绍的Distroy这个去中心化的算力与训练的系统, 他们在这个上面他们其实想做一个, 我记得是叫Psyche的一个network, 放在Solana的区块链上, 对吧? 是的, 他这个其实已经做出来了, 对, 因为这个等于说是一个crypto的产品了, 并且这个东西我印象很顺势, 因为他是和被大家认为是beaten sum的一个杀手, 一个最核心的竞争对手, 所以讲到news research, 我第一反应就是他是一个币圈的团队, 对, 我觉得您总结是对的, 今天可能我们的话题是说, 加密团队如何跑到AI领域里边去做出很好的成绩吗? 其实news research, 他不算跑到AI领域里边, 他现在依然是一个根弧瞄正的加密团队, 只不过他把两边掰得很清楚, 我面向消费者的, 面向AI领域的, 我就不提什么tokenomic, 我不提什么区块链, 我就给你一个很好的产品, 我就给你一个很牛的一个算法改进, 我来证明自己, 如果有一天我需要加密这端的时候, 不管是我曾经需要这边的资金, 还是我需要这边的一些初始的市场支持, 这是这边的事, 我跟这帮人讲, 我不跟你们讲, 他这两边分得很轻, 对, 所以我的印象中就是, 他的币可以发在哪里, Sychic肯定是可以发币的一个方式, 而最近Hermas Agent火了之后, 我听到很多crypto的朋友都会说, 这个团队, 这个产品火了之后, 他们之前那个币可能就不发了, 但是听你们这么讲, 我觉得其实这个团队, 从他的根本上面, 其实是一个相信crypto, 并且相信去中心化, 和开源和社区贡献的这么一个团队, 反而可以想象在未来, 他有很多种, 有可能会让大家吃惊的做法, 埋了很多有意思的线索, 这个团队完全出乎我的意料了, 他能够在AI领域和crypto领域, 都能够非常清晰, 和漂亮的展示自己的优势, 挺棒, 但是如果说到发币, 我还确实不确定他会不会发, 我知道他团队内部, 对这个是有不同的意见的, 那他现在其实做的这么成功, 毫无疑问, 未来他可能会融更大的, 不管是来自加密圈的资本, 还是来自于主流AI圈的, 如果说你看他现在能分的这么轻, 是因为他很清楚的知道, 什么是对的, 什么是不对的, 我做一个东西, 我如果上来就把加密这个东西融进去, 可能我就失去了更主流的一些用户, 付出发币我肯定不知道, 但是如果说他们在没想清之前, 我觉得他们一定不会贸然的去搞这个事情, 肯定的, 在今天我觉得crypto web3这个世界, 是一个黯然无光的环境, 但是AI心性向荣, 今天讨论他们的发币, 确实是一个逆反天纲的事情, 谁知道世界还是往前走的, 最起码我在去年看到他们内论融资的时候, 是非常清晰的, 他们的估值是以token来估值的, 好处就在于这个团队真的在交付, 另外一点的好处在于, 其实他们有非常强的, 在思想的领域上面, 其实是相信社区, 相信开源, 相信那种去中心化的这些文化, 而他们又知道有非常漂亮的打法, 其实有很大的可能性, 在这个技术上, 其实我们快速的可以看一下openreut, 相比之下openreut, 他们的背景和他们的发展, 就应该更加的简单和明了了, 因为他的创始人, 我印象中是18年, 作为一个也算是一个创始团队的重要成员, 以CTO的身份加入的openc, 18年的时候openc, 默默无名, 之后在NFT的周期里面, openc大放异彩, 但是呢, Alex我印象中, 他应该是在22年的年中, 也就是NFT最火爆, 往下走的那个时间节点了, 他算是离开了, 23年了他就创立了openreut, 因为王上老师应该跟他们有过交流, 其实我挺有关心, 就是您什么时候关注到openreut, 以及你为什么觉得这个东西值得关注, 因为我看到的现象, 就是因为龙虾的火, 让大家突然开始关注到大模型, 特别是大模型API的生意, 这时候openreut突然就出现在每个人的事业里面了, 但是我相信其实他很早就已经成为一个挺重要的基础设施, 因为我看到之前A16Z好几次在讲, 通过openreut这种业务和数据的分析, 是可以观测出来大模型和大模型公司, 他们的交付的能力和他们的用户的吸引力的, 对,openreut我注意到它其实非常早, 是因为在23年大模型刚开始火的时候, 有很多API你想掉它是很麻烦的, 你比如说你想掉XGBT的API, 你要搞定付款问题就给你霸掉一层皮, 当然也有很多国内的中展站什么的, 其实你又发现它的体验不够好, 甚至里面有一些假模型, 所以openreut当时就作为一个算是提供中立第三方的公司, 就已经很好用了, 所以关注到它应该是23年底吧, 就关注到它, 也在看它的量, 因为它是有排行榜的, 它的量一直是在公布有多少多少量, 也能看到它的量一直在上升, 所以这时候也就了解到openreut背后是Alex, 这不就是当年OBC的那哥们吗, 所以就想着, 哎, 要不要约他聊一聊, 就找了很多人, 最后终于其实是约到了, 但是就聊完以后, 一是人家对我问, 可能也没有太感兴趣, 在一个当时它也没有在融资, 所以就不了了之了, 到后面其实它变成了一个非常重要的生态位, 因为这本质上其实是提供了一个通用的模型调用层, 其实它在圈子里边的位置就已经算是一个很重要的位置了, 能看到比如说XI GPT, GPT 4.1当年是先在openreut上发布的, 起了个假名字, 让大家猜来猜去, 最后猜出来原来是XI GPT的最新的模型, 也包括今年小米的那个模型, 也是上去放了两个模型, 让大家猜, 大家当时都知道是中国模型, 猜来猜去谁也没猜到是小米, 也就是它现在实际上在生态上面已经非常重要了, 一个重要的分发层了, 对对对分发层, 对于开发者来说, 它也是非常重要的一个非常稳定的通用的模型能力层, 你可以不用管不同的模型里面有各种各样的这些摩擦, 各种各样不同的这些乱七八糟的东西, 你到它这儿反正就有一套API, 就一个接口, 就一个key, 你愿意接谁就接谁, 而且它永远不会封你, 这个故事本身我觉得没有太多的曲折, 就是它从day one就要做这个, 一直从小到大做到现在, 变成了行业里边的, 可以说巨头我觉得也没问题, 所以故事本身很平淡, 但是如果说我们再往后延伸一步, 梳理Alex这个人, 我觉得这个还是挺有意思的, 他跟Bankless录过一期播客, 我觉得他之所以能转身转得这么的, 你说他优雅也好, 说他坚决也好, 瞬间从web3跳到了AI, 并且做得这么好, 我从他的那个播客听到了几个东西, 第一个东西是, 他在做加密的时候, 他没有意识形态, 他不像比如说news-ray-search, 他是有很强烈的意识形态, 很强烈的这些东西在这边的, 在他看来呢, 我做open-sea是做的什么事呢, 我不是做NMT, 我不是做这些什么改变世界的东西, 我只是把NMT领域里面, 这么做复杂的, 不同格式的, 代表着不同意义的, 这些乱七八糟的东西, 给他梳理了出来, 在他看来, 这些东西不是NMT, 这些只是一个异构形态的库存, 我是把这些异构形态的库存, 做了一个聚合器, NMT只是他碰巧聚合的这个东西, 对, NMT只是一个展现的形式, 只是一个大家感受到的东西, 但是实际上他在open-sea处理的, 都是大量的数据和这个NMT背后的, dataset的这种处理, 对的, 对的, 所以在他自我认知里面, 这也是他自己的原话, 他说, 我其实只是做的异构数据的聚合, 那NMT是异构数据, 那现在各种各样不同的模型, 每家厂商自己的模型的, 这些很具体的参数, 然后调用的方法, API的格式, 计费的方式, 就乱七八糟都不一样, 这不就是一个事吗, 对吧, 那我来这边搞一搞吧, 对, 这个确实把他的这两份工作的, 核心的点给挑明了, 其实是一回事, 对, 就非常有意思, 而且他之所以下定决心搞这个事情, 从源头上来说, 跟news-resource的触动是一样的, news-resource是, TechName那哥们看到Lama, 出了一个微调版, 说觉得哇, 他能做我也能做, 他就做了, Alex看到以后, 他的反应是不一样的, 说哇, 600美元就能搞一个微调版, 那将来一定会有千千万万个模型, 那这就是我的事, 我就可以把这千千万万个模型, 通过一套东西聚合起来, 所以同一个事情, 不同的人判断出了不同的东西, 做了不同的事情, 最后都获得了很厉害的成功, 但是他有一个很有意思的一个习惯, 就是说他为什么能捕捉到这些东西呢, 他其实也是有很重的社区化的思维的, 他的社区化思维就是去识别出, 这些奇奇怪怪的社区, 就比如说当时OpenSea, 为什么能做出来, 因为他们发现, 在2017年底的时候, 为什么突然有一帮奇奇怪怪的人, 花那么多钱去买一只猫的图片, 这是啥玩意儿, 他们为什么在搞这个东西, 所以他就跑到这个社区里面去研究这个玩意儿, 研究完了就发现, 那如果这个东西能长出来的话, 那他们肯定是需要上面有一层东西来服务他们的嘛, 这其实就是OpenSea的模式, 但他现在发现, 哇, 有这么多奇奇怪怪的社区在搞各种各样的模型, 也包括很多大模型公司, 我能做点啥服务他们呢, 这是他的另外一个, 算是他自己的方法论吧, 先去识别这种奇奇怪怪的新的东西, 沉进去, 看大家在聊什么, 看大家在做什么, 然后想一个问题, 就是他们这个东西如果能做大, 他们需要什么, 最后就基于这个方法论做出来了两个产品, 这个是OpenReuters和Crypto和Web3的一些关联, 我不知道Jack有没有要补充的, 或Jack你有什么问题需要提出来? 对, 我之前看OpenReuters他的故事的时候, 我就觉得可能他确实还是在加密圈, 不是很有代表Web3圈子的人, 然后来到AI世界之后, 大放异彩那种案例, 很典型的就是一个技术男, 他把任何东西, 无论这个东西是Web3还是Web集, 对他来讲都是在解决一个特别深的, 实际的这个技术问题, 以及从这个问题上看到的一个商业前景, 这么去理解的。 对, 然后AI给他了一个更大的平台, 这个让我想起来, 之前我看CZ好像有一段视频, 他这个可能跟AI没有关系了, 他在一个论坛上面说, 他为什么要做币安, 他就说他自己没有选择, 因为他在这个之前一直在做交易平台, 他觉得他自己是一个被训练了很多次的一个猎人, 然后他觉得这个目标出现的时候, 他一定要扣班机, 我觉得就是有比较相似的这种逻辑在里面, 对我来讲就是一个领悟, 就是我们也别那么舔着脸, 人家成功了, 就马上跳出来说, 你看他们就来自币圈, 来自web3, web3的人干AI就会更有优势, 这个判断其实都是太过于片面, 以点待面, 其实是非常武断和不科学的, 我们应该脸红。 我觉得刘老师在咱们开始录之前讲的那个观点确实挺对的, 可能crypto的很多从业者, 或者是币圈的人, 用AI或者是在AI里边发现机遇, 确实会更灵敏一些, 这个让我想起来, 前两天发现cloud code代码泄露的超凡, Fuzzland的这个创始人, 他其实也是crypto的背景, 然后这个Fuzzland本身也是在crypto领域做安全的, 所以我觉得其实还是有那么一点东西。 超凡也是一个非常酷的极客, 开始为什么我想给出来一个之前被大家问到的问题, 我觉得我们很难这么去下结论, 因为有一些热门的AI产品是由之前跟币圈相关的这些创始人, 或者是团队做的, 就来判断web3对AI是有影响的, 或者说web3的人其实他们有一定的优势, 可是我觉得有一些事实我们可以提出来就是, 毕竟在上一个周期, 特别是22年左右的一个周期, 在AI的爆发来到之前, crypto和web3是吸引了最优秀的年轻人的。 这些年轻人当他们在那个时间节点, 他们勇敢地跳到crypto的这个tool洞里面去尝试, 之后又看到了更有意思的一些事情, 他们转到了AI, 这是很容易理解的一个发展的过程, 所以可能确实你会看到有不少现在在AI领域, 不断地尝试新东西的人, 是从上一轮的crypto世界转向过去的。 另外还有一个现象就是, 你会发现因为crypto在很常时期内, 它并不是一个被主流很容易接纳的领域, 它会吸引很多非常的有开放的想法, 甚至有可能可以用对反这种传统的这些人, 他们会勇敢地去尝试。 所以crypto的这个世界里面很多用户, 还有开发者, 他们本身就带有那种很强的冒险精神和开放的思维。 所以你会看到在很多现在AI的新产品的应用上面, crypto世界的这些朋友们, 他们用得非常的好。 我会发现在web3世界的一些朋友, 他们很快地就会去拥抱OpenCloud, Hermers Agent, 拥抱一切变成爆款之前, 就已经开始尝试。 我觉得这都是有些内生的原因的。 但是我也在想就是, 我们不应该那么快地去下结论, 就是web3来的人在AI世界还蛮能够大放异彩, 我觉得这个判断确实过早了。 但是我觉得我们也可以再看一看, 还有哪些有过web3的背景, 现在在AI世界还做得不错的团队或者是创业者。 我印象中MultiBook, 就是让agent去社交的产品的创始人, 他应该其实很早就是一个比特币社区的活跃分子, 并且也做过像MemeCoin这样的东西, 是之前在crypto世界挺活跃的。 另外一点, 我能够想到的还有就是stability AI的创始人Amert, 他好像之前也是在crypto这个领域, 也算是很早涉足的。 Amert其实他在加密领域里面涉及的还挺深的, 只是不是很被人所知, 而且最早stability AI差点就做成一个道, 但他幸亏没做成道, 做成了道就不会有那么成功了。 Stability AI, 他在AI领域里面, 他的影响力是什么样一个情况? 另外我觉得您可以展开讲讲, 他的最早的创始人Amert在crypto的背景, 和对这个公司的影响。 当然了, 今天Amert应该已经被踢出局了, 他现在又回到web3。 就是Amert这个人本身算是非常有争议嘛, 比如说创始人的纠纷, 比如说张口就来, 然后经常吹牛。 对, 不诚信。 对, 就是这些问题, 但是曾经的stability, 我觉得他在AI领域, 尤其是在开业员为AI领域, 我觉得他的历史上的贡献, 是非常大的, 而且说他是标志型的, 都不为过。 那他过去, 其实在他进入AI领域之前, 他当时做金融的, 那段历史我不是很了解, 他还试着做过一个加密的, 跟公共事业有关的项目, 但是那个项目是用了加密技术, 但是没有发币, 也没有做成。 但是其实他对于加密技术, 以及加密的这些意识形态什么, 他其实是很认可, 也很熟悉的。 所以后面的Stability创始的时候, 因为Stability至上也是一个社区化的启动, 也是由全世界很多地方的贡献者, 大家一块儿, 传出来的一个事情嘛, 所以当时他就想过说, 这个其实应该是一个道, 甚至是一个道 of道, 就是由Stability去支持各种各样的开源AI社区, 去做很多东西。 这故事他自己讲过不止一遍, 那最后他说为什么没做呢, 是因为他觉得记录设施啊, 这大的都不成熟, 那我觉得他的观察是很对的, 他幸亏没这么做, 否则可能就没有Stability了。 那Stability我觉得他最重要的贡献, 是他在应该是2022年的夏天, 他把自己的拳头的模型开源出来了, 现在我们觉得开源模型这玩意儿, 好像是理所应当的, 就是有一个流派在搞这个事情嘛, 其实在这之前不是的, 在业界有非常大的争论说, 我们应该做模型开源什么的, 如果我做了有哪些法律问题, 有这那的一锥各种各样的东西, 反正Stability光脚都不怕穿鞋了嘛, 反正到了那个时间, 当然可能也是他的自己的一个商业策略, 同时可能也有一些意志形态的东西, 所以他就把这个东西开源了, 就一下实际上AI的浪潮, 其实是早于Chai DPT的, 在Chai DPT之前就有很多极客, 有很多尤其是可能对于这种图像生成, 有兴趣的人, 已经在玩的不亦乐乎了, 这一切我觉得可以说都始于他对于这个模型开源的商业决定, 在23年真正AI活起来以后, 那时候他也是在主桌上的, 我记得当年美国有一个总统, 还是白宫, 还是国会的一个晚宴, 邀请了可能10个左右的人吧, Elon Musk, Sam Altman, 然后其中就有Ema的这个人, 所以这是正经上过主桌的一个男人, 只不过后面的这个东西的可持续性不够, 再加上可能有各种各样我们不知道的问题, 就慢慢他就下了牌桌了嘛, 又回来, 所以他其实不是一个从AI圈跑到加密圈来做了一个项目的人, 他是从加密圈先跑到了AI圈, 做出了一些成绩, 但是AI圈嘛, 没有长城将军, 他又不行了, 又跑回来再琢磨其他的事情了, 他现在在做些什么呢? 我记着他是现在做的东西, 是跟去中心化的AI的技术设施是相关的, 我上一次关注他也得有个小两年了, 因为当时他是号称辞职啊, 从Stability, 但实际上应该是可能被股东踢出来的, 他只是没有承认, 踢出来以后第一时间我们就约他了, 因为他最早的投资人有一个熟朋友, 跟他很熟, 所以就约到他聊, 那时候就吹了一头牛逼嘛, 就跟我们讲什么主权AI, 什么not your mind, not your model什么这些东西, 大概的意思就是我要做一套技术设施, 我支持全世界的任何一个国家, 或者任何一个组织, 任何一个团体能够去信念自己的模型, 讲了这么一个故事, 后面呢他启动了他的项目以后, 大差不差还是在这个故事的方向上, 可能细节略有区别, 但再后来他好像确实没做出什么东西啊, 就感觉消失了, 我能记住就是他写的本书, 叫做The Last Economy, 是讲AI打破的稀缺性, 让大家的生活足够富足之后的一些选择, 请未来感, 但是我没有看书, 只看过一些书评, 另外的话呢, 我还能印象中就是他一直和Render这个项目非常的紧密, Render就是做去中心化, GPU渲染的那个团队, 他在这个团队里面一直是做顾问, 另外一点的话呢, 我能看到就是他经常会讲一讲关于开源呀, 去中心化, 给大家赋予, 做主权的这些讨论, 产品确实也没有看到, 但是可以确定就是他其实在Web3的世界啊, 在做一些尝试的, 所以Stability AI确实是, 跟Crypto是多少有些关系的, 我们刚才讲到了Hermers Agent和Web3的关联, 还有OpenReuters的关联, 简单的提了一下MountBook, 然后呢也讲了讲Stability, 此外的话, 我能看到大家经常会讲的其实就是FTX, 当时推出的FTX Future Fund, 里面请了一些特别酷的年轻人, 这些年轻人现在在AI这个世界里面是大放异彩的, 但是我个人认为这些人跟Crypto的关系并不紧密, 是因为在22年的那个时代, 因为FTX有钱, 他可以砸钱去找那种最前沿的这些聪明人, 他当时无论是请的这些高级管理人, 还是招募的这些Future Fund的这些年轻人, 都是独当一面的, 那今天的话比较著名的就是, 后来去了OpenAI, 后来写了一篇长文, 太是感知, 因为写了这个长文, 拿到了很多大佬的钱, 做了一个态势感知对冲基金的那个小伙子, 名字特别长, 利奥波德, 阿胜布伦纳, 他前段时间也是在媒体吵得特别多, 这个哥们是从FTX Future Fund出来, 后来又去了OpenAI, 后来又做了对冲基金, 非常火, 还有一个就是辽波德的女朋友, 阿维塔尔, 他现在是在Authropic做幕僚长, Chief Staff, 他也是当年他们一起都在FTX这个Future Fund, 经常大家把这两位是当成Crypto到AI的浓潮, 不过我倒是觉得这样讲并不是特别的准确, 因为在那个时代他们就是聪明的年轻人, 因为FTX给了钱, 所以为他们工作去投资未来, 并且Narrative就那个故事讲的也非常的吸引一些有想法的年轻人, FTX一直在讲, 就是说我要做利他主义, 提倡开放的慈善, 要为长期投资, 为人类的未来做不计回报的投资, 那肯定能吸引到一群非常具有前卫思想的年轻人, 所以我不觉得他们是真正的Crypto出来的天才, 其实我觉得算力转AI确实还是有跑出来的一些公司, 像那个CoreV, 对这是另外一个故事, 对对对, 如果我们不说Web3而是说加密, 其实这是一个房间里的大象, 我们不应该忽略它, 因缘即会吧, 他们其实对于现在整个美国AI的格局, 其实都有一个结构性的影响, 如果不是当年这帮人在这预先去搞定了这些电力的审批, 一些前期的建设, 那美国现在电力的紧张要比现在要紧张得多得多, 你现在现建可能2030甚至2032年之前都建不到东西了, 可能就一两个月之前吧, 大家对这些算力公司转AI的态度都还是说, 你看这些公司比较追随热点潮流, 都是在玩弄热点, 但确实还是有跑出来的一些, 他们还是对美国的这些AI的基建做了挺大贡献的, 我觉得CoreV真的还是现在挺大放异彩的, 我们之前在整理相关的这种每天的热点的时候, 它经常就出现, 之前要不然是Meta想要去收购它, 或者就是前两天好像那个Jane Street也给它投了60亿美元, 再花10亿美元给它做事, 就是想要用这个CoreV去做量化方面的一些训练, 或者是业务, 我觉得确实这个是加密和AI这个领域联系比较紧密的一个部分。 你先不管它追不追热点, 但实际上大家对于他们的重要性是完全低估的, 他们绝对不是大家想象的说, 一帮矿工现在AI火了, 来转AI来转AI的钱, 我觉得他们在整个, 或者说至少在北美的AI算力领域里面, 他们是占据了一个非常重要的生态位的, 而且他们占这个生态位不是因为他们, 因为现在AI火了我转过来, 这里边当然有历史的巧合, 比如说2022年当时以太坊Morge以后转POS, 一下就腾出来了很多资源, 紧接着两个月之后, 拆GPT发布接上了, 还有像那种大规模的电力的建设, 审批长期勾电协议电网的这种协调, 也包括数据中心的建设与快速的部署, 也包括大规模GPU集群的管理, 这实际上是非常非常专业的技能, 等于是这帮公司批量的产出了很多很多专业人才在这边, 他们只要稍微培训培训, 把有一些特定的东西转成使用AI的方式, 就能用了, 同时他们在那个历史时期, 谈成的一些电力的比如说审批建设, 这个东西是非常宝贵的, 我记得之前天规谷101说过, 说有的估算是大概, 他们掌握了可能6到10G瓦, 但是有更乐观的估算, 说是大概在20到25G瓦, 一G瓦差不多是一个大型核电站, 差不多的规模, 这已经是一个天量的一个电力容量了, 如果没有这帮电力容量在这儿, 几乎是ready的可以转接AI, 现在我们可能每天用的Cloud的配额, 我们用的CHATGPT, 我们用的GemNet, 可能用不了三分钟就告诉你说, 对不起你这个要限额了什么了, 其实它的战略意义是非常非常高的, 我们不管它是因为商业上的原因, 还是因为什么原因, 他们只是想挣钱, 但是他们在历史上面, 实际上是起到了非常重要的一个作用, 也就是说, 无论他们当时是为了追热潮, 还是正好历史在那个齿轮转的时候, 无论是以太坊的POW转POS, 或者说是Crypto的周期, 逼着他们不得已往AI方面去转, 但是他们的作用是显而易见的, 尤其在今天, 其实在美国AI的算力中心的建设, 还并不是大家想的那么简单, 你拿钱就能开建, 无论的这种各地电力的供应的问题, 各个地方社区政客对这个问题不同的认知和存在的博弈, 其实反而这些从Crypto, 从挖矿过来的这些计算中心, 还撑起了一些的AI算力的机梁, 所以这个是一个积极的影响, 对对对, 不仅是一个积极的影响, 而且如果我们去升为也能升出来, 因为这些技能还真的是有用的, 不像我们在Web3项目, 我们想来想去, 想了半天好像也想不出来有什么东西, 是过去有优势的, 这个是真的能用起来, 这是实际是真正交付的能力, 总之我们已经把我们能想到的, 这些Crypto世界出来的人才公司技术, 甚至是电能给AI领域提供的贡献都梳理了一遍, 然后我们就可以回到一些相对而言更大的讨论, 挺想问一下王超老师, 因为刚才讲到的这些, 其实背后有一个绝对不容忽视的现象就是, 其实Web3的人才在流向AI, 无论是那些为了趋利, 还是为了逐热点的这些人才, 更核心的是曾经Web3或者Crypto是对奥赛的金牌得主, 最顶尖的大学的这些计算机系数学系密码学系的研究人员, 这些最聪明的年轻人, Crypto是最好的一个选择之一, 但是在今天我们在做这个节目的时候显然不是了, 最好的选择都是AI了, 我们怎么看这样的变化? 也不沉重, 就好像是非常自然的一个事情, 因为加密世界现在发展到这样的一个规模, 你说这里边还有多少人是真的因为对于密码朋克的那些认可, 对于去中心化的这些东西来到这个领域呢? 大家来这儿, 绝大部分人是为了利益嘛, 这里边当然利益又分几个嘛, 有些人觉得这里边其实是很容易起项目, 我在这边想到一个好的东西, 我去把它做出来, 我可能获得回报会很快, 那更多的人实际上是来参与这个投机, 或者是还有很多人就是来骗的嘛, 无论是他没有这个底层价值观的这些优秀的人才, 当然这里边我们没有任何诈指的意思啊, 你觉得加密好, 你是个人才, 你来这儿, 现在加密市场走到熊市, AI更好玩, 你去AI这是一个很自然的流动, 所以其实也没有多深层去能解释的事情嘛, 因为我觉得过去呢, 大家觉得加密节奏快和AI节奏快, 这两个事不是一样的快, 其实加密是一个很慢的一个事情, 因为加密你对抗的东西是制度, 而你在AI想做一个东西, 你对抗的只是工程, 你对抗的只是一些技术上的一些东西, 一些产品实现, 所以从这个角度来说, 加密是极慢的, 但AI是极快的, 之所以加密让人觉得快, 是因为这个圈子赚钱太容易了, 骗钱也太容易了, 钱又太多了, 是赚钱的周期快, 而交付的周期是很慢的, 火爆的产品非常快, 但是被淘汰的速度就非常的快, 真的想做出来一个好的Crypto, 或者Web3的产品, 而不是单纯的三分钟热一下的这种, 就可以发币快速圈钱的产品, Crypto其实, 其实这个周期会慢很多的, 所以第一我是觉得, 为什么要拿出来讨论呢, 这是一个事实, 并且好几个老朋友, 他们在各个地方写长篇的文章, 在呼吁Web3应该留住人才, 我都觉得, 哎, 这种想法我是理解的, 但是我是觉得, 人才的流动就应该是自由的, 当AI的世界真的在交付, 真的在不断的快速的进行技术迭代, 大量的新的应用, 新的想法在涌现, 而用户在这里面, 真的是可以感受到, 对我们的生活, 对我们的工作, 对很多的社会, 都带来真正的冲击的时候, 我觉得人才留在那里, 真的很好, 而相反的话, 其实Crypto的世界, 在今天的话呢, 你会看到, 它更多的涌现的是在, 传统金融对Crypto的结合, 那这有它自己的魅力所在, 我觉得刚刚王超老师讲的那段, 真的让我非常的有感触, 今天可能讨论这问题, 我们会觉得有点点伤感, 但是好在一点就是, 我还是很相信, 因为Crypto本身, 特别Crypto, Native, 就是原生的这些思想, 我相信它有它的非常强的, 和Solid的那种稳固的根基, 这些东西呢, 在一定的时间, 在合适的历史的周期里面, 它一定会带来巨大的推动力, 我是相信到那个时候, 人才还会再留回来的, 当然了, 前提就是, 我们真的能看到, 突破性的技术的进展, 外部世界的巨大的冲突, 可能让我们周边的整个社会, 都重新在认识, 一些社会的文化层面的变革, 我倒觉得那时候, 可能我们再去看, 人才做的这种选择, 可能会更有意义, 而现在的话呢, 应该为这些, 从Web3到AI世界, 又能大方异彩的这些人才, 给他们鼓掌, 我觉得刘老师说的特别好, 我们不去讲那些意志形态什么的, 在我看到的这些成功者, 他们其实就是一帮有着强烈的好奇心, 和非常好建设能力的人, 他们在不同的技术浪潮之间流动, 当有一天可能市场回暖, 当有一天可能我们又找到了更好玩的东西, 我相信还有很多人才会回来, 然后会建设出更好玩的一些, 我们现在想象不到的东西, 非常感谢王超老师帮我们梳理了, 两个热门的团队, Hermeless Agent和Open Reuters, 他们的发展历程, 他们对AI世界带来的影响, 还有他们跟Web3世界的联系, 当然我们也快速地看了看, Web3涌现出来的这些人才, 给AI领域带来的一些影响, 回到最初的问题, 其实我们不应该那么乐观, 骄傲, 甚至有点特别片面地在判断说, Web3就能带来AI的这种热门的产品, 其实这都是个例, 但是非常高兴就是我们最后可以讨论讨论, 目前AI的浪潮给我们带来的一些反思, 这期节目我觉得真的收获挺多的, 谢谢王超老师, 谢谢刘老师, 谢谢佳克老师, 再见, 再见, 这就是本期节目, 欢迎大家通过Spotify, 苹果博客, 小宇宙等渠道订阅Web3101这档节目, 如果您习惯在视频平台收听播客, 也可以在YouTube上搜索Web3101播客来关注我们, 更期待大家通过邮件或者留下评论的方式给我们提出建议, 我是刘芬, 谢谢大家的收听,