S2E53 Hermes Agent 深度上手！Opus 4.7 到底有沒有變強Ｗ?

Today we are going to talk about OpenClaw, the killer of crayfish, Hermes Agent. I recently bought a Mag Studio and wanted to run my own personal AI assistant on it. So I chose Hermes Agent this time because of the crazy discussions on the community. I found its installation experience, and its factory capabilities are really great. You can also communicate with him on Telegram or any Messaging App you like, so I currently think he can completely replace Crayfish. In addition, I have also mentioned to you before that you can build your own personal assistant based on your own common sense. I found that even though the Hermes Agent experience is already very good, there are still some situations where you build your own Agent. Some advantages will be shared with you in this episode. In addition, I will talk to you about the recently launched Opus 4.7. Currently, it seems that its performance on Benchmark is indeed better than 4.6 is even better, but one thing you have to pay attention to is that the cost of the same Prompt in 4.7 will be higher than that in 4.6. I will tell you why in this episode. In addition, haven’t everyone been talking about Opus 4.6’s intelligence declining since last month? We will also discuss with you in this episode whether there is really a decline? What tricks did they do? Why were they discovered? We will all reveal it to you in this episode. If this is your first time coming to our channel, hello, I am Kenji In this channel, we will share some of my thoughts on AI and the technology industry. I hope to learn with you and find out more important information in the world bombarded by crazy information about AI. For those who have not yet subscribed, please subscribe and like the video. Okay, let’s get started without further ado! Welcome to the Silicon Valley Talk Just Kidding Tech. I am Kenji and I am Ke Ke. Here you will hear first-hand experience sharing from the Silicon Valley technology industry. Let’s continue to learn and grow together in the ever-changing technology industry. We will discuss software development, quality development, life in the United States, and news and gossip about technology companies in a relaxed way. If you want to understand the latest trends in the Silicon Valley technology industry, don’t miss it! Hey everyone! Remind everyone that we will attach chapter content of this episode at the bottom of each episode of the video and podcast, so if you are particularly interested in any part, you can jump directly to listen. Last week, our family traveled to the Arizona Grand Canyon for a week because the children happened to be on spring break. American schools are very strange. They may be in nearby schools. But when we traveled this time, it was very interesting from the moment we were at the airport. You saw that almost every group of people had children, from the smaller ones like us who were only two and four years old to the larger elementary and middle schools. There were a lot of people traveling with children. The whole atmosphere was very happy. Then when we arrived in Arizona, we stayed in a pretty good resort for the first two days, which is JW Merritt. We were there, just playing in Piaopiao River and then playing in the water. Then the whole thing was very chill and very happy. There were a lot of children there and everyone was playing crazy. The difference in the weather was very big. When we left Seattle, it was still quite cold. Then when we arrived in Phoenix, the weather became quite warm. I thought wow. It suddenly went from the feeling of autumn and winter to summer again. I thought, wow, actually you have been cold in Seattle for a long time, and you also miss this kind of summer and comfortable weather. In short, we were in this better hotel at the beginning. After playing for two days, I slowly moved towards the suburbs. In the next two days, I went to Sedona. Wow, you can see the red rocks. I think the scenery is really shocking. When you saw it for the first time, you thought, wow, there are really many scenes that you may have only seen in books or in movies. But no matter how many times you see it, you still have to go and see it in person to have that sense of presence. Then you will think, wow, the scene is really magnificent. Then I ran to the Grand Canyon in the next two days. The level of the Grand Canyon was even higher, which made me think, wow, there are really such magnificent views. This very magnificent scenery, you really need to take in more. I will always remember that when I first came to the United States, when I went to Yosemite, I was also blown away by that scene. That is to say, wow, I have never seen such a magnificent scenery since I was a child. The first time was in Yosemite, and I was shocked first. Then the second time I felt this way was when I went to Banff a few years ago. Wow. In Banff, just driving on that highway, I saw the whole mountain. Then when I actually went to their national park to see it, I also felt that the scenery was really beautiful and irreplaceable. Then this was the second time, and this time I went to the Grand Canyon. I think it was the third time. You thought, wow, how can nature really have such miraculous skills? How can a mountain be cut into such a beautiful shape? The top is all flat, and you are looking down from above. Then you see its different canyons and different terrains, and you will feel that many times, why do people say that you just want to go see nature regularly? Because you will have this mentality that you are very small, and you will feel like, wow, we usually have this kind of personnel troubles, what are we doing in the company over there, who has more AI tokens? Then go and get away. Although this is not our choice, it's a bit like because the child is on vacation. You can't stay at home with him for a week, because we don't have the habit of hiring a nanny. I think that instead of staying at home for a week, we all go out to play. And this time I found a big advantage, because when we go out to play, everyone sleeps together. You don't have partitions. Only occasionally one or two days, the hotel where we stay has partitions, but most of the time, we are almost from the beginning to the end of their vacation. We are together for about nine days and ten days, 24 hours a day, and the whole family is drowned together. In this case, it will become more normal when you start to do it. We may all go to bed at nine or ten o'clock together. We may sometimes be a little later, but it is also before ten or eleven o'clock. But what about your spirit, because you get this kind of supplement from nature, and then you don’t have to work in the office like before. You have to be in the company, which consumes a lot of mental energy, so mentally you feel quite happy, so this is an additional benefit. And you also scroll less on your phone. As I have mentioned to you before, we try not to scroll on our phones in front of children, so the time of scrolling is shortened. Then we look at nature, go to bed early and wake up early, and feel like, wow, it really makes the body and mood better. But all this happened after we came back. In fact, it's back to the original point. I found that people are really interesting. It's the good habits you develop there for a week. After you come back, it quickly becomes twelve o'clock, so sometimes you even go to bed later. So this thing is true. We have to find a balance. You just need to go outside to replenish some energy in a timely manner. After you come back, even if you return to a normal routine, you will still have more energy to handle different tasks. Then let's move on to our topic today, Hermes Agent. I first saw this on the English community, and then I discovered that in the past week or two, I also saw on the Taiwanese community that many people started to pay attention to it, share it and use it. Then I tried it myself because I just bought a Mac Studio recently. I originally wanted to run my own Local LOM, but I haven’t tried it since I got it last week. I thought, otherwise I will install and watch Hermes Agent first to see if the real experience is as good as what everyone said. If I put it in one sentence, I think it is really good and I really recommend it to everyone. If you haven't had a chance to try crayfish before and you originally wanted to, don't try it. Just install the Hermis Agent directly because it is almost painless to install. Just make a few comments. Its interface is actually quite similar to clock code. They are both through COI. After you make a few comments and set it up, you can chat with it immediately. Then you can also connect it to your Messaging App. For example, I use Telegram, so I can also communicate with it directly on my phone through the Messaging App. Then I found that what it does quite well is that like clock code, it has a lot of fancy terms, such as dispatch, remote control, and teleport. It communicates and exchanges many sessions between the cloud, mobile phone, and desktop. But actually I think Hermis Agent does a good job, that is, it can also do such things, but its key point is that it all exists in Local and on your computer. So in fact, if you chat on the phone, you can switch back. You can do it on your computer through CLI. So there is no problem with these switches. It just saves a lot of nouns to describe it. But you can do it if you want to. But I later discovered that the demand for this is not very high because it has a good memory and a good ability to search across conversations. It is equivalent to saving almost all your conversations. So sometimes I chat on the computer. After chatting, I switch to the phone and tell him what we just talked about in the conversation. Let’s continue from there, so I don’t need to directly switch to the complete conversation. It can also catch the previous key point, and then we can continue the conversation and continue to complete what we were originally going to do. So I think this is the biggest difference between it and the crayfish, which is its memory ability. I have complained several times before that the memory of the crayfish is very unstable. I have told it several times that I don’t understand it at all. But I think Hermes Agent’s Harness Engineering actually does a pretty good job. The memory is very stable. I have hardly seen any mistakes so far. Occasionally, I need to be reminded once or twice. You have some big decisions to ask me. Don’t make random decisions. But in most cases, I think its performance is very stable. One reason is because I have also connected one of their memory providers, that is, they allow you to choose different providers. Then I chose one called Hong Cho. Because I did some research, Hong Cho seems to be the best provider among all memory manufacturers. It is currently the state of the art and is the best provider. I found that it does not only look at similar past memories through register search. It will also have some in-depth inferences and reasoning based on the context of your chat. It is not just literal reading, because in the past, for example, the Man Zero I mentioned to you before or other solutions to memory problems, the concept is very similar, that is, before you want to send this conversation, it will look at our memory to see if there is anything similar to the register structure. If so, it will inject it into this System Prompt. Then after you finish the conversation, it will put some emphasis on it. If there is something important, you can write it down and put it into the database of this register, so it is more like literally checking to see if there are similar things and then injecting it into the memory. But according to its explanation, Hong Cho's Provider I don't know how much impact it actually has on my chat content. But in addition to doing the above things, it will go into depth to understand the explanation. For example, I said something in one place and another thing in another place. In theory, the words look different, but the principle behind it may be the same. It has a way to deduce it. So this is what I think looks pretty powerful. In short, I integrated these memory functions. I have been using them for about four or five days now. I think the experience is very good. I have mentioned to you before that I have made my own AI Agent because I wanted to try to be my own personal assistant before. I found that my experience was actually pretty good because it has a very big advantage. Let me first tell you how I did it. I used Anthropic’s SDK. They have two SDKs. One is called Client SDK, which is a general chat interface. The other is Agent SDK. In fact, you can think of Agent SDK as the version of Cloud Code's SDK. The method I later studied was that I want to have a main Agent that can chat with me very smoothly all the time without being interrupted. Then my main Agent uses Anthropic's Client SDK because it can eat various file modes, including text and picture files. Then it only needs to do something every time. They will all be delegated to his Sub Agent. I ask him to use all his Sub Agents. Using the Agent SDK means you have a more complete tour capability. Whether you are doing research, writing programs, or debugging, you can do several things at once through this Sub Agent. Spawn multiple Sub Agents. Then I found that this experience is actually very good. You can chat with my Man Agent at any time. Oh, he will not be like us at Cloud Code. Or this will happen in Hermes, that is, he will be interrupted or your message is Cue in it. Because I now find that I don’t like the feeling of waiting, so I designed my main Agent program so that it can chat with me all the time. Even if there are many Sub Agents in its background, it will not affect the execution of the Sub Agent. How do we change our own programs to make ourselves better? So actually the overall appearance is good. So why do I still switch to Hermes Agent? Because I found that Hermes Agent has some benefits. I have to do additional things. For example, it downloads a lot of good Skills or generates better Skills by itself. What about these things? If I want to use my own Agent to do it, I have to reinvent the wheel. I need to find out which Skills are better now and then make them myself. So the first problem is that I have to reinvent the wheel. The second problem is because I am completely using Anthropic's SDK, so I have no way to switch to other vendors, such as OpenAI Gemini or my own Local LLM. Hermes does not have this problem. So, although I think the Agent I made seems to be pretty good, but I found that because Anthropic is now we will talk about it later, he has actually been criticized by everyone recently. He demoted this Opus. We will talk about it later, so I don’t want to say it. I just locked in a certain manufacturer. Then I thought about Hermes Agent, which should be the personal assistant I will use for a long time. I don’t know if everyone is like me. When I saw the name for the first time, I thought that Hermes is not Hermes. Isn’t it that luxury brand? Then I checked it and found that my knowledge is too shallow now. In fact, Hermes is one of the gods in Greek mythology. He is the god of merchants. He is also the god of merchants. He is also a messenger between humans and gods. He is responsible for the bridge of communication between humans and gods. So this is why this Agent is named Hermes Agent. He is to act as a bridge between us and the big circle model. Because this is a part of Harness Engineering, that is, he can talk to us through them in the middle, and then through some of his detachments and loop mechanisms with the big circle model, they try to find ways to make this Agent feel smarter. So it makes sense for them to choose this name. Well, I think this allusion is quite interesting. There is a pretty good design in it, that is, he will take the initiative to add or modify existing Skills. When you talk to him, he will find that he has a lot of distractions and needs to use many tools, and he will say that I want to package it into a Skill. So I think the good thing is that every time I talk to him, after a few sentences back and forth, he will write a Skill by himself or update his memory, so he makes this Agent. He is actually quite proactive and can know your preferences. Then as your conversation gets longer and longer, he will know what your preferred thinking principles and thinking directions are. I will give him a principle: I don’t need you to cater to me or deliberately cater to me. It is best to use the results of your research. If I have different opinions from yours, you can still tell me what you think is good because I found that there is a similar situation in Cloud Code. If you find his program and you tell him a problem, he will say that you are right. Right? What you said is right. I should change it to another one. Now, I think in Hermes Agent, because I have joined this, the main principle is not to cater to me. In fact, his performance is pretty good. Of course, you can do it in Cloud Code. It’s just because I was using the company’s repo in Cloud Code at that time. I did not add this principle. But I still want to remind everyone here that the speed of AI progress is really very fast. I will tell you now that in this period of April 2026 Do you think Herms Agent is good? Maybe the world will be completely different in three months, but I think it doesn't matter. The process will change. But I think the core principles have not changed. That is how they do Harness Engineering and how to make this Agent easy to use. It is actually quite worthwhile to see how they do it and learn how to learn it. Recently, in fact, OpenAI's Codex announced that they are going to make a major revision. They announced that they are improving in a new direction. That's what we are talking about now. Is Harness Engineering what we are talking about is this AI? The middle layer of Agent is different from the big circle model. It's just divided into two layers. But they said that Harness Engineering should be divided into more details, into the Harness part and the Computer part. The Harness part is to make this looping mechanism. Then how to use these tools to communicate with Da Yuan. It is equivalent to the closest layer outside the head of the Da Yuan model, which is the Harness layer. What is Computer? This is actually the execution of a certain comment. For example, if you download a Git comment or NPM Comment of install a certain package. This is the layer where your Round time is somewhere. I don’t know if you can’t understand this episode, or if you don’t have much technical background. Anyway, most of the existing designs are Harness and Journey, which is your Harness layer. You can execute this layer now, and everything is placed in the same place. For example, my own Mac Studio. Its Hermis Agent runs on it, so how does it communicate with the big circle model? How does it optimize Harness and Journey? And it is actually executed in the same place, so that makes sense. If you do it in the same place, of course it is more efficient. You have everything. But what are the risks? If a certain place breaks down today, your secrets may be directly leaked. OpenAI’s latest Agent SDK improvement means that we should be separated. My brain thinks that this thing is running in a certain container and a certain environment. Then your secrets will also run in it. Then when you actually want to execute it, you open another sandbox, so if you have several comments to execute and several tours to do, you can open several different sandboxes. In fact, these sandboxes do not have your secret. They just know that it is going to execute a certain comment, so it failed to execute today. Or if a certain container was hacked today, it actually has no way to know where your main harness logic and your password are. So this is a concept they proposed. Then we can continue to observe and see if this kind of design trend is a good direction. I personally think that if it is a simple personal project, there is actually no difference, because we are now wrapping everything together for simplicity and convenience. But if you are going to enterprise production today, I think such a separation system is safer. I think it is also possible that this kind of agent SDK will develop in the future, because I have found that the strategies of Cloud Code or Anthropik are similar to OpenAI actually copies many things from each other, especially Cloud Code. We have talked about it before. After the source code leaks, I believe that these designs will soon be seen in Codex. Didn’t their Cloud also change their desktop version of the App recently? It will also look slightly more similar to Codex. If you actually use it, you may feel it. So I think it is really interesting now. In the past 10 years ago, Google and Facebook were competing. The two giants are competing. Now, the AI is actually Open AI and Anthropik, the two giants are competing, so it is very interesting. You can see the evolution of this in the past. It used to be the confrontation of mobile phones and social media. Now it has become the confrontation of AI. We can see how these two can sustain their own interesting sparks and news. We will continue to share with you on this channel. Follow. Let me share a little story with you. When I originally bought Mac Studio, I wanted to run Gemma 4. Because it was released about two weeks ago, everyone's initial comments seemed to be good, that is, there is finally a model that can run locally without too much memory. That's because its parameters are only 31B. So in theory, you may have about 24GB or 36GB of memory, and it can actually run pretty well. But just this week, another stronger open source model came out, which is Quen 3.6 launched by Ali. Its parameters are actually the same as Gemma 4. The strongest model is almost the same. Quen is 35B and Gemma 4. The strongest one is 31B. But currently, I see that their Benchmark, at least in the Coding part, Quen 3.6 completely crushes Gemma 4. Of course, this is reasonable, because after all, Gemma 4 is a more general model. It is not a Coding model. It does not say so. It just means that if you want to Coding today, what if you run Local LUM? Quen 3.6 This is really a complete crushing, not just by an order of magnitude. I remember that many Benchmarks were worse than ten or twenty points. At least on the Benchmark, it seemed to be a lot worse. Then I thought, wow, these two The release of Quen 3.6 is actually only 12 days away. The competition for open source models is too fierce. I haven’t installed it in just 12 days. My hands and feet are really too slow. There is a new model that can completely crush Gemma 4 in terms of coding. Then Quen 3.6 uses NOE this time, which is a hybrid expert mode. So every time you make inferences, you only need to enable the parameter of 3 Billion. What does this mean? That is, your inference speed will be very fast because you don’t have to enable so many parameters at one time, so what you actually feel is that your motion response will be faster, especially when you run it locally. In theory, your resources are not as good as those running in the cloud. These servers are so powerful, so using NOE will make you want to run these big circle models locally. Those people, including me, may be able to feel a better experience. After talking about the open source model, of course we can’t talk about it this week. The most powerful B-circle model Opus 4.7 is officially released. Please note that this 4.7 is not the Mythos mentioned before. The Mythos mentioned by Anthropik before is actually still in the internal testing stage. There are only a few big companies. They have the right to use this strongest model with them because they feel that this model is too powerful and they are afraid that it will be abused if it is released to the public. So this 4.7 can be thought of as a safe version of Mythos. A castrated version, but it is on many Benchmarks. Every indicator is ahead of 4.6. One of them performs best, which is the accuracy of image recognition. It jumps from 55% in 4.6 to 99% in 4.7. I feel that they should find a way to jump directly or retrain to improve their recognition ability by more than one level. In addition, there is one thing that needs to be paid attention to because many people may not have any special thoughts. Then you switch from 4.6 to 4.7 and you find out how much your Token usage costs. How about the cost of each API use? Let's talk about it. Why is it all used up all at once? There is a reason for this. They changed to another new Tokenizer in 4.7, which is the tokenizer. The tokenizer means that you take a piece of text and how do you cut its Token? So they say the same Prompt comes in. What about the tokenizer of 4.7? 1 times to 1.35 times of 4.6. In other words, for the same prompt, you may consume more Tokens, so everyone should pay attention here. After you switch to 4.7, you should feel that your Token consumption speed becomes faster, and the money you spend becomes more. Here you can take a look at some of Enthloppy’s recent very criticized moves, because OPS 4.6 was released in February, but since 3 In the past month, everyone started to feel that they have become stupid after using it. Why can't the problems that OPS 4.6 could solve before could not be solved later? This is actually true because someone later analyzed 7000 Sessions in Cloud Code and found that in March, the Thinking Token of these models was actually about 60% to 75% less. There was a very large drop. Many people also analyzed it from different angles and discovered that, in fact, when it was used in March, OPS 4.0 almost stopped thinking. Then someone discovered that Enthloppy was silently changing the Thinking Effort from High to Medium, which would force the model to be in the Default state and use fewer Tokens to think. So you will find that he often changes it without even looking at the program code. Or there are some tools of his. I didn’t look at the Documentation at all. He just wanted to use those Tools directly, which would cause a lot of different errors, so the overall feeling became stupid. Like me, I happened to not encounter this problem because my own approach is like I am too lazy to ask him for Skip Permission every time, so I set Effort, for example, to Max by default. That is, every time I open a new Session, I will be the largest Thinking Effort, and I will always use Bypass Permission. status so that I don’t have to change it back and forth. Of course, everyone should pay attention to this when using it. I think because I will observe some of his status and if he does something today, I will change his direction in time. So there is still a certain degree of risk. If you want to use it, you will remember it. So I myself have not felt the difference in the change of Thinking Effort because I always set it to Max. But if you have not changed it, you may open a new Session. It is originally Thinking. Effort is High and now it becomes Medium. It is not so thorough. You will naturally feel that its ability has become weaker physically. Then with the release of Opus 4.7, they added a new Level. It's between High and Max. He added an Extra High Then bring up a new model and at the same time increase the depth of this thinking. Then you will naturally feel good. In addition, as we just mentioned, if you have the same Prompt, your Token consumption will increase, so the Token consumption will increase. Then you have to think more. You may have the same Prompt overall. When converting from 4.6 to 4.7, you may spend more money. So this is why everyone starts to feel a little bit more and more unhappy about this Enthloppy. Some people even switch to it. Codex, like OpenAI, their chief operating officer said it in an internal meeting. He said that everyone’s love for Enthloppy is almost a belief, right? Now everyone wants to say that you can use Quark Code directly. But we know from past experience that if you support a certain vendor or company too much, how do you know that it will not become bad in the future? Just like OpenAI, in the beginning, it seemed that everyone was using Chepgy and using it well. After that, some turmoil broke out one after another, and Sand Almond's popularity may have declined recently, so everyone went to Cloud Code first, and then went to Enthloppy first. But will you find out later, one year later, two years later, that Enthloppy is now worshiping his own character and doing some things that seem a little bit shabby? Because things like their change to the Effort Level were actually not announced anywhere, and everyone discovered it on their own. So it’s a little weird. Of course we can interpret it as a good direction. They may be entering a startup stage now. The focus is to generate more Features. They may have made this change without thinking too much. It was just an unintentional mistake. But if you look at history, some companies, such as Meta or other technology companies, may be discovered in a few years. It turns out that many people there already knew about this matter at the time. They deliberately wanted to be bad. Of course we don’t know this. It’s just the phenomenon I observed. Enthloppy seems to be losing character a bit. Some of the decisions they have made recently are very poor in their own stability. So we will see if this will give Open AI a chance to catch up. Because we are really paying a lot of Premium to give Enthloppy the same model. In fact, others are doing pretty well, but if you use Opus, it will be very expensive. So I think cost will be a big consideration later on when everyone really continues to import it on a large scale. Are you willing to pay this Premium now? Will you wait until Local LLM or other manufacturers can get the same level of capabilities, but in the absence of cheaper prices? In fact, these companies or these individuals will definitely try to find a way to save money and get high model capabilities. So we will see how long this trend lasts. We will continue to observe it for everyone. Okay, I am very happy to chat with you today. If you are interested in the content of this episode, you are interested in Hermes Agent, Crayfish, the Agent built by yourself. Or this kind of open source and closed source model. If you have any ideas, please tell me below. I hope to see everyone's discussion, because in our discussion area, I believe there are many Crouching Tigers, Hidden Dragons. If you ask these questions, maybe others have some better answers. You can share with each other below. Okay, I am Kenji. Let's see you next week. Bye bye.

今天我們來聊 OpenClaw 小龍蝦的殺手 Hermes Agent 最近我買了一台Mag Studio 就想要在上面跑我自己個人的AI助理那我這次選擇的是Hermes Agent 因為在社群上大家瘋狂的討論我發現它的安裝體驗, 它出廠的能力真的是體驗太好了比起 OpenClaw 小龍蝦我之前痛苦的設定是好非常非常的多它的記憶力非常的好你在不同對話之間講的內容他都記得你除了可以在電腦裡面用CLI 跟他在裡面的內容 你也可以在 Telegram 或是任何你喜歡的 Messaging App 上跟他溝通所以我目前覺得他可以完全取代小龍蝦另外我之前也跟大家提過某個自己常識建立自己的個人助理那我發現即使Hermes Agent 體驗已經算很好了但是還是有一些情況是自己建立的 Agent 有一些優勢都會在這集跟大家分享另外會跟大家聊到最近推出的 Opus 4.7目前看起來它在 Benchmark 上的表現的確是比 4.6 還要好 但是有一點你要注意的就是同樣一段 Prompt 在 4.7 所花的費用會比 4.6 還要高我這集就會告訴你為什麼 那另外最近大家不是都在談說 Opus 4.6 從上個月開始智力就下降了嗎?我們也會在這集裡面跟大家討論真的有下降嗎?他們到底做了什麼手腳?為什麼會被大家發現?我們都會在這集跟大家揭露 那如果你是第一次來到我們頻道的話你好我是 Kenji 在這個頻道我們會分享我對 AI 對科技業的一些心得跟分享 希望跟大家一起學習在 AI 瘋狂資訊轟炸的世界裡面能夠找出比較重要的資訊提供給大家 那還沒訂閱的朋友請你訂閱起來並且把影片按下喜歡 好 我們廢話不多說我們就開始吧! 歡迎收聽矽谷輕鬆談 Just Kidding Tech 我是 Kenji 我是柯柯 在這裡會聽到來自矽谷科技業第一手的經驗分享 一起在瞬息萬變的科技業持續學習與成長 我們會用輕鬆的方式討論 軟體開發 質額發展 美國的生活以及科技公司的新聞和八卦 想要了解矽谷科技業最新趨勢的你 千萬不能錯過喔! 嘿大家好! 提醒大家我們每一集的影片跟Podcast底下都會附上本集的章節內容 所以如果你有對哪一段特別有興趣的話 您直接跳過去收聽 那上個禮拜呢 我們全家人去Arizona 大峡谷旅遊了一個禮拜 因為小朋友剛好放春假 美國的學校很奇怪 就他可能都是在附近的學校 他放的時間不一樣嘛 像我們學校是上個禮拜 那也有聽到一些朋友是什麼上個月或是兩個禮拜前的 感覺上是為了分流啦 但是這次我們去旅行的時候 就從在機場的時候就非常有趣 你就看到幾乎每一組人 都是有帶小朋友的 從比較小的 像我們這樣只有兩歲四歲的 到這種比較大的國小國中的都有 就是一堆都是帶著小朋友出遊的 那整個氣氛就是非常的歡樂 然後我們到了Arizona以後呢 我們前兩天先住一個還不錯的resort 就是JW Merritt 我們在裡面呢 就是遊玩啊 飄飄河啊 然後在裡面玩水 然後整個就是非常的chill 很開心啦 那裡面也是很多小朋友大家都玩瘋了 那這個天氣的差異就非常的大 我們離開西雅圖的時候還是蠻冷的 然後到了鳳凰城的時候 整個天氣是變得還蠻暖的 就覺得哇 整個一下子從那種秋天冬天的感覺 瞬間又跑到夏天了 就覺得說 哇 其實你在西雅圖冷久了 也是有點懷念這種比較夏天 比較舒服的天氣 那總之呢 我們就是一開始在這種比較好的飯店 先玩了兩天以後呢 慢慢玩比較郊區的方向移動啦 後來的兩天跑去了Sedona 哇 可以去看到它那種紅色的岩石 我覺得那種景色真的是 你第一次看的時候真的是覺得蠻震撼的 就覺得說 哇 原來真的有好多景象是你可能以前只有在書本上看到 或是在影片上看到的 可是看再多你還是必須要實體的去現場看 你才有那個臨場感 然後你就覺得說 哇 真的是景象非常的壯闊 然後後來的兩天又跑到大峽谷 那大峽谷的這個level又更往上一截了 就讓我覺得說 哇 就是真的有些這種大景 這種非常壯闊的景色 你真的是定局需要補充攝取一下 我永遠記得就是我第一次到美國的時候 那時候去優勝美帝 Yosemite 也是被那個景象震撤住 就是說 哇 我從小到大從來沒有看過這麼壯闊的景色 然後第一次就是在優勝美帝 先被震撤到 然後第二次有這種感覺是前幾年去班夫 哇 在班夫 光是在那個高速公路上開 就看到全山環繞 然後實際上去他們國家公園看的時候 也覺得這個景色真的是很漂亮 無法取代 然後這個是第二次 然後這一次去大峽谷 我覺得算是第三次 你就覺得說 哇 怎麼大自然真的是有那種鬼斧神功 怎麼可以把一片山可以割成這麼漂亮的形狀 上面全部都是平的 然後你是從上面往下看 然後就看到它不同的峽谷 不同的地形 你就會覺得說 很多時候就是 人為什麼會說你就是定期要去看一下大自然 因為你就會產生這種你很渺小的心態 你就覺得說 哇 我們平常這種人事的煩惱 我們在那邊公司瞎忙 在那邊搞什麼 誰的AI token稍比較多 這件事情在大自然底下一點意義都沒有嘛 你光是為了生存 或者是你光是看到這個美景 你就會覺得 哇 這個東西是非常的值得了 所以我就像這樣子定期去放假 然後去抽離一下 雖然說這個也不是我們選擇的 就有點像是 因為剛好小朋友放假嘛 你總不可能跟他關在家一個禮拜嘛 因為我們也沒有請這種保姆的習慣嘛 我覺得與其關在家一個禮拜 不如我們大家一起出去玩 而且這次出去玩 我就發現有一個很大的好處 因為我們出去玩的時候就大家都睡在一起嘛 你沒什麼隔間嘛 只是偶爾有一兩天我們住的飯店是有隔間的 但大部分的時間我們就是 幾乎是從他們開始放假到結束 我們大概在一起九天十天 二四小時 全家人全部溺在一起 那在這個情況下呢 你們做起就會變得比較正常 就我們可能都是一起九點十點睡 那我們可能有時候會晚一點點 但是也是十點十一點以前 你就一定會上床睡覺 然後你就發現 哇 其實早睡真的是會致百病耶 你雖然整體的休息時間沒有變多 然後你整天在外面也玩了一天 物理上就是你身體是會 還是會勞累 還是會想睡覺 可是你精神上呢 因為你得到了這種大自然的補充嘛 然後你又不用像之前做辦公室 你就是要在公司裡面就是腦力要消耗很多 所以精神上反而是覺得是還蠻開心的 所以這算是額外的一個好處啦 然後你也比較少滑手機 因為之前也跟大家提過 我們在小朋友前面是儘量做到不要滑手機 所以滑手機的時間變短 然後看看大自然 然後早睡早起就覺得說 哇真的是讓身體跟心情變得比較好 但這一切呢就在我們回來之後 其實又回歸到原點了 我發現人真的是很有趣 就是你在那邊一個禮拜養成的好習慣 回來之後呢 很快又回到變成是十二點 所以甚至有時候搞到比較晚一點才睡覺 所以這件事情是真的 我們要去抓這個一個平衡啦 你就是適時的去外面補充一點能量 回來以後呢即便是回到正常的routine 你還是可以有比較多的能量 可以去處理不同的工作啦 那接下來進入我們今天的主題 Hermes Agent 這個呢我一開始是在英文的社群上先看到 然後發現過去一兩禮拜呢 在台灣的社群上也看到蠻多人開始注意它 分享它使用它 然後我自己就嘗試了一下 因為我最近就是剛好買了Mac Studio 本來是想要跑自己的Local LOM 但是一直上禮拜拿到以後都還沒嘗試 就想說那不然我就先來裝裝看Hermes Agent 看真的體驗有沒有大家講得這麼好 那如果一句話來講的話 我覺得是真的不錯 是真的推薦大家 如果你之前還沒有機會嘗試過的小龍蝦 你本來想嘗試的先不要 你直接去裝Hermis Agent就夠了 因為它幾乎是無痛安裝啦 你就下幾個comment 它的介面跟clock code其實也是蠻像的 都是透過COI 你下幾個comment設定好之後 你就可以馬上很快的跟它聊天 然後你也可以去串接你的Messaging App 像我就是串Telegram 所以我也可以透過訊息的App 直接在手機上跟它溝通 那我發現它做的蠻好的一點就是 像clock code它有很多花俏的名詞 比如什麼dispatch、remote control、teleport 就是把很多雲端、手機跟桌面之間的Session 互相溝通互相換 但其實我覺得Hermis Agent做得很好一點 就是它也可以做這樣的事情 但是它的重點就是它全部都存在Local 存在你的電腦上面 所以其實你在手機上聊的你可以切換回去 你在CLI你的電腦上面做 所以這些切換都是沒有問題的 只是它省了比較多這種名詞來描述它 但是你要做到這件事情也是可以的 但是我後來發現這件事的需求其實不太高 因為它的記憶力很好 然後跨對話搜尋的能力也很好 它等於是把你所有的對話幾乎其實都存起來 所以像有的時候我本來在電腦上聊 聊一聊之後我換到手機我就跟他說 我們剛剛那個對話我們講的什麼 我們從那邊繼續開始 所以我也不需要直接切到完整的對話 它也可以抓到之前的重點 然後我們也可以繼續對話 然後繼續把我們本來要做的事情繼續完成 所以我覺得這一點就是體現它跟小龍蝦最大的不一樣 就是它的記憶能力 我之前也抱怨過好幾次 就是小龍蝦的記憶力是非常的不穩定 跟它講了好幾次完全不懂 但是我覺得Hermes Agent呢他們的Harness Engineering其實做得蠻好的 記憶力是非常的穩定 我目前幾乎沒有看到什麼錯誤啦 偶爾會有一兩次我需要提醒說 哦 你有一些大的決定要問我不要隨便亂決定 但是大部分的情況下我覺得它的表現算是非常的穩定 那有一個原因是因為我也串接了他們的一個記憶力的Provider 就是他們可以讓你選不同的Provider嘛 那我選了一個叫做Hong Cho 因為我研究了一下 Hong Cho好像是所有記憶力廠商裡面 它目前算是state of the art 就是最好的Provider啦 我發現它不只是透過語域搜尋去看類似的過去的記憶 它還會根據你的聊天的脈絡 有一些深度的推論跟推理啦 就它不是只是看字面上 因為以前這些比如說我之前跟大家提過的Man Zero 或是其他這種解決記憶力呢 它的概念都很像 就是說在你要傳這個對話之前呢 它去看我們的記憶力裡面 有沒有存過類似的語域結構的東西 有的話就把它注入到這個System Prompt裡面 然後你聊完這個對話之後呢 它又會稍微把一些重點 如果有重要的東西就可以記下來 記到這個語域的資料庫裡面 所以都比較像是字面上去看有沒有相似的東西 然後注入到記憶力裡面 但是呢這個Hong Cho的Provider 根據它的講法 我不知道實際上對我的聊天內容影響多大 但是它除了會做以上那些事情之外 它會去深度了解說 比如說我在某個地方講了某件事 另外一地方講了另外一件事 它理論上是字看起來不一樣 但是它背後的原理可能是一樣 它有辦法去推論出來 所以這就是我覺得看起來還蠻厲害的一點啦 那總之呢 那我就整合了這些記憶力功能 那目前大概用了四五天左右 我覺得體驗都非常的好 那之前我跟大家提過嘛 我自己是不是有做一個自己的AI Agent 因為我之前想嘗試做自己的個人助理 那我發現呢其實我那個體驗也算是蠻好的 因為它有一個非常大的優勢 我先跟大家講我是怎麼做的 就是我是使用Anthropic的SDK 他們有兩個SDK 一個叫做Client SDK 就是一般的聊天介面 另外一個是Agent SDK 那Agent SDK其實你就把它想成是 Cloud Code的SDK的版本嘛 那我後來研究的方法是 我要有一個主Agent可以一直非常順暢的跟我聊天 不會被打斷嘛 那我這個主Agent呢就是使用Anthropic的Client SDK 因為它可以吃各種不一樣的檔案模式嘛 就是文字圖片檔案都可以 然後呢他每次只要做什麼事情 都會下放給他的Sub Agent 他所有的Sub Agent呢我都是請他使用Agent SDK 那就是有比較完整的Tour的能力啊 你不管是去做Research啊寫程式啊 或是Debug 都是透過這個Sub Agent 那可以一次做好幾件事情 Spawn好幾個Sub Agent都可以 那我就發現這樣的體驗其實 非常的好 就是你隨時跟我的Man Agent聊天 喔他完全不會像我們在Cloud Code啊 或是在Hermes也會發生這個事情 就是他會被打斷 或是你的訊息是Cue在裡面的 因為我現在發現我不喜歡等待的感覺 所以我那時候就把我的主Agent設計程式 可以一直跟我聊天對話 那即便他背景有很多Sub Agent 子Agent在執行的時候都不會影響 那子Agent如果做完了 他就會馬上回報給我說 我們在目前的狀況是怎麼樣啊 我就可以第一時間收到訊息 那這個體驗其實蠻好的 而且我也有設計一些 他會自己去進化改進的一個機制啦 就是他看到我們這個地方Mess 他還是會自己想說 我們要怎麼自己更改自己的程式嘛 讓自己變得更好 所以其實整體的樣子是不錯的 那為什麼我還是會換到Hermes Agent呢 因為我發現Hermes Agent有一些好處是 我還要額外去做的 比如說他自己有下載很多好的Skill 或是他自己產生的比較好的Skill 那這些東西呢 我如果是要自己用自己的Agent去做的話呢 就是還要重造輪子啦 就是要去看說現在哪些Skill比較好 然後自己再做出來 所以第一個問題是我必須要重造輪子 那第二個問題呢 是因為我完全是使用Anthropic的SDK 所以我沒有辦法去換成其他Vendor 比如說OpenAI Gemini 或是自己的Local LLM 那Hermes就沒有這個問題啦 所以雖然說呢 我覺得我自己做的Agent好像也還不錯 但是我發現因為Anthropic現在 我們等一下也會聊到 他其實最近有被大家抨擊 就是他把這個Opus降置嘛 我們後面會展開再聊 所以我就不想要說 我直接Lock in某一個廠商 那後來想想Hermes Agent 應該是我長期會使用的這個個人助理啦 那我不知道大家有沒有跟我一樣 我第一次看到這個名字的時候 我就想說Hermes不是愛馬仕嗎 不是那個奢侈品牌嗎 那後來我查了一下 發現我現在是知識實在太淺薄了 那其實Hermes呢 是希臘神話裡面的其中一個神 他是商人之神 那他同時呢 也是介於人跟神界裡面的信使啦 就是他負責的是人跟神裡面溝通的橋梁 所以這就是為什麼 這個Agent要命名為Hermes Agent 他就是要扮演我們跟大圆模型之間的橋梁嘛 因為這算是Harness Engineering 的一個部分嘛 就是他可以在中間透過他們跟我們的對話 然後透過他跟大圆模型的一些脫扣啊 loop的機制 想辦法去讓這個Agent 感覺是更聰明 所以他們取這個名字算是有其道理的 那我覺得這個典故算是還蠻有意思的 那他裡面有一個蠻好的設計呢 就是他會主動去新增或是修改既有的Skill 就是當你跟他對話的時候 他發現他有很多脫扣 要用好多工具的時候 他就會想說那我是要把它包成一個Skill 所以我覺得蠻好的一點是我每次在跟他對話 進行幾句來來回回以後呢 他就會自己寫一個Skill 或者是把自己的記憶更新 所以他就讓這個Agent呢 他其實算是蠻主動積極的可以去知道你的偏好 然後隨著你們對話越來越久 他就知道說原來你喜歡的思考原則是什麼 思考方向是什麼 像我就給他一個原則就是 我不需要你迎合我 刻意迎合我 你最好是用你研究過來的結果 如果我跟你的意見不一樣 你還是可以把你覺得好的想法跟我講 因為我發現其實像是在Cloud Code 也會有類似的狀況嘛 你發現他的程式嘛 有一個問題你跟他講 他就會說你說的對 對不對 你說的對我應該換成另外一個 那現在呢我覺得在Hermes Agent裡面 因為我有加入這個主要的這個原則 就是不要迎合我以後呢 其實他的表現算是還不錯啦 那當然你在Cloud Code也可以啦 只是因為那時候我在Cloud Code 是用公司的這種repo去做 我就沒有加這個原則了 但是我在這邊還是要提醒大家 現在AI進步的速度是真的非常的快 我現在跟大家說 在這個2026年4月這個時期 你覺得Herms Agent不錯 說不定再過三個月 這個世界是又完全不一樣了 但是我覺得沒關係 這個工序會化嘛 但是我覺得核心原則是沒有改變的 就是他們怎麼做Harness Engineering 怎麼把這個Agent弄得好用 其實是還蠻值得去看他們怎麼做 怎麼學習的 那最近呢 其實像OpenAI的Codex 他們就宣布他們要大改版了 他們宣布他們往一個新的方向改進 就是我們現在呢 講的Harness Engineering 是不是就是把這個AI Agent 中間這層跟大圓模型分開 分成兩層而已嘛 但他們說其實應該要把Harness Engineering 分得更細 分成是Harness的部分跟Computer的部分 Harness的部分呢 就是去做這個Looping的機制啊 然後怎麼樣去使用這些工具 跟大圓溝通 它等於是在大圓模型的這個腦袋以外 最接近的這一層啊 就是Harness這一層 那Computer是什麼 Computer呢 這個其實就是它實際上在執行某一個comment嘛 比如說你去下Git的comment 或者是NPM install某個package的comment 這個就是你的Round time在某個地方這一層 我不知道這一集大家會不會聽不太懂 或是Loss 如果你沒有太多技術背景的話 反正呢 現有大部分的設計都是Harness and Journey 就是你的Harness這一層呢 跟你現在執行這一層 全部都放在同一個地方嘛 就比如說我自己的Mac Studio呢 它的Hermis Agent就跑在上面的 所以它怎麼跟大圓模型溝通 它怎麼優化的Harness and Journey呢 以及它實際上執行 它都是在同一個地方嘛 那這樣也合理嘛 你在同一個地方做 你當然是效率比較快 你擁有所有的東西 但是呢 會有什麼風險 就是今天呢 如果某一個地方壞了 你的可能 你的secret就直接洩漏出去了 那OpenAI他們最新的Agent SDK的改進是說 我們應該要分開 就是我大腦覺得這件事在某一個container 某一個環境去跑的嘛 那你的這些secret也是跟著在這裡面去跑 然後你實際上要執行的時候 是開另外一個sandbox 所以你有好幾個comment去執行的話 你好幾個tour要做的話 你是可以開好幾個不同的sandbox 那這些sandbox裡面 其實它並沒有你的secret 它只是知道說 它要去執行某一個comment 所以它今天執行壞了 或者是它今天某一個container被hack的話 它其實沒有辦法知道 你主要的這個harness的邏輯 跟你的密碼在哪裡 所以這是他們提出的一個概念啦 那我們就可以持續觀察看 這樣子的設計趨勢 會不會是一個好的方向 那我個人認為呢 如果是對個人的簡單的project來講 其實沒什麼差嘛 因為我們現在就是為了簡單方便 所以把所有東西都包在一起 那但是如果你是今天是要跑到 企業用的production 那我覺得這樣的分離制度是更安全的 那我覺得也是有可能 未來這種agent SDK 會發展的趨勢啦 因為我就發現像現在 Cloud Code 或是說 Anthropik 他們的策略 跟 OpenAI 其實他們很多東西 也是會互相抄來抄去啦 尤其是像Cloud Code 我們之前聊過嘛 原始碼洩漏之後我相信這些設計裡面 很快就會在 Codex 裡面去看到 那他們Cloud 最近不是也 把他們的桌面版本的 App 也改了嗎 那也是跟 Codex 看起來會稍微比較像一點啦 那這個大家如果實際上去使用到 可能就有感覺 所以我認為現在真的很有趣啦 就是之前10年前的時代是 Google 跟 Facebook 在做競爭 這個兩大巨頭在競爭 那現在這個 AI 實際上呢 是 Open AI 跟 Anthropik 這兩大巨頭在競爭 所以非常有趣 你就可以看到這個以前的演變 以前是 Mobile 手機 社群媒體的對抗 現在就變成是 AI 的對抗 我們就可以看這兩個呢 能夠持續自己怎麼樣的有趣的火花跟新聞 我們都會在這個頻道持續跟大家 Follow 那這邊跟大家分享一個小故事 就是我本來買 Mac Studio 的時候呢 是想要跑 Gemma 4 因為它大概是在兩禮拜前發表嘛 大家的一開始的評價看起來都是不錯的 就是終於有一個在本地不用太多記憶體就可以跑起來的模型 那因為它的參數只有 31B嘛 所以理論上呢 你可能記憶體大概 24GB 或者是 36GB 其實都可以跑得還不錯 那但是呢就在這個禮拜 有另外一個更強的開源模型出來了 就是 Ali 推出的 Quen 3.6 它的參數的量其實跟 Gemma 4 最強的模型差不多 Quen 是 35B 然後 Gemma 4 最強的是 31B 但是呢目前我看他們的 Benchmark 至少在 Coding 的部分 Quen 3.6 是完全碾壓 Gemma 4 那這當然也合理嘛 因為畢竟 Gemma 4 是比較走一個通用的模型 它不是走 Coding 的模型 它也沒有自己這樣講 只是說如果你今天要 Coding 跑 Local LUM 的話呢 Quen 3.6 這真的是完全算是碾壓 不只一個數量級 我記得蠻多 Benchmark 都是差到十幾二十分以上的 至少在 Benchmark 上看起來是差很多 然後我就覺得說 哇 他們這兩個的發布其實才差 12 天 就是開源模型的競爭也太捲了吧 短短 12 天的時間我都還沒安裝好 我的手腳真的太慢了 居然就有一個新的模型 你可以完全在 Coding 方面是碾壓 Gemma 4 了啦 那 Quen 3.6 它這次使用 NOE 就是混合專家模式嘛 所以你每次在推論的時候 其實只需要啟用 3 Billion 的參數 這代表什麼意思呢 就是你的推論的速度會很快嘛 因為你就不用一次啟用這麼多的參數 所以你在實際上的感覺是 你的體感上它的回復的速度會更快 尤其是你在 Local 跑的時候 你的資源理論上是沒有這些在雲端跑的 這些伺服器這麼的厲害 所以它採用 NOE 就會讓你想要在 本地跑這些大圓模型的這些人包含我呢 可能就可以感受到體感更好的一個體驗啦 那講完開源模型呢 那我們當然不能講這個禮拜 最強的 B 圓模型 Opus 4.7 正式發表 那這邊要注意一下 這個 4.7 並不是之前提到的 Mythos 之前 Anthropik 說的 Mythos 其實還在內部測試的階段 也只有幾個大公司呢 他們有權利跟他們一起去使用 這個最強的模型 因為他們覺得這個模型的能力太強了 怕如果釋出給大眾的話會被濫用 所以這個 4.7 應該可以想成是 Mythos 的安全版 閹割的版本 但是它在很多 Benchmark 上面呢 每一個指標也都是領先 4.6 的啦 那其中有一項的表現是最好 就是圖片辨識的準確率 從 4.6 的 55% 跳到 4.7 的 99% 這個我感覺他們應該是找到一個方式去直接跳躍 或者是重新訓練 就把他們的辨識能力整個提升了不止一個檔次 那另外呢 這邊有一點要非常注意 因為很多人呢 可能就是沒有特別想 然後你就從 4.6 換到 4.7 你就發現你的 Token 使用的費用呢 你每一次 API 使用的費用呢 你聊一聊 怎麼一下子就用完了 那這個是有原因的 他們在 4.7 是換了另外一個新的 Tokenizer 就是分磁器啦 分磁器就是說你把一段文字 你怎麼去切它的 Token 的 所以他們就說一樣的 Prompt 進來呢 4.7 的分磁器呢 會是 4.6 的 1 倍到 1.35 倍啦 也就是說一樣一段 Prompt 進來 你可能會消耗更多的 Token 所以這邊大家要自己注意一下 你在換到 4.7 以後呢 你應該會感覺到你的 Token 消耗速度變得更快 然後你花的錢變得更多了 那這邊就可以順便帶到 Enthloppy 最近讓人非常詬病的一些舉動啦 因為像 OPS 4.6 是在 2 月釋出的嘛 可是從 3 月呢 大家在使用就陸續去開始覺得說 欸 怎麼好像變笨了 這個 OPS 4.6 怎麼以前可以解決的問題 後來都解決不了 那這件事情其實是真的 因為後來有人去分析 他在 Cloud Code 的 7000 個 Session 發現了在 3 月的時候呢 這些模型的 Thinking Token 就是他思想的花的 Token 居然是少了 60% 到 75% 左右 有一個非常大的一個降幅 那很多人也從不同的角度去分析 就發現 喔 其實真的 當時在 3 月使用的時候 這個 OPS 4.0 幾乎是變得沒有在思考了 那後來有人就發現 原來 Enthloppy 他們在默默的把這個 Thinking Effort 從本來是 High 改成 Medium 就會迫使這個模型在 Default 的狀態下呢 去使用比較少的 Token 去思考 所以你就會發現他常常連程式碼都不去看就去改 或者是有一些工具他的 Documentation 根本沒有去看 他就想說去直接去使用那些 Tool 就會造成蠻多不一樣的一些錯誤 所以整體的感覺就變笨了 那像我呢 剛好我是沒有遇到這個問題 因為我自己的做法是像 我很懶得每次都要跟他下什麼 要 Skip Permission 這件事情嘛 所以我就把 比如說 Effort 都 預設都設成 Max 就是我每次開一個新的 Session 我就是最大的 Thinking Effort 然後我永遠是 Bypass Permission 的狀態 這樣我就不用改來改去了 那當然這個自己大家使用上要注意一下 我是覺得說因為我會觀察他的一些狀況 如果他今天做了一些事情 我是會及時的去改變他的方向的 所以這還是有一定程度的風險 大家如果要使用的話會記得 所以像我自己是並沒有感受到 這個 Thinking Effort 的改變的差異 因為我就是 都是把它設成 Max 可是如果你沒有改的話呢 你可能開一個新的 Session 本來是 Thinking Effort 是 High 現在就變成 Medium 它就沒有想得這麼徹底 你自然就會感覺體感上他的這個能力變弱了 那 隨著他們在 Opus 4.7 的釋出呢 他們加了一個新的 Level 就是在 High 跟 Max 之間呢 加了一個 Extra High X High 那 他把這個預設都切成 X High 所以 這就是很吊詭嘛 你之前把他的思考能力降級 然後就是你放了一個新的模型 然後你又把思考能力 他的 Effort 變高 那當然體感上大家就覺得說 哇 這聰明好棒棒 好厲害哦 所以就有人就覺得說 Enthloppy 是不是有一點壞壞的 就是先讓你覺得這個模型變差了 然後再提一個新的模型 然後同時把這個思考的深度又變高了 那你自然會覺得好嘛 那再加上我們剛剛提到的 你的同樣的 Prompt 你的 Token 消耗會變多 所以 Token 消耗變多 然後你要思考更多 你可能整體一樣的 Prompt 在 4.6 到 4.7 的轉換的時候 你可能又花更多錢了 所以這是為什麼大家開始會 對這個 Enthloppy 越來越有一點 覺得不開心 有的人甚至就換到 Codex了嘛 那像 OpenAI 他們的營運長 自己在內部的會議有講嘛 他說現在大家對於 Enthloppy 的這種喜歡呢 幾乎是信仰了 對不對 現在大家 Default 都是想說 你直接去使用 Quark Code 但是我們從過去的經驗就知道 你如果太過支持某一個 Vendor 某一家公司廠商的話 你怎麼知道它後面不會變得壞壞的 就像是 OpenAI 一開始 好像看起來大家也都是用 Chepgy 用得好好的嘛 那後來陸陸續續有爆出一些風波 然後 Sand Almond 最近可能風頻也下滑了 所以大家就先往 Cloud Code 去走了 先往 Enthloppy 去走了 但是你會不會到後來發現 一年後 兩年後 發現 Enthloppy 又自己現在在 把自己的人品拜光光 做了一些看起來有一點 好像有點猥瑣的事情嘛 因為像他們改這個 Effort Level 這些事情 其實是沒有在任何地方有宣布的 是大家自己去發現的 所以就有一點怪怪的啦 那當然我們可以把它解釋成是往好的方向 他們可能就是現在是進入一個 Startup 階段嘛 重點是產生更多 Feature 他們可能沒有想太多就做出這個改動 只是一個無心之過 但是呢 你如果看歷史的話 有一些公司 比如說像 Meta 或是其他科技公司 可能幾年後就會有人發現 原來很多人當時就在裡面已經知道這件事情 他們是故意要壞壞的 那這個我們當然不知道 只是我觀察到的現象是 Enthloppy 好像有一點在敗人品 他們最近做出的一些決定啊 他們自己的穩定度都非常的差 所以我們就看這會不會給 Open AI 機會來追上嘛 因為我們現在真的是付了蠻多 Premium 去給 Enthloppy 同樣的模型呢 其實其他也做得蠻好的 但是你使用 Opus 就是會非常的貴 那我覺得成本這件事情呢 到後來呢 大家真的繼續在大規模導入的時候 一定會是一個蠻大的考量的 現在大家是願意去付這個 Premium 會不會等到 Local 的 LLM 或者是其他廠商 可以得到一樣程度差不多的能力 可是缺用更便宜的價格的時候 其實這些企業或是這些個人呢 一定會想辦法去找出一個比較好用省錢 可是又可以得到高模型能力的方法的 所以我們就看這個趨勢可以稍多久 我們就持續為大家去觀察下去 好啦 今天很開心跟大家聊天 如果你對本集的內容 你對 Hermes Agent 小龍蝦 自己建的 Agent 或者是這種開源閉源的模型 有什麼想法的話 歡迎在底下告訴我 我很希望可以看到大家的討論 因為我們的討論區 我相信有蠻多臥虎藏龍的 你提出這些問題 說不定別人有一些更好的答案 你可以在下面互相交流分享啦 好啦 我是 Kenji 那我們下周見囉 掰掰