Hermes Agent's style

Welcome to the Tingshu Podcast channel produced by Arbitration Engineers. Imagine this scene. You wake up today and find that the open source project you worked so hard to write is suddenly criticized by thousands of people on the Internet. Hundreds of thousands of people are crazy to repost an article, accusing you of code laundering and saying that you stole other people's hard work. Wow, this is really stressful. Yes, the point is. When we really peel back this layer, it is this kind of public anger, and sneak into the underlying source code to compare it line by line. But it turns out that the core part of what you were accused of plagiarizing is that you don’t even have blood. Well, you don’t even have blood. It only uses less than 600 lines of code in total to call a third-party tool. This is the one that has caused a lot of fuss in the AI agent circle recently. The suspected plagiarism between Norris Research Hermes Agent and the Chinese team Evil Map Evilver engine. Yes, this matter has been discussed very much recently. Really, what about this in-depth analysis? Let’s not talk about emotions, only look at the evidence. Well, let’s guess what this allegation seems to be. In fact, the development trajectory of this incident is very intelligent and interesting. If you look at the popular accusation article, his wording can be said to have hit the emotional pain points of the community extremely accurately. How can you say what words he used? He directly claimed that Hermes’ ten-step group patrol pills are exactly the same. Then he said that the terminology has been completely changed, but the architecture is not fixed. Wow, this sounds really like Stone Legs, right? For the general public who don’t have time to read the original code. When you see such an accusation with pictures and text, your first reaction is definitely to think, wow, this is simply irrefutable evidence. Indeed, but this is the most paradoxical part. If that article has provided such a seemingly perfect comparison chart, why did we get this one today, which is an in-depth analysis note of three related open source code bases line by line, but we will come to the completely opposite conclusion. This is the core problem. Yes, the most basic first step when you want to accuse others of plagiarism is to expose the code of both parties in the sun and then avoid it line by line, right? Logically it should be like this, but the first and honestly the most fatal blind spot revealed in this in-depth note is right here. What blind spot is the core application code that Evolver is accused of attracting plagiarism. The outside world cannot independently verify it. Wait, it cannot be verified. Why is the code not open source? Because it contains key files such as Evolve. If you are a listener today, you may not understand what this concept is. Can you explain it to me? To put it simply, it is a program that humans can understand and convert it into garbled code that can be run by machines but that humans can read. The report specifically mentions that Evolver's Evolve.js is a huge file of up to 728KB. It is a 728KB text file. It is very big. It is very large. And if you open it, the first line will look like this. I will read it to you. It says Constant 0x2b4 E3 equals 0x1cc9 and is followed by hundreds of thousands of garbled codes. Oh my god, this is like, this is like someone accusing you of stealing his ancestral secret recipe, but when the judge asked to see the original secret recipe, he handed over a locked safe and said, Hey, judge, you don’t have to look inside, just believe my description. He definitely stole it. This metaphor is very accurate. Since Chen’s horse is locked, I don’t understand. What is the ten-step cycle comparison in the article? How is it done? This is a very keen question. The answer is that the comparison benchmark of that article is completely based on the text description of the Evolve.js team themselves in their blog. It only relies on the text description. That’s right. So when the matrix side asks the community to trust their architectural description instead of opening the source code for everyone to check, this technical discussion has deteriorated. What has become? It has become a test of faith. Under the strict standards of open source, you absolutely cannot use a black box to prove that others have plagiarized your black box. This is a big flaw in the matrix logic. I understand. It’s because we can’t see the true appearance of A at all, so we can only use B, which is Hermes’ fully public program code, to compare with A’s self-claim. That’s what it means. Well, this note actually looked at the EvolveSkill PY file of Hermes. Did it really copy those ten steps perfectly as the accusation said? What’s fascinating here is that in actuality, not only is it not a perfect copy, but it is different step by step. Every step is different. Can’t even match one step? Let’s look at a few core steps in depth. Evolver claims that their first step is Insure Asset Files. What does this do? It is mainly used to initialize empty files. But Hermes’ actual first step is to find and parse a file in a specific format called SQL.MD. This sounds completely different. One is to see the empty file, and the other is to read a specific configuration file. That’s right. Then in the second step, Evolver does a very complex three-layer signal extraction. What about Hermes? Hermes is building a data set for model evaluation. You see, what the two sides are doing are basically two completely different tracks in terms of engineering logic. But the article must be somewhat similar to convince so many people. Is it true that everyone blindly follows the entire network and forwards it without even reading it? Well, if you insist on looking at it from an extremely abstract perspective, both of them do comply with a big framework. What big framework? That is, first load the data, then optimize, then verify the results, and finally save it. What to stop? That's it. Load, optimize, verify, save, right, that's it. Wow, this is like saying that all compilers have copied the world's first compiler because they have all gone through lexical analysis, syntax analysis and then code generation. This is just the basic process, right? Or it means that when you write an article, you must have an inheritance, a transformation, and a combination, so when you write an article, you plagiarize Li Bai. This is just a basic structure that any data processing pipeline, that is, a pipeline, must follow. Yes, this points out the complex irony that people face when they face it. A common cognitive bias is that people easily mistake the inevitability of a common architecture for the iron needle of plagiarism. Well, it is indeed easy to be misled. Just like we will never accuse a new car of plagiarizing an old car, just because they happen to have four wheels and a steering wheel. Because the car originally looks like that, right. This is determined by the underlying logic of physics and engineering. This is not someone’s original creation. Well, even if the large framework of the process cannot be considered plagiarism, I still want to mention one point. The popular article is the most convincing. The most guilty accusation is that he said that Hermes had systematically replaced 12 pairs of terms. Oh, this part was really hotly debated. You deliberately changed all the variable names. I must admit that this trick is very, very convincing visually, that is, the comparison table. This is indeed the most provocative part of the entire article. It is the so-called AI code-washing textbook. But when we put these 12 pairs of terms back into their original programming context to understand what they are actually doing, what happened? This accusation collapsed instantly like a house of cards. Wow, it's so exaggerated. Let's just take a look at some of the most commonly discussed examples. The article accuses Gene in Evolver, which is G-in, of being replaced by Hermes as SKL.MD. Yes, it sounds like he deliberately replaced an abstract concept with a bad name for an entity to avoid danger. But if you really check it, you will find that SKL.MD is not a word invented by Hermes to deceive others. What is it? This term is actually the Agent Skills open standard released by Anthropic in December 2023. To date, more than 32 tools in the industry have adopted this standard. Which tools are included? It includes mainstream tools such as Cloud Code and Cursor that everyone uses every day. Well, so this is not a replacement term at all, but Hermes has chosen to support the latest common standard in the industry. It is absolutely correct. This is too ridiculous. It is like when I developed a website today and decided to use the standard tags of HTML5. As a result, you came to sue me and said that I deliberately changed the name to cover up the plagiarism of your city. This logic is too jumpy and even more far-fetched. The article said that Evolver is used to track the success status and the complete record of Capsule. The word capsule, right? The article said that it was replaced by the Execution Record in Hermes. The execution record sounds quite similar in function. It sounds very similar, right? But the problem is if you go to the Hermes Self Evolution code warehouse You can't even find the word Execution Record. Wait a minute, the word doesn't even exist. Yes, there is no such variable or category name. So what exactly is Hermes using to record? It can't be recorded. What Hermes actually uses is a scoring data structure called Fitness Score. Asana energy score. Haha. Yes, literally. But it is actually just a very simple scoring mechanism. This is completely unequal in terms of functionality and implementation with Evolver's huge and complex Capsule recorder. So in order to gather these 12 pairs, the accusers even made up some words in their own minds. Yes, they have carried out a very serious cross-system migration program. What is the migration program method? Can you give me another example? For example, the accusation mentioned that Evolver has a very complicated Solidify release pipeline, saying that it was replaced by Hermes' Skill Manage Create. These two sounds completely different. Not only do they sound different, but their actual functions are also very different. The problem is Skill Manage. Create is just the most basic CRUD gadget in the Hermes main repository. It is the most basic database operation tool like adding, reading, updating and deleting. It is just a small script. It is not even in the self-banned repository that was accused of plagiarism. But the most ridiculous thing is not this. There are even more ridiculous things. The most ridiculous thing is that they equate Evolver's Selector with Hermes' Skill Listing function. Wait a minute. I remember writing this paragraph in my notes. Evolver's Selector sounds like some kind of weighted selection engine, used to calculate overlap. Right, that sounds very high-end. So what is the actual code of Hermes' Skill Listing doing? Its function is just to read the folder and make a list of the SKLMD file names in it. That's it. Hey, print out the file names in the directory. A list action that reads the file directory is accused of plagiarizing your weighted selection engine. Yes, it is so far-fetched. This is too ridiculous. No wonder there is that magical metaphor in the note. In other words, it is like accusing React of plagiarizing View. Because the component of React corresponds to the component of View, and the state of React corresponds to the Data of View. This metaphor is really vivid. This is simply a sneaky change of concepts, forcibly tying two unrelated things together. Just to make up a beautiful term replacement comparison table. This is why we have repeatedly emphasized that you must not just look at the lazy comparison chart. When you really peel away these terms to see the ultimate function of the underlying layer, you will find that they are two completely different species. OK, the details are completely inconsistent, and the terminology is also made up. But I still want to play the devil's advocate. OK, you said, the article mentioned an accusation that sounds very technical. He said that the three-layer memory system of the two systems has achieved exact correspondence. Three levels of memory, that is to say, they all have simple knowledge, attachable skills, and searchable history. Even if the previous details are all coincidences, this high degree of consistency in the architectural level cannot be said to be completely without reference. If we connect this with the larger context, you will understand why. Imagine if you were an engineer developing an AI Agent today, sitting in front of the computer. What is the biggest hardware and software limitation you face? The biggest limitation? Well, it is a large language model, which is the context box of LLM. Yes, it is the Context Window. It is impossible for the model to remember all the project details, the huge code, and every conversation you had in the past at once. He will definitely forget it, so you have to take them apart and classify them. To be precise, you have to do this. At the level of information theory, in order to solve this context bottleneck, the most natural classification of knowledge is these three layers. Which three layers? The first is to turn the rules that you must remember permanently into simple knowledge, that is, the vital memory. The second is to encapsulate the actions that can be performed repeatedly into skills, which is procedural, and the third is what happened in the past. Passed into a searchable history, this sounds like human long-term memory. Muscle memory is the same as short-term memory. It is very similar. So this is not a flash of inspiration by a genius team at all, but a kind of inevitable success under the architectural constraints of LLM, which is engineering necessity. So this is the proper term you mentioned in your notes. Independent convergence Because everyone faces the same wall, in the end they can only carve out doors with very similar shapes. Yes, this is the concept. In fact, if you look at the mainstream agent frameworks on the market now. You will find that everyone is doing the same thing. For example, like land chain, it has physical memory, tool and conversation history, orGBT has long-term memory, skill instructions and interaction history. Crew AI also has a similar three-layer division. Wow, so everyone has these three layers. Yes, everyone is building a house under the same set of physical laws. Then the configuration of the living room, bedroom, and kitchen will naturally be highly similar. Treating this industry-common engineering solution as a plagiarism is really a misunderstanding of the technological development context. Okay. Let's dismantle it all the way. The process is universal, the terminology is a patchwork, and the architecture is an engineering necessity. That's where the really interesting part comes. If Hermes didn't plagiarize Evolver's complex system at all, what exactly is installed in Hermes' own city code library? How do they implement the so-called self-evolving engine? Well, when you really go through the Hermes original code, you will find that the truth actually has a little bit of it. How should I put it? Black humor. Black humor. Why. Because of Hermes' self-evolving engine. In essence, it's just a Stanford DSP Y. Right, they didn't move the bricks and cement themselves. They just wrote a specification and handed it over to the general contractor. So what is the core logic of Hermes? It's actually only a two-line program. Wait a minute, two lines. Yes, the core is just two lines. The first line is set to call an optimizer called GPEA in the DSP Y. By the way, this GPEA is a top academic achievement published by the University of California, Berkeley and Stanford University at ICOR2026. Wow. Directly use the results of the Dinghui. The second line is to execute the compilation. It is as simple as two lines of core code. The entire Hermes handwritten Python code. How many lines does it add up to? According to the line-by-line statistics of the notes, it is only about 600 lines. Only 600 lines. These 600 lines also include the basic loader, scoring function and some execution scripts. They even have several planning stages. Currently, the directory is still empty. Wow. What about Evover, which is accused of plagiarism? How big is its magnitude? Evover is a complete GPE protocol engine, which is a gene expression programming engine, built from scratch, brick by brick, using No.js. So it is very big, very huge. There is only one Signals.js file responsible for processing signal extraction. This file alone has 660 lines. It is equivalent to one file being worth your entire project. Yes, this does not include the huge mutation engine behind Evover. There is also the Multilingual Source module. This makes absolutely no sense in terms of technical logic. This is tantamount to accusing a person of using 600 lines of specifications to call a third-party general contractor to wash away tens of thousands of lines of textbooks for another team to build a huge system from the underlying protocol. It sounds ridiculous, right? Is it basically taking the manual of a remote-controlled aircraft to say that he copied the engine design drawings of the Boeing 747? This is why we have always emphasized that as long as you have a little engineering background and are really willing to read the cities on both sides. This will be self-defeating. Indeed, in terms of their engineering depth, implementation path, and even the language ecology they use, one is Python and the other is Node.js, which are products of completely different magnitudes and different universes. But this brings up a question that I can't figure out. What is the problem? If Hermes is really There is absolutely no use of Evolver. Why did that popular article accuse Hermes of seven material principles in an extremely angry tone? Zero mention and zero attribution to Evolmap. If I were a developer of Evolver today and saw a project that was working on a similar direction to mine without even mentioning me, I might also feel a little unhappy. Isn’t it a basic courtesy to pay tribute to each other in the open source world? Here we need to clarify a key open source ethics. Where is the boundary of attribution and citation? Where is the boundary? Imagine you wrote an academic paper on PiTouch deep learning today. Do you need to thank the TensorFlow team in the references? Of course not. Although everyone is doing deep learning, the underlying framework is completely different. Why should I thank it? Absolutely correct. Then let’s check out the actual technology station of Hermes. It relies on DSPI, which is the MIT open source protocol. It relies on the GEPA optimizer and MIPROV2. These also have their own authorizations. Hermes has clear references and tributes to these actual underlying tools in their projects. So they did not pretend that these 600 lines were a magical machine invented by themselves. They honestly explained that they were standing on the shoulders of Stanford University. Yes, they did not mention Evolver because they have no intersection at all in the history of technology. Not at all. Not at all. They did not use Evolver's code, did not call their GEP engine, and even the signal extraction mechanism is different. Just because everyone is studying the self-evolution direction of AI agents, they require mandatory attribution. This is completely untenable in the software engineering world, but the author of that article obviously didn’t think so. He played up this zero attribution as irrefutable evidence that the other party had a guilty conscience and deliberately erased traces. In fact, the most ironic thing about this whole incident is the official attitude of EvolMap. What did the official say? If you bypass the vivid Internet celebrity article and go directly to the official technical blog published by the EvolMap team, you will see a completely different narrative of what they wrote. EvolMap's official wording is actually very rigorous. They admit that the Agent Skills standard preceded them on the timeline, and they also admitted that Hermes established the group advocacy library half a year earlier than they did. Wow, this is very objective, right. They even clearly pointed out that Stanford's GEPA and the GEP they developed are completely different technical routes, etc. So the official team themselves are actually very aware of the technical differences between the two. Not only that, EvolMap's official blog also wrote down the concept we just discussed in black and white. They admit that the possibility of independent convergence exists. They themselves also mentioned independent convergence, yes, and at the end of the article they also emphasized that the text is not legally plagiarized. So what does all this mean? This is crazy. The official team clearly maintains the objectivity and customization of the technical staff and even reserves the space for independent convergence. However, after the fermentation of social media and the translation of self-media that lacks traffic, it actually turned into a solid plagiarism of textbooks and laundering code. It even gave rise to something like Hermes being deceived and deleting its account with a guilty conscience. This kind of rumor is completely groundless. Oh my God, this is really a mistake. That's why we have to do this in-depth analysis. It's not just to help clarify a 600-line project, but also to point out a huge potential crisis faced by the current open source community. Are you referring to the backlash of traffic culture on technical discussions? Yes, imagine if you are a passionate developer today and you find a common pain point in the industry. Then you piece together a simple and elegant solution based on existing open source tools, such as DSPY. Just like what the Hermes team did, yes, just because you face the same engineering constraints and come up with a similar architecture that meets industry standards, you will face the huge risk of being publicly interrogated by Internet celebrity articles and being accused of being a city code thief. This will definitely cause an unimaginable chilling effect. If the architecture is similar or just solves the same problem, it will be vilified by the entire network. Who dares to open source their own projects? Who dares to put their inspirations on Github for everyone to see? Right. Let’s all just shut down and return to the most closed dark age of the software industry. To be honest, in this era of whistle-blowing culture, the threshold for the truth seems to be getting higher and higher. When the catharsis of emotions and the temptation of click-through rates override the rigor of the main bank’s review of the city code, are we protecting the intellectual property rights of open source, or are we personally killing the most precious trust foundation of open source collaboration? Yes, this is a battle because people are willing to calm down and read the city code. Just looking at the sensational headlines makes people indignant. What the open source world needs is more people who are willing to open the compiler and understand the underlying logic, rather than judges who can only forward lazy people's packets. This is indeed the case. What this incident left us is definitely not just a tea and after-dinner topic. It is a serious wake-up call for the entire community ecology. Okay, that's all for today's discussion. Thank you for listening. Welcome to subscribe to the Sharing and Listening Book Podcast channel. See you next time.

歡迎收聽聽書Podcast頻道由仲裁工程師製作 那個你想像一下這個畫面喔 就是你今天一覺醒來 然後發現你辛苦寫的開源專案 在網路上突然被千夫所指 幾十萬人瘋狂轉發一篇文章 指控你什麼教科書籍洗代碼 說你偷了別人心血 哇這壓力真的超大 對啊那重點是喔 等我們真的去扒開這層 就是這種群情激憤的外衣 潛入底層的原始碼去逐行比對的時候 卻發現你被指控抄襲的核心部分 你根本連血都沒血 嗯連血都沒血 對總共才用了不到600行程式碼 去呼叫第三方工具而已 這就是最近在AI代理 也就是大家常聽到的 AI Agent圈子裡面 鬧得沸沸揚揚的那個 Norris Research Hermes Agent 跟中國團隊Evil Map Evilver引擎之間的抄襲疑運 沒錯這件事最近真的討論度非常高 真的 那這道深度解析呢我們不談情緒 只看證據 好我們來猜解一下這個 這個看似鐵證如山的指控 其實這起事件的發展軌跡喔 非常智得玩味 你看那篇爆紅的指控文章 他的用詞可以說是極度精準地踩中了社群的情緒痛點 怎麼說他用了什麼詞 他裡面直接聲稱說 Hermes的十步組巡丸一模一樣 然後還說什麼術語全換了 但架構文思不動 哇這聽起來真的很像石腿 對嘛 對於一般沒有時間去翻閱原始碼的大眾來說 你看到這種圖文並茂的指控 第一反應絕對是覺得 哇這簡直就是鐵證如山了 確實但這就是最吊詭的地方了 如果那篇文章已經提供了看似這麼完美的對比圖表 為什麼今天我們拿到這份 就是逐行審查三個相關開源代碼庫的深度分析筆記 卻會得出完全相反的結論 這就是最核心的問題 對呀 你要指控別人抄襲最基本的第一步 不就是應該把雙方的程式碼攤在陽光下 然後一行一行並避對嗎 邏輯上確實應該是這樣 但這份深度筆記揭露的第一個 老實說也是最致命的盲點 就在這裡 什麼盲點 就是Evolver被指控招人抄襲的那個核心應勤程式碼 外界根本無法進行獨立驗證 等一下無法驗證 為什麼程式碼不是開源的嗎 因為包含像是Evolve.js啦 或是Selector.js 這些關鍵檔案 全部都是被混淆過 也就是Ophouscated的JavaScript 欸 等一下混淆過 如果今天你是一個聽眾 可能不太了解這是什麼概念 可以解釋一下嗎 簡單來說 就是把人類看得懂的程式嘛 轉換成機器能跑 但人類看了一籌污水的亂碼 報告裡面特別提到 Evolver的Evolve.js 是一個高達728KB的巨型檔案 728KB的文字檔 那很大欸 超大 而且你打開來 第一行是長這樣的 我唸給你聽喔 它寫Constant 0x2b4 E3等於0x1cc9 然後後面跟著幾十萬字的亂碼代號 我的天啊 這就像是 這就像有人控告你偷了他的祖傳秘方 但當法官要求看原版秘方的時候 他卻遞上一個鎖死的保險箱說 欸 法官你不用看裡面 你只要相信我的描述就好 他絕對偷了 對 這個比喻非常精準 既然陳氏馬是鎖死的 那我就不懂了 那文章裡面的什麼十步循環對比 到底是怎麼做出來的啊 這是個非常敏銳的問題 答案是 那篇文章的對比基準 完全是建立在Evolve.js 團隊自己在他們部落格裡面的 文字描述上 蛤 只憑文字描述 沒錯 所以當矩陣方要求社群 信任他們的架構描述 而不是開放原始碼 讓大家檢驗的時候 這場技術探討就已經變質了 變成什麼 變成了一場信仰測試啊 在開源的嚴謹標準下 你絕對不能用黑箱 來證明別人抄襲了你的黑箱 這在矩陣邏輯上是有很大瑕疵了 我懂了 就是因為我們根本看不到A的真實樣貌 所以我們只能拿B 也就是Hermes完全公開的程式碼 來跟A的自我宣稱作比對 對 就是這個意思 那好 這份筆記實際去看了 Hermes的那個EvolveSkill 打PY檔案 它真的有像指控說的那樣 完美複製了那十個步驟嗎? 這裡令人著迷的是 實際攤開來看 不但沒有完美複製 反而是步步不同 步步不同 連一步都對不上嗎? 我們深入看幾個核心步驟就知道了 Evolver聲稱他們的第一步是 Insure Asset Files 這個是做什麼的? 主要是用來初始化空文件的 但是Hermes 實際對應的第一步動作 卻是去尋找並解析一個名為 SQL.MD的特定格式檔案 這聽起來完全不一樣啊 一個是見空檔案 一個是去讀特定的設定檔 沒錯 然後到了第二步 Evolver是做非常複雜的三層信號提取 那Hermes呢? Hermes卻是在構建給模型評估用的資料集 你看 兩邊在做的事情 在工程邏輯上根本就是兩條完全不同的軌道 可是 那文章能說服這麼多人 總得有點相似之處吧 難道大家真的是全網盲從 連看都不看就轉發? 呃 如果硬要從一個極度抽象的角度來看的話了 兩者確實都遵訊了一個大框架 什麼大框架? 就是先加載資料 接著進行優化 然後驗證結果 最後保存下來 停下什麼? 就這樣 加載 優化 驗證 保存 對 就這樣 哇塞 這就像是說 所有的編譯器都抄襲了世界上第一個編譯器 因為他們都經歷了詞法分析 語法分析 然後代碼生成 這只是基本流程吧 對呀 或者是說 你寫文章一定有起承轉合 所以你寫文章就是抄襲了李白 這只是一個任何資料處理管線 也就是pipeline都必須遵循的基本結構吧 沒錯 這就點出了人們在面對複雜戲諷的時候 常犯的一個認知偏誤 大家很容易把通用架構的必然性 誤認為是抄襲的鐵針 嗯 確實很容易被誤導 這就像我們絕對不會指控 一輛新車抄襲了舊車 只因為它們剛好都有四個輪子 和一個方向盤 對 因為車子本來就長那樣嘛 對呀 這是物理與工程的底層邏輯決定的 這不是誰的獨創 好 就算流程這種大框架不能算抄襲了 但我還是想提一個點 那篇爆紅文章裡面 最讓人覺得實錘 感覺最作賊心虛的指控是 他說Hermes做了12對術語的系統性替換 哦 這個部分真的討論得很熱烈 你故意把變數名稱全改了 我必須承認 這招在視覺上 就是那個對比表格列出來 非常非常有說服力 這確實是整篇完章中最具煽動性的一環啦 就是所謂的AI洗代碼教科書 但當我們把這12對術語放回他們原本的程式碼脈絡裡面 去理解他們實際上到底在做什麼的時候 發生什麼事 這個指控就如同紙牌屋一樣瞬間崩塌了 哇 這麼誇張 那我們直接拿最常被討論的幾個例子來開刀好了 文章指控Evolver裡面的Gene 也就是G-in 被Hermes替換成了SKL.MD 嗯 對 這添起來很像是為了避險 故意把一個抽象概念 換成一個實體的擋爛名稱啊 但如果也真的去查證 你就會發現SKL.MD 根本就不是Hermes 為了掩人耳目而自己發明的詞彙 那它是什麼 這個詞其實是Anthropic 在2023年12月發布的Agent Skills開放標準 時至今日 業界已經有超過32個以上的工具採用了這個標準 包含哪些工具 包含大家每天都在用的Cloud Code 還有Cursor這些主流工具 欸 所以這根本不是什麼替換術語 而是Hermes選擇支援了業界最新的通用標準 完全正確 這也太扯了吧 這就像我今天開發一個網站 我決定使用HTML5的標準標籤 結果你跑來告我說 我是為了掩蓋抄襲你的城市嘛 才故意把名字改掉 這邏輯太跳躍了啦 還有更牽強的喔 文章裡面說 Evolver用來追蹤成功狀態 還有完整記錄的那個Capsule 膠囊那個詞對吧 對膠囊 文章說它被替換成了Hermes裡面的Execution Record 執行記錄聽起來功能蠻像的啊 聽起來很像對不對 但問題是如果你去翻邊Hermes那個Self Evolution程式碼倉庫 你根本連Execution Record這個詞都找不到 等一下 連這個字眼都不存在 對 根本沒有這個變數或類別名稱 那Hermes到底是拿什麼東西在做記錄 總不可能不記錄吧 Hermes實際使用的是一個名為Fitness Score的評分資料結構 體式能分數 哈哈 對 字面上是這樣 但它實際上就只是一個非常單純的打分機制 這跟Evolver那個龐大而且複雜的Capsule記錄器 在功能和實作層面上完全是不對等的東西 所以為了湊齊這12對 指控者甚至自己腦補了一些詞彙 沒錯 他們進行了非常嚴重的跨系統移花節目 怎麼個移花節目法 你再舉個例子 例如說 指控中提到Evolver有一個很複雜的Solidify發布流水線 說它被換成了Hermes的Skill Manage Create 這兩種聽起來完全不一樣啊 不僅聽起來不一樣 實際作用也差了十萬八千里 問題是Skill Manage Create只是Hermes主倉庫裡面一個最基礎的CRUD小工具 這是新增讀取更新刪除那種最基本的資料庫操作工具 對 就只是一個小腳本而已 它甚至根本不在被指控抄襲的自禁化倉庫裡面 但最荒謬的還不是這個 還有更荒謬的 最荒謬的是 他們把Evolver的Selector等同於Hermes的Skill Listing函數 等一下 我記得筆記裡有寫到這一段 Evolver的Selector聽起來像是一個什麼加權選擇引擎之類的 用來計算重疊度 對吧 那座聽起來非常高大上 那Hermes的Skill Listing實際的程式碼到底在幹嘛 它的功能就只是去讀取資料夾 把裡面的SKLMD檔名列成一個清單 僅此而已 蛤 就把目錄下的檔名印出來 把一個讀取檔案目錄的列表動作 指控為抄襲了你的加權選擇引擎 對啊 就是這麼牽強 這也太扯了吧 難怪筆記裡面會有那個神比喻 就是說 這就像是指控React抄襲了View 因為React的component對應View的component 然後React的state對應View的Data 這個比喻真的很傳神 這根本就是偷換概念嘛 強行把兩個不相干的東西綁在一起 就只為了湊出一個漂亮的術語替換對比表格 這就是為什麼我們一再強調 絕對不能只看那種懶人包的對比圖表 當你真的撥開這些術語去看底層的作用極致 它的How and Why 你就會發現它們根本是兩種完全不同的物種 OK 好 細節完全對不上 術語也是硬湊的 但我還是想扮演一下魔鬼代言人了 好 你說 文章裡面提到了一個聽起來極具技術深度的指控 他說這兩個系統的三層記憶體系達到了精確對應 嗯 三層記憶 對 就是說都有什麼樸實知識 可附用技能 還有可搜尋的歷史 就算前面的細節都是巧合 這種架構層級的高度一致 總不能說完全沒有借鑒吧 如果我們將這與更大的脈絡連結起來看 你就會明白為什麼了 你想像一下 如果你今天是一位正在開發AI Agent的工程師 坐在電腦前面 你面臨的最大硬體與軟體限制是什麼 最大限制 嗯 是大語言模型 也就是LLM的上下文框嗎 沒錯 就是Context Window 模型不可能一次記住所有的專案細節 龐大的程式碼 還有你們過去的每一次對話 他一定會忘記 所以你必須把它們拆開來分類處理 精確地說你是不得不這麼做 在資訊論的層面上 為了解決這個上下文瓶頸 知識最自然的分類法就是這三層 哪三層 第一把你必須永久記住的規則 變成樸實知識 也就是生命性的記憶 第二把可以重複執行的動作 封裝成技能 這是程序性的 然後第三把過去發生過的事情 傳成可搜尋的歷史 這聽起來就像人類的長期記憶 肌肉記憶跟短期記憶一樣啊 非常類似 所以這根本不是哪一個天才團隊的靈光一閃 而是在LLM的架構約束之下 所產生的一種功成必然 也就是engineering necessity 所以這就是你在筆記裡面提到的 那個專有名詞 獨立收斂 independent convergence 因為大家面對著同一堵牆 最後只能凿出形狀很相似的門 沒錯 就是這個概念 事實上 如果你現在去檢視 市場上主流的agent框架 你會發現大家都在做一樣的事 例如說呢 像是land chain 它有實體記憶 工具和對話歷史 orGBT呢 有長期記憶 技能指令和互動歷史 crew AI 也有類似的三層劃分 哇 所以大家都有這三層 對呀 大家都在同一套物理法則下蓋房子 那客廳 臥室 廚房的配置自然會高度相似嘛 把這種行業共通的工程解法 當作抄襲的鐵震 這真的是對技術發展脈絡的言通誤解啊 好 我們一路這樣拆解下來哦 流程是通用的 術語是東拼七湊的 架構則是工程必然 那真正有趣的地方來了 什麼地方 如果Hermes根本就沒有抄襲 Evolver那套複雜的系統 那Hermes自己的城市碼庫裡面 到底裝了什麼東西 它們到底是怎麼實現所謂的自進化引擎的 這個嘛 當你真的去越路Hermes原始碼 你會發現真相其實帶有一點 怎麼說呢 黑色幽默 黑色幽默 為什麼 因為Hermes的自進化引擎 本質上它只是一個 Stanford DSP Y 對呢 他們沒有自己去搬磚頭 搬水泥 他們只是寫了一份規格書 然後交給這個總包商去處理 所以Hermes的核心邏輯到底寫了什麼 其實只有兩行程式嘛 等一下 兩行 對 核心就兩行 第一行 設定去呼叫DSP Y裡面的一個叫做GPEA的優化器 順帶一提哦 這個GPEA是加州大學伯克萊分校和史丹佛大學 在ICOR2026發表的一項頂級學術成果 哇 直接拿鼎會的成果來用 那第二行呢 第二行就是執行編譯 就這麼簡單 兩行核心代碼 那整個Hermes 他們自己手寫的Python程式碼 全部加起來到底有幾行 經過筆記的逐行統計 大約只有600行左右 才600行 對 這600行裡面還包含了基礎的加載器了 打分函數了 還有一些執行腳本 他們甚至有幾個規劃中的階段 目前在目錄裡面還是空的呢 哇塞 那被指控遭抄襲的Evover呢 它的量級有多大 Evover是一個從零開始 一磚一瓦自己用No.js构建出來的 完整GPE協議引擎 也就是基因表達編程引擎 所以很大咯 非常龐大 光是裡面一個負責處理信號提取的 Signals.js檔案 單獨這一個檔案就有660行了 等於人家一個檔案 就抵你整個專案了 對呀 這還沒算上Evover背後那個龐大的突變引擎 還有多語言之源模組呢 這在技術邏輯上完全說不通啊 這等於是指控一個人他用600行呼叫第三方總包商的規格書 去教科書籍洗高了另一個團隊幾萬行 從底層協議自己建立的龐大系統 聽起來很荒謬對吧 這根本是拿遙控飛機的說明書 去說他抄襲了波音747的引擎設計圖嗎 這就是我們為什麼一直強調 只要稍微具備一點工程背景 並且真的願意去閱讀兩邊的城市嘛 這個止顧就會不攻自破 確實 他們在工程深度 實現路徑 甚至使用的語言生態上 一個是Python 一個是Node.js 就完全是不同量級 不同宇宙的產物啊 但這就帶出了一個我百思不得解的問題了 什麼問題 如果Hermes真的完全沒有使用Evolver的東西 為什麼那篇爆紅文章 要用極度憤怒的語氣去指控說 Hermes的七份材料理 零提及零歸屬Evolmap 如果今天我是Evolver的開發者 看到一個跟我在做類似方向的專案 連提都不提我 我可能也會覺得有點不爽吧 在開源界 互相致敬不是一種基本禮貌嗎 這裡要釐清一個關鍵的開源倫理 歸屬於引用的界限到底在哪裡 界限在哪 你想像一下你今天寫了一篇關於PiTouch深度學習的學術論文 你需要在參考文獻裡面去感謝TensorFlow團隊嗎 當然不用啊 雖然大家都在做深度學習 但底層框架完全不同啊 我幹嘛謝它 完全正確 那我們來檢視一下Hermes實際的技術站 它依賴DSPI 這是MIT開源協議 它依賴GEPA優化器和MIPROV2 這些也都有各自的授權 Hermes在他們的專案裡面 對這些實際呼叫的底層工具都有清楚的引用與致敬 所以他們並沒有假裝這600行是自己從頭發明的神機 他們很誠實地交代了他們是站在斯丹佛大學的肩膀上 是的 那他們沒有提及Evolver 是因為在技術史作上 他們完全沒有交集 一點都沒有嗎 一點都沒有 他們沒有使用Evolver的程式碼 沒有呼叫他們的GEP引擎 連信號提取機制都不一樣 僅僅因為大家都在研究AI agent的自進化方向 就要求強制署名歸屬 這在軟體工程界是完全站不住腳的 但那篇報文的作者顯然不這麼想啊 他把這種零歸屬渲染成了對方是做賊心虛 刻意抹除痕跡的鐵證 其實啊 這整起事件中最具諷刺意味的 反而是EvolMap官方自己的態度 哦 官方自己怎麼說 如果你繞過那篇生動性的網紅文章 直接去讀EvolMap團隊發布的官方技術部落格 你會看到一個完全不同的敘事 他們寫了什麼 EvolMap官方的用詞其實非常嚴謹 他們承認Agent Skills標準在時間線上早於他們 也承認Hermes建立組倡庫的時間比他們早了半年 哇這很客觀呢 對啊 他們甚至明確點出 Stanford的GEPA與他們研發的GEP是完全不同的技術路線 等等所以官方團隊自己其實很清楚兩者的技術差異 不僅如此EvolMap官方部落格還白紙黑字寫下了我們剛剛討論的那個概念 他們承認獨立收斂的可能性是存在的 他們自己也提到了獨立收斂 沒錯 並且在文章的最後他們還特別強調 文本不做法律層面的抄襲定性 所以這一切到底意味著什麼這太瘋狂了 官方團隊明明保持的技術人員的客觀與客制 甚至保留了獨立收斂的空間 結果這個資訊在經過社群媒體的發酵 還有那種缺求流量的自媒體轉譯之後 居然變譯成了實錘抄襲教科書籍洗太碼 甚至還生出了什麼Hermes被垂後心虛刪除帳號 這種完全無中生有的謠言 天哪這真的是烏龍一場 這就是為什麼我們必須做這趟深度解析 這不僅僅是為了幫一個600行的專案澄清 更是要點出當前開源社群面臨的一個巨大的潛在危機 你指的是流量文化對技術討論的反噬嗎 沒錯 你想像一下 如果你今天是一個滿懷熱忱的開發者 你發現了一個業界共同的痛點 然後你基於現有的開源工具 比方說DSPY 你拼湊出了一個簡潔優雅的解決方案 就像Hermes團隊做的那樣 對 結果只因為你面對相同的工程限制 得出了一個符合行業標準的相似架構 你就要面臨被網紅文章公審 被指控為城市碼小偷的巨大風險 這絕對會引發難以想像的寒蟬效應啊 如果說只要架構相似 或者只是剛好解決了同一個問題 就要被全網列污的話 那誰還敢開源自己的專案 誰還敢把自己的靈感放到Github上給大家看 對呀 大家乾脆全部閉門造車退回到軟體工業 最封閉的那個黑暗時代好了 老實說 在這個爆料文化盛行為時代 真相的門檻似乎變得越來越高了 當情緒的宣洩 還有點擊率的誘惑 凌駕於主行審查城市碼的嚴謹性之上的時候 我們究竟是在保護開源的智慧財穿權 還是在親手扼殺開源協作最珍貴的信任基礎 沒錯 這是一場因為民願意靜下心來去讀城市碼 只看著聳動的標題就群起激憤 而引發的烏龍事件 開源世界需要的是更多願意打開編譯器 讀懂底層邏輯的人 而不是只會轉發懶人包的審判者 確實是這樣 這起事件留給我們的 絕對不只是一個吃瓜的茶與飯後話題而已 它是一個對整個社群生態的嚴肅警鐘 好啦 那我們今天的探討就到這邊 謝謝收聽 歡迎訂閱分享聽書Podcast頻道 我們下次見