Hermes Agent-Nous Research Self-transformation AI technology and technology

Welcome to collect money and listen to books - the Podcast channel is produced by a vegetable engineer. Well, imagine that the AI assistant you use every day is not just helping you write code or replying to letters. Yes, he is actually staring at you silently. Yes, he is secretly observing every one of your work habits and then directly packages your unique problem-solving logic into a digital file. It is even said that all of this may be preparing for a super large decentralized network. Does it sound like the beginning of some cyberpunk novel? Really, but this is the technological change that is happening in the open source community now. Indeed, we are actually witnessing the transformation of AI from a relatively passive question and answer machine to a digital entity that really has memory and can evolve on its own. So this is the protagonist we will take you to analyze in depth today. This open source AI agent framework developed by NOS Research is called Hermes Agent. Yes, this project has become really popular recently. This time, we combed through Github There are official technical documents on the website and in-depth discussions on many developer forums. In addition, the latest reports of these technology media such as 36kla. It has only been released for less than two months and the number of emerging units on Github has soared past 47,000. It is really exaggerated. What is even more indicative is that the V0.8.0 and V0.9.0 versions they released in April directly triggered a wave of large-scale relocation. Where did the large-scale relocation come from? It is from the benchmark project Open Claw that currently dominates the market. Many developers have decided to switch jobs and use Hermes. You know, in the open source world, you have to let the engineer put away a whole set of workflows he has already developed and switch to a brand new framework. The conversion cost is very high. Yes, this usually means that this new tool really solves a fundamental problem that makes everyone miserable. OK, let's unpack this. This problem that makes everyone miserable is actually AI amnesia. Yes, the whale brain of AI. Most of the Agents on the market today are stateless, that is, as soon as the Stateless task is over, well, his head is cleared instantly. You have to tell him the background again every time and redefine your city code style. It is super annoying and will really make people crash when using it. Hermes claims to have built a set of memories that can span time. How does it do it? In fact, the most fascinating thing here is that What's festivating here is that it does not just rely on a single database. It has established a very three-dimensional three-layer memory system. Three layers, which sounds very complicated. The first layer is cross-session memory. It uses an engine called FTS-5 and then it is paired with the important technology of large-scale language models. Wait a minute, what is FTS-5? FTS-5 is actually an expansion module in the SQ-Lite database specifically designed for extremely fast full-text retrieval. So it is like a Google search engine built directly into the AI's head? Well, it can be understood that way, but it is actually a little smarter than that. How to say? Because if there is only FTS-5, it can only do a rigid version of keyword comparison. But Hermes will throw the historical records retrieved by FTS-5 to LLM for contextual semantic analysis. In other words, if I tell it, it will also help me handle the database error reported last Tuesday. Yes, it can accurately correlate to the context of your conversation last Tuesday. It will know which database you are talking about and which link you were stuck at at that time. Wow, this really completely solves the pain point. What about the second and third layers? The second layer is called Skill Documents. Skill Documents will directly convert your previous successful experiences into Markdown files and then store them in the local folder of your computer. It is quite safe to store them locally. Yes, the third layer is also life-enhancing. It is called Honcho Dialectical User Modeling. The name sounds very hard-core. Simply put, it will dynamically depict who you are every time it interacts with you. Do you want to study me? Yes, it will secretly record whether you have a high tolerance for errors, or whether you prefer to write the architecture first, or whether you like to revise as you write. It will even record your tone preference when communicating. Wow, so it is not only solving the problem at hand, it is also studying me. This reminds me of a metaphor. Traditional AI is like a super intern with only a goldfish level of memory. It has to be retaught every time. Yes, every time you hand over a task, you have to smash the thick employee handbook to it. But what about Hermes? It is more like a new colleague who will write down his successful experiences as SOP notes after completing the task. This metaphor is quite accurate. Next time he encounters a similar problem, he can just read his own notes. But, wait a minute, how does it turn the bunch of chaotic conversations in our terminal into a structured SOP? It can't be retraining its underlying Shenjun Network. Oh, you pointed out a very key technical detail. It is definitely not modifying the weight of the large model. The computing power cost is terrible. Our ordinary computer can't run at all, so how does it do it? In the Hermes system, this is set up as a five-step learning cycle, that is, observation, extraction, skill building, refinement, and finally push. What exactly does this look like at the level of program code? Imagine it. You may have just spent about half an hour asking AI to help you solve a very tricky shooting bug. Usually the terminal screen has long been washed out by various error reports. Yes, it is full of various failed attempts and invalid instructions. This is actually the first step. That is to observe, and then enter the second step of extraction. Hermes will secretly start a relatively small summary model in the background, such as MiniMax M2.7. It will review the entire chaotic Log Ridge you just went through. Oh, so it is equivalent to sending another small AI to act as a referee to find out which step in the half hour is the key to solving the problem. That's it. It will extract the Python command that finally succeeded and the previous environment settings, and then enter the third step. Write it into a Markdown file that complies with open standards. This file is actually a high-density prompt word, right? That's right. If it encounters the same situation next time, it will read this Markdown and skip the previous half-hour of fuss. Then the subsequent experience and push will be that it optimizes these notes by itself when your computer is idle and turns the notes into a more general long-term memory. Right, solidify it, but this brings up a problem that I think is quite fatal. If it keeps writing its own rules based on seemingly successful experiences, how do we control it? This is why the developer community is now divided into two factions. Everyone is comparing Hermes with the very popular OpenCloud at the beginning of the year. In fact, the two represent completely opposite AI philosophies. This is a duel of evolutionary adaptation. Is the philosophy of absolute control colliding? It can be said that the core concept of OpenCloud is a high degree of certainty. Its capability boundaries are all pre-written by human developers. It is a password and an action. It relies on traditional RAG tools, that is, subtractive enhancement generation. Humans first prepare the knowledge base. When AI encounters a problem, it goes inside to find the answer and then executes it very accurately. It sounds like OpenCloud is working in a highly regulated factory without cities. Every action must fully comply with the specifications. Yes, so it is very suitable for enterprise-level tasks with extremely low fault tolerance. But Hermes takes the emergent route. Emergent means that skills are grown by themselves. Yes, they are born from countless practical experiences. But won't this cause serious overfitting? Overfitting I am in Hugging I saw a lot of senior engineers warning about this on the Face forum. What did they say? For example, in order to solve a network connection error today, I violently turned off the firewall of the entire server world. As a result, the connection was successful. It really happens very often in reality. Then Hermes Cohen found this result. Will he write in his notes that he will turn off the firewall when encountering network problems in the future? If he regards this accidental and even destructive stupid method as truth. In the long run, won't his head betray the noise? Will he eventually go crazy? Isn't stability more important than this kind of romantic autonomy? Well, this actually brings up a very important question. This is the core trade-off we must face when deploying any autonomous agent. Do you want pre-characteristics or adaptability? You can't have your cake and eat it too. The warning above Hogging Face is very accurate. This kind of noise accumulated in long-term memory is indeed the fatal flaw of the current self-evolving system, as long as it relies on experience learning. There must be a risk of mistakes and experiences being magnified. So does Hermes have any way to solve this bottleneck? We can't always let him write wrong SOPs. Their strategy is actually quite clever. Hermes system has a built-in command called Hermes Claw Migrate Migrate. The meaning of moving is yes. He also said that he does not want to eliminate OpenClaw. He can even painlessly absorb all OpenClaw configuration files with one click. Wow, if you can't beat it, absorb it. That's right. So in the actual engineering structure Now many developers have explored a dual-engine complementary model. How the dual-engine works is to let OpenClaw be responsible for the bottom layer, the kind of precise execution that requires strict permission control, and then let Hermes sit on the upper layer and be responsible for higher-dimensional strategic planning to respond to the vague new needs. Oh, I understand. In other words, use OpenClaw as the chassis and brake, and then use Hermes. Mes is used as a brain navigation system, and the risk of trial and error is controlled in the box. Yes, this architecture is really imaginative, and it's really cool. But when it comes to deployment, this is where it gets really interesting. Here's where it gets really interesting. It breaks the physical limitations we have on AI. How do I say that with traditional AI assistants like the Chat GPT or Github Co-Pilot we are using, you usually have to open a special web page or is it bound to your script? That's right. If you leave it, you can't use it. Yes, the AI falls asleep as soon as my webpage is closed. But Hermes is different. Because traditional AI essentially requires you to connect to other people's servers. But Hermes supports six different mid-end backend deployments. Six types. Yes, it can not only run Docker boxes on your computer, but it even supports Serverless environments like Modo or Datana, that is, serverless environments. What does this mean for Ichiban Tingzou? It means that the Agent is waiting for your instructions. When it is not doing any calculations, it is in a hibernating state, so you don’t have to spend money. There is almost no need to pay for the idle time of the server. But as soon as you issue a command, it will wake up instantly. Wow, this serverless design has really completely changed our personal cost structure for deploying powerful AI. Not only that, it also has a built-in multi-platform gateway. You can use only one command to seamlessly connect it to your Telegram, Discord, Slack, and even 15 communication platforms such as WhatsApp. This is the most sci-fi part. In this way, it is no longer just a gadget that is locked in my computer screen. It directly becomes a big butler in the cloud. Yes, you can call it at any time. I can walk on the road and send a voice to it using Telegram. Help me check the LOG record of the remote server last night. Fix the error. By the way, send me a briefing. Then it will wake up in the remote serverless environment and silently help you get all this done. Oh my god, but what about the cost? In practice, it would be crazy to have an autonomous agent continuously observe and extract in the background and call the language model. It would be crazy to burn API tokens. I may not be able to pay the bill. Oh, this is where the News Research team is very smart. They provide a very strategic function. What function? Hermes supports direct integration of your existing Cloud Pro or Cloud Max subscriptions through the OAuth mechanism. Wait, you mean that it can directly use my monthly subscription quota to run. That’s right. It is equivalent to letting the Agent directly apply the fixed amount you have paid, which perfectly bypasses the super expensive pay-per-use developer API model. Wow, this is definitely a huge game changer for heavy users. It saves everyone a lot of money. But oh, when we put this extremely smart and ubiquitous AI that can directly control the terminal directly to the cloud, the security mechanism will become very scary. I will indeed be a little worried if he wants to fix a small bug today. Suddenly decided to issue an order to delete my entire database. What should I do? Don’t worry too much about this. Hermes has introduced a smart authorization mechanism similar to Codex. Smart Approvals means humans must intervene, right, that is, humans in the loop. When Hermes’ scolding actions include destructive instructions, such as changing system permissions, deleting blocks, etc., yes, the system will be forced to pause, and then it will ask you for the order through Telegram or Slack that you have bound. Is it okay to say that I am ready to execute this command? He will only continue to run when you click Agree. What if he has gone crazy and the entire logic collapses? You can type Stop in the chat room at any time and he will forcibly cut off all processes. Okay, this sounds much safer. What about privacy? He keeps extracting my logs into notes, which must accidentally contain my API medicine or customer confidential information. They have also thought of this. He has built-in mandatory PII, which is a coding mechanism for personally identifiable information. There is also confidentiality masking for gold medicine, so it is processed before uploading to the big model. Right, before sending the logs to LOM, he will replace all these sensitive strings on your local side. Even if he needs to crawl data on the Internet, he also supports anti-detection browsers like Camel Fox to ensure that he will not be easily tracked when he automatically surfs the Internet. Wow, the protection is very thorough. So what does this all mean? We just talked about how powerful its technology is, how flexible its deployment is, and it has saved us a lot of money. But to build such a perfect open connection technology machine, are they really doing charity? This is very deep. Right. When we dig into the background of this development company NOS Research, things become very intriguing because this company has a very strong Web3, which is the gene of decentralized blockchain. Yes, this is indeed another dimension that cannot be ignored in understanding the Hermes project. Their boss is also a person in the currency circle. Their CEO Jeffrey Quesnell He was previously the chief engineer of the Ethereum infrastructure Eden Network, and they recently received a financing of up to US$70 million. US$70 million? There is another important point. This financing is not a new equity financing like traditional Silicon Valley. It is priced using tokens. No wonder this immediately triggered a huge controversy in the community. I see that an unofficial token called NOS has appeared in the cryptocurrency market. And countless people have begun to expect in the community that the official will issue short-seller rewards in the future, which is to record short positions. What’s more important is that when you spread out the bottom layer of Hermes, look at it. Has it been encouraging you to hang up and run locally for a long time and continue to accumulate your success track in solving problems? Yes, and there is a name in the system that we didn't mention just now. It says that it supports exporting the training data of Atropos RL. Can you explain to our audience what this is? No problem. This RL stands for Reinforcement Learning. So Atropos is actually a framework used to collect human preference data. Collect human preferences? In other words, Hermes does not just write down your habits as notes for himself to read. It can also package your programming habits and trial-and-error logic into a standard data format that can be directly used to train larger AI models in the future. Well...this is where I think the most impressive thing is. Is this the perfect fit with Web3's logic of contributing data to prove the workload? Proof of Work does have this flavor. As an ordinary engineer who just wants to use AI to improve efficiency, I am very confused now. Should I treat it as an open source tool that single-handedly improves productivity? Or should I realize that when I am happily enjoying this free butler over there, you have actually become a miner working for others, right? Have I unknowingly become a free worker in a huge network, helping to produce high-quality training data? If we connect this to the bigger picture, we actually have to calm down and separate the practical value of technology from the operating logic of capital. Will this help Glass? Aren’t the two tied together? In this open source system, it can actually be separated. For developers, Hermes is undoubtedly the most beautiful architecture and the most thorough agent system that implements closed-loop learning. Well, technically it is really strong. And because its code is open source, this means that you can completely cut off its external network and only run it in your local sand river. You can not export any Atropos RL data or pass it to anyone. The mechanism of the tool itself is reasonable. But what about the crazy hype of those webster capitals? Or those farming articles on the Internet that teach you how to hang up Hermes to earn rewards. You must be extremely cautious, just don't get cut. That's right, because the official has not clearly announced any reward mechanism. When we play with new technology, the focus should be on whether the memory mechanism of this set of tools can bring substantial negative effects to your personal work. This is the value that you can really hold in your hands. This is really very important. We should arm ourselves with tools instead of turning ourselves into data sources in capital stories. Okay. Let’s summarize today’s in-depth analysis. We are actually experiencing a paradigm shift in AI capabilities. Yes, in the past, we all paid monthly to OpenAI or Anthoropic to rent a standardized smart brain. But Hermes Agent showed us one thing. AI is transforming into a personal digital asset that we can truly own and keep in our own hard drives to accompany us. He has redefined the dimensions by which we evaluate AI. The focus is no longer on how smart it is on the first day it is unboxed, but after it has been fighting side by side with you for half a year and a year. How well can he understand you? So at the end of the show, I want to leave you with a thought-provoking question. Imagine that in the next year, through constant observation and extraction, your Hermes Agent will perfectly copy your circuit of thinking about problems. Your unique naming habit when writing about cities. He may even know better than you yourself about the logical blind spots you are most likely to make when working. Really, by that time, he is no longer just a tool, but a highly customized version of you. Then the problem comes when one day you decide to buy a new computer. Or maybe there is a new and awesome new framework on the market. When you are ready to back up and transfer, the folder of DOT Hermes Skills is right. The folder loaded with all your behavior patterns this year. This file may only be a few MB in size. Is it still a pure software asset? Or has it actually become an extension of your personal mind in the digital world? When we finally cured the amnesia of AI, did we also inadvertently create a digital clone with fragments of our souls? I’ll leave this question to you to explore slowly by yourself. Thank you for listening. Welcome to subscribe to the Share Listening Book Podcast channel. See you next time.

歡迎收金, 聽書-Podcast 頻道由種菜工程師製作 呃,想像一下哦, 你每天在用的那個 AI 助理 他不只是在單純的幫你寫 code 或者回信而已 沒錯, 他其實一直在默默盯著你看 對, 他在暗中觀察你的每一個工作習慣 然後把你獨特的那個解題邏輯 直接搭包成一份數位檔案 甚至說這一切可能還是在為一個 超龐大的去中心化網路做準備 聽起來超像什麼賽博胖克小說的開頭嗎? 真的, 但這就是現在開源社群裡面 正在真實上演的科技變革哦 確實, 我們現在其實正在見證 AI 從一個比較被動的問答機器 轉變成一個真的有記憶, 還會自己進化的數位實體了 所以這就是我們今天要帶你深度解析的主角啦 由NOS Research 開發的這個開源 AI 代理框架 叫做 Hermes Agent 沒錯, 這個專案最近真的超紅 我們這一次呢, 梳理了 Github 上的官方技術文件 還有很多開發者論壇的深度探討 加上像 36 克啦, 這些科技媒體的最新報導 它才發布不到兩個月耶 Github 上的新興數就狂飙突破四萬七千顆了 真的非常誇張 而且更具指標意義的是 他們在四月發布的 V0.8.0 和 V0.9.0 版本裡面 直接引發了一波大搬家方潮 大搬家是從哪裡搬過來? 就是從目前稱霸市場的那個標杆專案 Open Claw 很多開發者都決定跳槽過來用 Hermes 哇, 要知道在開源世界裡喔 你要讓工程師放去他已經弄好的一整套工作流 轉去用一個全新的框架 那個轉換成本是很高的耶 沒錯, 這通常就代表這個新工具 它真的解決了一個讓大家痛不欲生的根本性問題 好, 那個我們來好好拆解一下 OK, Let's unpack this 這個讓大家痛不欲生的問題其實就是 AI 失憶症嘛 對, AI 的鯨魚腦 現在市面上的 Agent 大多數都是無狀態的 也就是 Stateless 任務一結束, 好, 他的腦袋瞬間清空 你每次都要重新跟他講背景 重新定義你的城市碼風格 超級煩的 真的會讓人用到崩潰 Hermes 號稱自己建立了一套 可以跨越時間的記憶體 它到底是怎麼做到的 其實這裡最迷人的地方在於 What's festivating here is 它不是只靠單一的資料庫而已 它是建立了一個非常立體的三層記憶系統 三層, 聽起來很複雜 第一層就是跨對話的記憶 Cross Session 它用了一個叫做 FTS-5 的引擎 然後搭配大型語言模型的栽要技術 等一下, FTS-5 是什麼? FTS-5 其實就是 SQ-Lite 資料庫裡面 一個專門用來做 極速全文檢索的擴充模組 所以它就像是在 AI 的腦袋裡面 直接內建了一個 Google 搜尋引擎嗎? 呃, 可以這麼理解 但其實比那個還要聰明一點 怎麼說? 因為如果只有 FTS-5 的話 它就只能做死版的關鍵字比對嘛 但是 Hermes 會把 FTS-5 撈出來的歷史記錄 丟給 LLM 去做上下文的語意分析 哦, 也就是說 如果我跟它講 也幫我處理上週二的那個資料庫報錯 對,它就能精準關聯到你上週二的對話脈絡 它會自己知道你說的是哪一個資料庫 還有當時到底卡在哪個環節 哇塞, 這真的完全解決了痛點耶 那第二層跟第三層呢? 第二層叫做技能文件, Skill Documents 它會把你之前成功的經驗 直接實體化成 Markdown 檔案 然後就存在你電腦本地端的資料夾裡面 存在本地端, 這蠻有安全感的 沒錯, 那第三層也是裡面聚隔命性的 叫做 Honcho 辯證式用戶建模 Honcho 建模, 這名字聽起來就很硬核 簡單來說 它會在每一次跟你互動的時候 動態去描繪你是誰 研究我這個人嗎? 對, 它會偷偷記錄你對錯誤的容忍度高不高 或者你喜歡先把架構寫好 還是喜歡邊寫邊修 甚至連你溝通時候的語氣偏好 它都會記下來 哇, 所以它不只在解決眼前的問題 它還在研究我耶 這讓我想到一個比喻 傳統的 AI 就像是一個記憶力 只有金魚等級的超級實習生 每次都要重新教 對呀, 每次交辦任務 你都要把那本厚厚的員工手冊 砸給它看 但是 Hermes呢 它比較像是一個做完任務之後 會自己把成功經驗寫成SOP筆記的新同事 這個比喻蠻精準的 下次遇到類似問題 它直接翻自己的筆記就好了嘛 不過, 等一下喔 它到底是怎麼把我們 終端機裡面那一堆混亂的對話 變成一份結構化的SOP啊 它總不可能是在重新訓練 它底層的神軍網路吧 喔, 你點出了一個非常關鍵的技術細節 它絕對不是在修改大模型的權重 那個算力成本太可怕了 我們一般電腦根本跑不動 那它是怎麼辦到的 在 Hermes系統裡面 這被定立為一個五步學習壁環 也就是觀察、提取、建立技能、精煉 最後是推送 這個在程式碼的層面 具體到底長怎樣 你想像一下喔 你剛剛可能花了大概半小時 讓AI幫你繳一個超刁串的拍艙bug 那通常終端機畫面 早就被各種報錯洗版了 沒錯 充滿了各種失敗的嘗試 還有無效的指令 這其實就是第一步 也就是觀察 接著進入第二步 提取 Hermes在背景會偷偷啟動一個 比較小的專門做摘要的模型 像是哪一種模型 像是MiniMax M2.7這一種 它會去回顧你剛剛在整段混亂的Log Ridge 喔 所以它等於是派另一個小AI來當裁判 去找出剛剛那半小時裡面 到底哪一步才是真正解決問題的關鍵 就是這樣 它會把最終成功的那個Python指令 還有前面的環境設定給提取出來 然後進入第三步 把它寫成符合開放標準的Markdown文件 這份文件其實就是一個高密度的提示詞 對吧 沒錯 下次遇到同樣狀況 它去讀這份Markdown 就可以直接跳過前面那半小時的瞎忙 然後後面的經驗跟推送 就是它趁你電腦閒置的時候 自己去優化這些筆記 把筆記變成更通用的長期記憶 對 固化下來 可是這就帶出了一個我覺得蠻致命的問題 如果它一直根據看似成功的經驗來自己寫規則 我們到底要怎麼控制它 這也是為什麼現在開發者社群會炒成兩派 大家都在拿Hermes跟年初那個很紅的OpenCloud做比較 其實這兩者代表的就是完全相反的AI哲學 這是一場演化適應對決 絕對控制的哲學碰脹嗎 可以這麼說 OpenCloud的核心理念是高度的確定性 它的能力邊界全都是人類開發者預先寫好的 就是一個口令一個動作 對 它依賴傳統的RAG機具 也就是減索增強生成 人類先把知識庫準備好 AI遇到問題就去裡面找答案 然後非常精準地執行 聽起來 OpenCloud就像是在一個高度管制的工廠無城市裡面工作 每一個動作都必須完全符合規範 沒錯 所以它非常適合那種容錯率極低的企業級任務 但是Hermes走的是湧現路線 Emergent 就是技能是自己長出來的 對 是從無數次的實戰經驗裡面自己生出來的 可是 這難道不會產生很嚴重的過度擬河嗎 Overfitting 我在Hugging Face的論壇上看到蠻多資深工程師在警告這件事 他們怎麼說 比如說了 假設我今天為了解決一個網路連線爆錯 我就直接暴力地把整台伺服界的防火牆給關了 結果 連線居然成功了 真的現實中很常發生 對呀 那Hermes Cohen查到這個結果 他會不會就在筆記裡面寫說 以後遇到網路問題就關閉防火牆 如果他把這種很偶然的甚至有種破壞性的笨方法 當成了真理 長期下來他的腦袋不就出賣了雜訊嗎 最後會不會整個走火入魔啊 穩定性難道不比這種浪漫的自主性還要重要嗎 這個嘛 其實帶出了一個很重要的問題 這就是我們在部署任何自主agent的時候都必須面對的核心取捨 你到底是要預特性還是要適應性 魚與熊掌不能兼得 對 Hogging Face上面的警告是非常精準的 這種長期記憶累積下來的雜訊 確實是目前這種自進化系統的致命傷 只要是依賴經驗學習 就一定要有錯誤經驗被放大的風險 那Hermes有什麼方法可以解決這個瓶頸嗎 總不能一直讓他寫出錯誤的SOP吧 他們的策略其實蠻妙的 Hermes系統裡面內建了一個指令 叫做Hermes Claw Migrate Migrate 搬家的意思 對 再表示他並沒有想要消滅OpenClaw 他甚至可以一鍵無痛吸收OpenClaw所有的設定檔 哇 打不過就吸收它 沒錯 所以在實際的工程架構上 現在很多開發者摸索出了一種雙引擎互補的模式 雙引擎是怎麼運作的 就是讓OpenClaw負責底層的 那種需要嚴格權限控制的精確執行 然後讓Hermes坐在上層 負責比較高維度的策略規劃 去應訴那種模糊不清的新需求 哦 我懂了 也就是說 用OpenClaw當作底盤和煞車 然後用Hermes當作大腦導航系統 把試錯的風險控制在煞盒裡面 沒錯 這架構真的非常有想像力 真的蠻屌的 不過說到部署這件事喔 接下來這個才是真正有趣的地方 Here's where it gets really interesting 它打破了我們對AI的那種物理限制 怎麼說 傳統的AI助理 像我們在用的Chat GPT 或者是Github Co-Pilot 你通常都得打開一個特并的網頁 或者是綁定在你的寫扣碗底裡面嗎 沒錯 離開了就不能用了 對啊 我網頁一關 AI 就睡著了 可是Hermes不一樣 對 因為傳統AI本質上是你連線去存取別人的伺服器嘛 但Hermes支援了六種不同的中端後端部署 六種喔 對 它不只可以跑在你電腦裡的Docker撒盒 它甚至支援像Modo或者是Datana這種Serverless 也就是無伺服器的環境 這對一番聽奏代表什麼意思 這代表說當這個Agent在等你的指令 沒有再做任何運算的時候 它是處於冬眠狀態的 所以不用花錢 幾乎不需要為伺服器的閒置時間付費 但只要你一發出指令 它就會瞬間被喚醒 哇 這種Serverless設計 真的徹底改變了我們個人又部署強大AI的成本結構 不只這樣喔 它還內建了多平台闸道 你可以只用一個指令 就把它無縫接上你的Telegram、Discord、Slack 甚至是WhatsApp等等15個通訊平台 這個才是最科幻的地方啦 它這樣就不只是一個被關在我電腦螢幕裡面的小工具了 它直接變成一個雲端大管家耶 沒錯 你隨時都可以呼叫它 我可以走在馬路上 然後用Telegram傳語音給它說 那個幫我檢查一下遠端伺服器昨晚的LOG紀錄 把報錯的地方修好 順便傳一份簡報給我 然後它就會在遠端的無伺服器環境裡面醒過來 默默幫你把這一切都搞定 天啊 不過成本的問題呢 實戰上要讓一個自主agent在背景不斷地觀察提取 還要一直調用語言模型 這API token燒起來可是很瘋狂的耶 那個帳單我可能付不起吧 哦 這就是News Research團隊很聰明的地方了 他們提供了一個非常具戰略意義的功能 什麼功能? Hermes支援透過OAuth機制 直接整合你原本就有的Cloud Pro或者是Cloud Max訂閱 等等 你的意思是 它可以直接用我每個月包月的額度去跑 沒錯 等於是讓Agent直接套用你已經付費的固定額度 完美繞過了那種按次即費超級貴的開發者API模式 哇塞 這對重度使用者來說 絕對是個超級大的遊戲規則改變者耶 幫大家省了超多錢 可是哦 當我們把這個極度聰明無處不在 而且還能直接操控終端機的AI 直接放養到雲端上面 安全機制就變得很恐怖了耶 確實會有點擔心啊 如果他今天為了結一個小bug 突然決定下指令把我的整個資料庫給刪了 那我怎麼辦? 關於這一點不用太擔心 Hermes導入了類似Codex那種風格的智慧授權機制 Smart Approvals 就是人類必須要介入對吧 對 也就是human in the loop 人類在回圈內 當Hermes歸罵的動作裡面包含具備破壞性的指令的時候 像是改系統權限啦 刪除擋難這種 沒錯 系統就會強制暫停 然後他會透過你綁定的Telegram或Slack 傳順序問你 說我準備執行這個指令了可以嗎? 只有當你按下同意他才會繼續跑 那如果發現他已經瘋掉了 整個邏輯崩潰呢? 你隨時可以在聊天室輸入Stop 他就會強制切斷所有的進程 好 這樣聽起來安全多了 那隱私方面呢 他不斷地把我的日誌提取成筆記 裡面一定會不小心混進我的API金藥 或是客戶的機密資料吧 他們也有想到這一點 他內建了強制的PII 也就是個人身份資訊的打碼機制 還有針對金藥的機密遮蔽 所以在上傳給大模型之前就處理掉了 對 他在把日誌送到LOM之前 就會先在你的本地端把這些敏感字串全部替換掉 甚至如果他需要去網路上爬資料 他還支援像Camel Fox這種反偵測瀏覽器 確保他自動上網的時候不會被輕易追蹤 哇 防護做得很徹底耶 所以說這一切到底意味著什麼 So what does this all mean? 我們剛剛聊了它技術有多強 部署有多靈活 還幫我們省了一大筆錢 可是建立一個這麼完美的開聯技術機器 他們真的是在做慈善事業嗎? 這個就水很深了 對呀 當我們去深挖這家開發公司 NOS Research的背景的時候 事情就變得很耐人尋味了 因為這家公司有著非常濃厚的Web3 也就是去中心化區塊鏈的基因 沒錯 這確實是理解Hermes這個專案 絕對不能忽略的另一個維度喔 他們的老闆也是币圈的人嘛 他們的CEO Jeffrey Quesnell 之前就是以太坊基礎設施Eden Network的首席工程師 而且他們近期還拿到了一筆高達7000萬美元的融資 7000萬美元? 還有一個重點喔 這筆融資不是傳統Silicon Valley那種新創的股權融資 它是用代幣也就是Token來定價的 喔 難怪這立刻就在社群裡面引爆了超級大的爭議 我看到現在加密貨幣市場上 已經出現了一個名叫NOS的非官方代幣類 對 而且無數人開始在社群裡期待 官方未來會發放空頭獎勵 就是去錄空頭啦 更關鍵的是喔 當你把Hermes的底層落機攤開來看 它是不是一直在鼓勵你長時間在本地端掛機運行 然後不斷累積你解決問題的成功軌跡 沒錯 而且系統裡面還有一個我們剛剛其實沒講到的名字 它說它支援匯出Atropos RL的訓練數據 你可以幫我們聽眾解釋一下這是什麼東西嗎? 沒問題 這個RL代表的就是強化學習 Reinforcement Learning 那Atropos其實就是一個用來收集人類偏好數據的框架 收集人類偏好? 也就是說Hermes不只是把你的習慣寫成筆記給他自己看而已 他還可以將你寫程式的習慣你試錯的邏輯 打包成一種可以直接用來訓練未來更大型AI模型的標準數據格式 呃...這才是我覺得最背脊發亮的地方欸 這是不是完美契合了Web3那種貢獻數據來證明工作量的邏輯啊 Proof of Work 確實有這個味道 身為一個只是想用AI來提高效率的普通工程師 我現在就很疑惑啦 我到底應該把它當成一個單權提升生產力的開源工具 還是我應該要意識到當我在那邊很開心享受這個免費管家的時候 其實你已經變成幫人家打工的礦工了 對啊 我是不是不知不覺變成一個龐大網路裡面 在幫忙產出高品質訓練數據的免費勞工 這個嘛 如果我們把這根更大的宏觀格局連起來看的話 If we connect this to the bigger picture 我們其實必須冷靜一點把技術的實用價值 跟資本的運作邏輯給剝離開來看 這有幫到玻璃嗎 這兩者不就是綁在一起的嗎 在這套開源系統裡面其實是可以分開的哦 對開發者而言 Hermes無疑是目前架構最漂亮 把學習閉環實作的最徹底的代理系統 嗯 技術上它是真的很強 對 而且因為它的程式碼是開源的 這代表說 你可以完全切斷它的外部網路 只在你本地的沙河裡面跑 你大可以不把任何 Atropos RL 的數據匯出 也不會傳給任何人 工具的機制本身是中理的 那面對那些webster資本的瘋狂炒作呢 或者網路上教你怎麼掛機 Hermes來賺獎勵的那些農場文 你必須保持極度的謹慎 就是不要被割韭菜就對了 對因為官方其實並沒有明確宣布過任何獎勵機制 我們在玩新科技的時候 焦點應該放在這套工具的記憶機制 到底能不能為你個人的工作 帶來實質的負利效應 這才是你能真正握在手裡的價值 真的這點非常重要 我們應該是用工具來武裝我們自己 而不是讓自己變成資本故事裡面的數據來源 好 我們來總結一下今天的深度解析喔 我們其實正在經歷一場AI能力的典範轉移 沒錯 過去我們都是向OpenAI或Anthoropic 每個月付錢去租借一個標準化的聰明大腦 但是Hermes Agent向我們展示了一件事 AI正在轉變成一種我們可以真正擁有 並且能留在自己硬碟裡面 陪著我們持續累積的個人數位資產 他重新定義了我們評估AI的維度 重點已經不再是他開箱的第一天有多聰明了 而是他陪你並肩作戰了半年一年之後 他能變得多懂你 對所以節目的最後我想留給你一個發人深神的問題 想像一下喔 在接下來的一年裡面 你的Hermes Agent透過不斷的觀察和提取 他完美複製了你思考問題的回路 你寫城市時候獨特的命名習慣 他甚至可能比你自己還要清楚 你工作時候最容易犯的那些邏輯盲點 真的 到了那個時候 他已經不只是一個工具了 而是高度客製化了另外一個你 那麼問題來了 當有一天你決定買一台新電腦 或者是市場上出現了下一個更屌的新框架 你準備要備份轉移的時候 那個DOT Hermes Skills 的資料夾 對 那個裝載的你這一年來所有行為模式的資料夾 這個可能只有幾MB大小的檔案 他到底還算不算是單純的軟體資產 還是說他其實已經變成了你個人心智 在數位世界的延伸了 當我們終於治好了AI的失憶症 我們是不是也無意間創造了一個 擁有我們靈魂碎片的數位分身呢 這個問題就留給你自己去慢慢探索了 謝謝說聽 歡迎訂閱分享聽書Podcast頻道 我們下次見