EP308 – Hermes Agent true battle: travel, change, speed, change, expand 24/7, adjust the number of workers to the top of the road!

Welcome back to share some good news with you. AI Lazy Newspaper also has a YouTube channel. If you want to know more about the topics discussed today, you can directly search AI Lazy Newspaper on YouTube. There are reference links in the information column under the video. Okay, let’s return to today’s topic. There are a bunch of new AI tools popping up every day. Do you often don’t know where to start? Don’t worry, here is AI Lazy Newspaper to help you select the five most discussed AI tool videos every day. Let AI enter your daily life in ten minutes every day. Let you unknowingly become an efficiency master. I think what I am going to talk about today will make you rethink the question of what exactly am I using AI? Because we are no longer just asking AI questions, but directing an automatically operating system. There is a tool that allows you to enter a line of instructions and then make coffee, lure dogs, and stare at the ceiling. When you come back an hour later, not only have you completed the task, but you have also tested it yourself and made your own mistakes. The visual materials have been generated. My first reaction is, wait, do I still have a job? The second reaction was immediately, this is exactly what I have been waiting for. Before talking about the specific tools, I want to help you build a framework first, because if you don’t understand the background, you will hear the following things. You can think of it as four levels of love and interaction. It's purely a suggestion, with no action at all. The next level up is that you string together tools and use platforms like M8N or CAPEER to set up data to automatically flow between different services. I've done this myself. You obviously just don't want to spend 10 minutes a week to manually compile a report. In the end, it took 6 hours to build a super complex automated process. You still feel that you are very smart. It's really fun when you run it, but the problem is that this static process doesn't think at all. As long as a situation occurs that he didn't expect, for example, the data format has changed a little bit. The whole process will be broken directly. He will only follow the steps you realistically followed. You are still the one who defined each step. But what is really scary is that the third layer AI starts to judge what to do next. You no longer define each step. You only define the goal and let the model figure out how to do it. You tell it to help me convert this script into another language and then write tests. He will push it by himself, execute it by himself, observe the results, adjust it over and over, pull the required files by himself, plan the instructions by himself, see the errors, and fix them by himself. You can think of it as the model is the brain. The outer execution framework is the hands and feet. It is responsible for actually opening files and planning instructions. It was changed to 4, but there is also a ceiling on the third level. Usually it is the AI that does one thing. In an independent working stage, if you close the window, it will forget all the things it has just learned. The fourth level is today’s focus. Autonomy with collaborative capabilities is the AI system. It is not just an AI, but a whole collaborative team to help you operate. AI has professional skills at this level, long-term memory across working stages, and continuous context. You are not giving the AI a task. You are managing an asynchronous digital employee. Okay, the framework is built. Let’s talk about a few tools that make this fourth layer really possible. Let’s start with Codex. We mentioned him before. At that time, he could already test web applications by himself and control the virtual mouse to click on the interface. But recently Codex launched an experimental feature called Ghost. This thing completely redefines how we handle long-running tasks. Have you ever tried to let AI build a complex project from scratch? You know the pain. The upper and lower micro-capacity began to be insufficient, and the model began to get mixed up, or it ran out of time. In the past, to make the AI run for several hours, you had to build a complex coordination layer yourself. Write a script, which probably means reading the command, looking at the current status of the project, performing an action, updating the status file, and repeating until the completion conditions are met. This method can be used, but you have to build all the printing racks yourself. What to do if the API rate limit is encountered? What to do if the AI is tuned? The basic loop will die directly. Codex calls will build this entire coordination layer directly into the tool. All you need is a slash command. Let me tell you what the actual process looks like. Suppose you want to build a complete application, like a first-class fighting game. You don't directly say to help me build a game. The first step is to use Codex's planning mode to put together a very specific blueprint. You define the game mechanical material requirements. The most important thing is to define clear verification standards. What your command says is that the definition of success is that the application can be built normally, the local development server can be started, and an automated test script can successfully load the screen. Simulate keyboard operations to trigger damage and then enter the victory state. After the blueprint is confirmed, you give an instruction and tell him to follow the plan. Then you can do other things. Codex runs a sophisticated integration loop in the background, constantly checking the internal state to confirm the progress. But what makes Ghost really smart is how he handles the budget limit. Let an autonomous AI run for an hour. Tokens will burn quickly. If you hit the preset usage limit, the script will usually abort, leaving you with a pile of half-written files. and broken code, but Codex Goals is smarter. At the end of each round, he will evaluate what to do next. If he finds that he is about to hit the limit, he will not block it directly. He will end the current round and produce a final report detailing which parts have been completed. After you authorize more degrees, what to do next. It is like a public student. Instead of just writing half the sentence and leaving when the work time is up, you will leave a detailed handover instruction on your desk. The results of the actual test really surprised me Codex You can run it for 45 minutes to an hour without your intervention at all. Use AI to generate original proof diagram materials, with transparent background layers and enemies generated by fine logic. You can also write tests to verify your results. All done automatically in the background. You can use Code in one window to do your main development work and at the same time in another window, let Codex run a large target, handle large-scale refactoring, or build a basic architecture. The two things can be done at the same time without interfering with each other. But Codex is still essentially a tool that lives on your computer. It is designed for you to sit in front of the keyboard and do in-depth engineering work. What about when you leave the computer? When you need a continuously operating system to monitor your infrastructure, manage your content, or run scheduled automation. You cannot leave the laptop with the window open 24 hours a day just to grab some information at three in the morning. This is why next we will talk about the latest evolution of Harmicization. If you are paying attention to the field of open source AI, you may also be following OpenCode OpenCo. de is very useful, but recently it has started to have a symptom of wanting to shoot everything, pushing large updates very frequently, and almost every update will break the system. If you are like me, what you want is to use your AI directly instead of spending 20 minutes every morning debugging your AI execution environment. Harmic has taken a different path, lighter and faster, and very focused on reliability and self-improvement. To understand why Harmic is suitable for situations where you are not next to the computer, you need to understand its five core things: memory, skills, personality scheduling tasks, and self-improvement loops. Let's talk about memory first. When an AI starts up, it's basically in a state of frustration. Harmic uses two persistent files to solve this problem. One is about you, recording your preferences, the way you work. You don't like long-winded answers, or you prefer a certain programming language. The other is about the environment and projects. You don't need to update these manually. When you interact with Harmic, it will automatically intercept important information in the background and update these files. Then there are skills. Think of skills as program memory. A reusable operating manual tells the AI how to perform a specific task perfectly every time. If you ask Harmic to generate a flow chart, it doesn't guess what to do. It will pull a specific operating manual and follow the steps. Then there is the personality setting. This file is to set the personality of the AI. If you want an AI to reply to social messages, you may want it to be enthusiastic and helpful. If it is your personal assistant and wants to communicate with you on Telegram, you may want it to be direct, concise and even with a sense of humor. But the function of scheduling tasks is This is the key to turning Harmic from a passive chat tool to an active employee. Because Harmic is designed to run on a virtual private server, it is always online. You can open Telegram while waiting in line to buy coffee and send a voice message saying Harmic. It will help me catch the latest AI news at 6 o'clock every morning, filter out the noise, and then send the summary to my slide channel. Harmic will automatically write the skills, schedule the tasks and then execute them. Every morning it will start an independent work phase. After running the task, the results will be sent back to you. Can you let it automatically synchronize every night and back up your program for it? You can have him monitor your YouTube comments and respond to you based on the tone of your brand. You can have him run a security scan on your server every night. This changes your infrastructure from static to dynamic. And finally, the self-improvement cycle. This is the most interesting part of the whole system to me. When you correct Harmic, let's say he used the wrong tone or messed up a script, he doesn't just say sorry, he actually rewrites his skill profile and updates his memory to ensure that he does not make the same mistake again. The more you use it, the more accurate it becomes. It sounds complicated to set up, but it's actually very intuitive. You open a cheap virtual private server on a platform like Hoster and deploy Harmic in a large container. This containerization method is very important for security and scalability. Imagine that you want to use Harmic to manage your personal finance and social media marketing at the same time. You definitely don't want a super big AI to have access to your bank account and your Twitter account at the same time. That is a time bomb waiting to explode. So you open a container for Finance Harmic and another completely independent container for Marketing Harmic. They each have independent profiles, their own memories, and their own tools. Your marketing AI does not need to know the highest-privilege password of your server. Your financial AI does not need to know how to generate dynamic graphics materials. Through these independent strengths, you can use different Telegram robots to communicate with each AI. You have a small board that can be put in your pocket. No matter where you are, you can directly give instructions in natural language. But here we go. I'm going to talk about a real-life issue directly because many people don't really think about it when planning this kind of system. You have Codecode running in-depth development work locally. You have Codex running 45-minute autonomous tasks in the background. You have three different Harmics running scheduled tasks every hour on the server. You know what this sounds like? Let an AI classify a customer service ticket, format a JSON file, or summarize an ordinary news article. This kind of task does not require the strongest model at all. This is the problem that Ocarouter is solving. And I think it may be the most important underlying tool that most people don't know yet. Ocarouter is a model routing platform. The concept is very straightforward. You don't point your program directly to Anthropic OpenAI, you point it to Ocarouter. You change a URL and a key. You don't need to understand the other code at all. When your AI sends a request, Ocarouter will analyze it instantly. It's a bit like automatically comparing prices when calling a car for you. It will call a cheap car for a short distance, and call a high-end car if you are in a hurry or have complicated road conditions. It will check more than 200 different models, the real-time status and pricing of 10 different suppliers, and then automatically send it to the cheapest model that is capable of handling the task. And the delay of the whole process is less than one millisecond. The cost savings are really exaggerated, on average 40% lower than the list price. In some processes, API fees can be cut by 60% to 80%, and the best part is that their business model is zero markup. You pay the original rate disclosed by the supplier. Ocarouter does not take a commission from your usage charges. You will also get audit records of each request, automatic switching of backup sources when the API hangs, and cost tracking for each request. Combine this with an autonomous AI system and think about it. If Harmic runs a scheduled task at three o'clock in the morning, it just reads a text file and extracts three keywords. Ocarouter It will dynamically send this request to a super-fast and super-cheap model. But if your Codex is running a complex target loop and trying to make an error, there will be a very troublesome asynchronous state problem in your application. Ocarouter will know to send this to the strongest model. You get two benefits at the same time. Use the top model when you really need top wisdom. Automatically save money when you don’t need it. You no longer have to manually change the model in the configuration file. Just set it and forget about it. Let it help you take care of your wallet. Okay. Now put all the things I talked about today together. We are no longer just chatting with AI, we are building and managing systems. You have tools like Codex, which can receive a high-level goal and then execute it autonomously for an hour, elegantly managing your own budget and status. You have digital employees like Hermis, who are constantly operating and improving themselves, living on your server and running automation to communicate with you through Telegram. You have smart low-level tools like Ocarouter, working under the hood to ensure that this large wave of autonomous AI requests will not bankrupt you. Enter software development The threshold for workflow automation and scale has never been so low. But the real change is not in the tool itself, but in the way you think. Your job is no longer to manually trigger backups or format data. Your job is to define the target design architecture and coordinate these AIs. You are changing from operators to commanders. And for those of us who have always believed that we can lie down instead of sitting down, this is the future we have been waiting for. If you feel that today's content has gained you something, please help me follow the like track on Apple Podcasts and leave a five-star review. If I have helped you clear up some pitfalls, please leave a five star so that more lazy people can find us. FBIG Feds Soul AI Lazy Newspaper can find me. The recent AI development is still so happy that it scares me. I don’t want to miss it and continue to listen to AI lazy people tomorrow. I am Kang Lalan. We will see you tomorrow. Bye.

歡迎回來跟大家分享一個好消息 AI懶人報也有 YouTube 頻道咯 想更深入了解今天討論的主題都可以直接在 YouTube 搜尋 AI懶人報 在影片底下的資訊欄都有附上參考連結 好,我們回到今天的主題 每天都有一堆新的 AI 工具冒出來是不是常常不知道該從哪裡開始 別擔心,這裡是 AI 懶人報每天幫你精選五個最多人討論的 AI 工具影片 每天十分鐘讓 AI 走進你的日常生活 讓你不知不覺成為效率大師 今天要聊的東西我覺得會讓你重新想一下 我到底在用 AI 做什麼這個問題 因為我們已經不只是在問 AI 問題 而是在指揮一個自動運作的系統 而且有個工具可以讓你打一行指令 然後去泡咖啡、誘狗、放空盯著天花板 一個小時後回來,任務不只做完了 還自己測試過,自己出錯 視覺素材都生好了 我第一個反應是等等這樣我還有工作嗎? 第二個反應馬上就是,這正是我一直在等的東西 在講具體工具之前,我想先幫你建一個框架 因為如果不懂這個背景 後面的東西你會聽得一頭污水 戀愛跟人互動這件事,你可以把它想成四種等級 最初階就是你問一句,他答一句 你打開瀏覽器,把文字貼進去 問他幫你整理或改寫 然後你把答案複製出來自己用 他不知道你在做什麼專案,不記得你昨天問的什麼 純粹是建議,完全沒有行動 再往上一層是你把工具串起來 用M8N或CAPEER之類的平台 設定讓資料在不同服務之間自動流動 我自己也做過這件事 你明明只是不想每周花10分鐘手動整了一份報告 結果花了6個小時,搭了一個超複雜的自動化流程 還覺得自己很聰明 跑起來的時候真的很爽,但問題是這種靜態流程根本不會思考 只要出現一個他沒預期到的狀況 比如說資料格式改了一點點 整個流程就直接壞掉 他只會照你當初寫實的步驟走 你還是那個定義每一步的人 但真正可怕的是第三層 AI開始自己判斷下一步要做什麼 你不再定義每一個步驟 你只定義目標 讓模型自己去想怎麼做 你跟他說幫我把這個腳本轉成另一種語言 然後寫測試 他會自己推力 自己執行,觀察結果,反覆調整 自己去拉需要的檔案自己刨指令 看到錯誤,然後自己修 你可以把它想成模型是大腦 外層執行框架是手腳 負責真的去開檔案,刨指令 改成4,但第三層也有天花板 通常就是AI做一件事 在一個獨立的工作階段裡 你把視窗關掉 他就把剛才學到的東西全部忘光了 第四層才是今天的重點 有協作能力的自主是AI系統 這不只是一個AI 而是一整個協作團隊在幫你運作 AI在這個層次有專業技能 有跨工作階段的長期記憶 有持續的背景脈絡 你不是在給AI一個任務 你是在管理一個非同步的數位員工 好,框架建完了 接下來聊讓這個第四層 真正變得可能的幾個工具 先從Codex說起 我們之前有提過他 當時他已經可以自主測試網頁應用程式 還會自己控制虛擬滑鼠去點選介面 但最近Codex推出了一個實驗性功能 叫做Ghost 這個東西完全重新定義了我們怎麼處理長時間執行的任務 你有沒有試過讓AI從頭開始建一個複雜的專案 那個痛苦你懂的 上下微容量開始不夠用 模型開始搞混 或是跑到一半整個當調 以前要讓AI連續跑好幾個小時 你必須自己搭一個複雜的協調層 寫一個腳本大概是說讀取指令 看現在專案的狀態執行一個動作 更新狀態檔案 然後重複 直到完成條件達成 這個方法可以用 但你要自己搭所有的印架 遇到API速率限制怎麼辦 AI當調怎麼辦 基本的循環就直接死掉了 Codex calls把這整個協調層 直接內建進工具裡 你只需要一個斜線指令 白說說實際的流程長什麼樣子 假設你想建一個完整的應用程式 就像一個第一格鬥遊戲 你不是直接說幫我做一個遊戲 第一步是用Codex的規劃模式 把一份非常具體的藍圖整理出來 你定義遊戲機制素材需求 最重要的是定義明確的驗證標準 你的指令要說的是 成功的定義是 應用程式可以正常建置 本地開發伺服器可以啟動 一個自動化的測試腳本 可以成功載入畫面 模擬鍵盤操作觸發傷害 然後進入勝利狀態 藍圖確認之後 你下一個指令 叫他照著計劃走 然後你就可以去做別的事了 Codex在背景跑著一個 精密的整合循環 不斷對照內部狀態確認進度 但讓Ghost真正聰明的地方 是他怎麼處理預算上限這件事 讓一個自主的AI跑一個小時 Token會燒得很快 如果你碰到預先設定的用量上限 一般情況下腳本就直接中止 留給你一堆寫到一半的檔案 和壞掉的程式碼 但Codex Goals更聰明 每一輪結束的時候 他會評估接下來要怎麼走 如果他發現快碰到上限 他不會直接擋掉 他會把當前這一輪收尾 然後產出一份最終報告 詳細說明哪些部分已經完成了 等你授權更多的度之後 接下來要做什麼 這就像是一個公讀生 不是工作時間到了就直接把句子寫到一半 然後走人 而是在你桌上留一份詳細的交接說明 實際測試的結果真的讓我有點驚訝 Codex 可以完全不需要你介入 跑個45分鐘到一小時 用AI生成原創的點證圖素材 帶透明背景的圖層 且由細邏輯產生敵人 還有寫測試來驗證自己的成果 全部在背景自動完成 你可以在一個視窗用Code 做你的主要開發工作 同時在另一個視窗 讓Codex跑一個大型目標 處理大規模的重構 或是建立一個基本架構 兩件事同時進行 互不干擾 但Codex本質上還是一個 住在你電腦上的工具 設計給你坐在鍵盤前 做深度工程工作的時候有 那當你離開電腦的時候呢 當你需要一個持續運作的系統 去監控你的基礎架構 管理你的內容 或是跑排程自動化的時候呢 你不可能24小時 都讓筆電開著視窗掛著 就為了凌晨三點去抓一些資料 這就是 為什麼接下來要聊Harmicization的最新進化 如果你有在關注 開源AI這個領域 你可能也有在追OpenCode OpenCode很好用 但最近它開始有一個 什麼都想射的症狀 非常頻繁的推大型更新 而且幾乎每次更新都會把系統搞壞 如果你跟我一樣 你想要的是直接用你的AI 而不是每周早上花20分鐘 除錯你的AI執行環境 Harmic走了一條不同的路 更輕更快 而且非常專注在可靠性和自我改善這兩件事上 要理解為什麼Harmic適合不在電腦旁邊的情境 你要了解它五個核心的東西 記憶、技能、個性排程任務 還有自我改善循環 先說記憶 當一個AI啟動的時候 它基本上是失意狀態 Harmic使用兩個持續存在的檔案來解決這個問題 一個是關於你的 記錄你的偏好 你的工作方式 你不喜歡囉嗦的回答 或是你偏好某種程式語言 另一個是關於環境和專案的 你不需要手動更新這些 當你跟Harmic互動的時候 它會自動在背景截取重要資訊 並更新這些檔案 再來是技能 把技能想成是程序記憶 一個可以重複使用的操作手冊 告訴AI怎麼每次都完美的執行某個特定任務 如果你叫Harmic生成一個流程圖 它不是自己猜怎麼做 它會去拉特定的操作手冊 然後照著步驟走 然後是個性設定 這個檔案是在設定AI的個性 如果你要讓一個AI去回覆社群留言 你可能希望它熱情 樂於助人 如果它是你的個人助理 要在Telegram上跟你溝通 你可能希望它直接簡潔 甚至帶點幽默感 但排成任務這個功能 才是讓Harmic從一個被動的聊天工具 變成主動員工的關鍵 因為Harmic被設計成在虛擬 私人伺服器上跑 它是一直在線上的 你可以在排隊買咖啡的時候打開Telegram 傳一段語音說Harmic 每天早上6點幫我抓最新的AI新聞 過濾掉雜訊 然後把摘要發到我的slide頻道 Harmic會自動寫好技能 排好排成任務 然後執行它 每天早上他會啟動一個獨立的工作階段 跑完任務把結果回傳給你 你可以讓他每晚自動同步 給他備份你的程式嗎? 可以讓他監控你的YouTube留言 並根據你的品牌的語氣朝你回覆 可以讓他每晚對你的伺服器 跑一次安全掃描 這把你的基礎架構 從靜態變成動態 最後是自我改善循環 這是整個系統最讓我覺得有意思的部分 當你糾正Harmic的時候 比如說他用了錯誤的語氣 或是搞砸了一個腳本 他不只是說聲抱歉 他會實際重寫自己的技能檔案 更新自己的記憶 確保他不會再犯同樣的錯誤 你用它越多 他就越精準 設定這個聽起來很複雜 但其實很直觀 你在hostinger這類平台上 開一台便宜的虛擬私人伺服器 然後把Harmic部署在一個大格容器裡 這個容器化的方式 對安全性和擴展性來說非常重要 想像你想用Harmic同時管理你的 個人財務和社群媒體行銷 你絕對不會想要一個超級大的 AI同時掌握你銀行帳戶的存取金鑰 和你的推特賬號 那是一個等著爆炸的定時炸彈 所以你開一個容器 給財務Harmic 另外一個完全獨立的容器 給行銷Harmic 他們各自有獨立的設定檔 各自的記憶 各自的工具 你的行銷 AI不需要知道你伺服器的最高權限密碼 你的財務 AI也不需要知道怎麼生成動態圖形素材 透過這些獨立的實力 你可以用不同的Telegram機器人跟每個AI溝通 你等於是有了一個 可以放在口袋裡的小型董事會 不管你在哪裡 都可以用自然語言直接下指令 但說到這裡 我要直接講一個現實問題 因為很多人在規劃這種系統的時候 都沒有認真想過這件事 你有Codecode在本地跑深度開發工作 你有Codex在背景跑45分鐘的自主任務 你有三個不同的Harmic 在伺服器上每小時執行排程任務 你知道這聽起來像什麼嗎 這聽起來像是月底一張 嚇死人的API帳單 如果你在每一個流程裡都應用最頂級的模型 你基本上是在用大炮打蚊子 為什麼你要付最貴的價格 讓一個AI去分類一張客服單 格式化一個JSON檔案 或是摘要一篇普通的新聞文章 這種任務根本不需要最強的模型 這就是Ocarouter在解決的問題 而且我覺得它可能是目前最重要 但最多人還不知道的底層工具 Ocarouter是一個模型路由平台 概念非常直白 你不是直接把你的程式指向 Anthropic OpenAI 你把它指向Ocarouter 你改一個網址換一個金鑰就這樣 其他程式碼完全不用懂 當你的AI發出一個請求 Ocarouter會即時分析它 它有點像幫你叫車時自動比價 短程就叫便宜車 趕時間或路況複雜才叫高級車 它會檢查超過200個不同模型 10個不同供應商的即時狀態和定價 然後自動送到最便宜 同時又有能力處理這個任務的模型 而且這整個過程的延遲不到一毫秒 省下來的費用真的很誇張 平均比定價低40% 在某些流程裡可以把API費用砍掉60%到80% 而且最棒的是他們的商業模式是零加成 你付的是供應商公開的原始費率 Ocarouter 不從你的用量費用裡抽成 你還會得到每個請求的審計記錄 API掛掉時的自動切換備源 以及每個請求的費用追蹤 把這個跟自主式AI系統結合起來想 如果Harmic在凌晨三點行來跑一個排程任務 只是讀一個文字檔 然後結取三個關鍵字 Ocarouter 會動態把這個請求送到一個超快 超便宜的模型 但如果你的Codex正在跑一個複雜的目標循環 試圖出錯你應用程式裡一個很麻煩的 非同步狀態問題 Ocarouter 就知道要把這個送給最強的模型 你同時得到兩個好處 真正需要頂級智慧的時候用頂級模型 不需要的時候自動省錢 你再也不用手動在設定檔裡換模型 設定好就不用管 讓它幫你看好你的錢包 好 現在把今天講的這些東西全部拉在一起看 我們已經不只是在跟AI聊天了 我們在建立和管理系統 你有像Codex這樣的工具 可以接收一個高層次的目標 然後自主執行一個小時 優雅的管理自己的預算和狀態 你有像Hermis這樣持續運作 會自我改善的數位員工 住在你的伺服器上 跑排成自動化透過Telegram跟你溝通 你有像Ocarouter這樣的聰明底層工具 做在這一切底下 確保這一大波自主AI請求不會讓你破產 進入軟體開發 工作流程自動化和規模化的門檻 從來沒有這麼低過 但真正的轉變不在工具本身 而在於你的思維方式 你的工作不再是手動觸發備份或是格式化資料 你的工作是定義目標設計架構協調這些AI 你正在從操作者變成指揮者 而對於我們這種一直相信能躺就不要坐的人來說 這正是我們一直在等的未來 如果你覺得今天的內容讓你有點收穫 那就幫我到Apple Podcast按讚追蹤 留個五星好評吧 如果自己有幫你掃踩一點坑 也拜託留個五星 讓更多懶人找到我們 FBIG Feds Soul AI懶人報就找得到我了 最近的AI發展依然快樂讓我害怕 不想錯過明天再繼續收聽 AI懶人爆囉 我是康懶懶 我們明天見 拜拜