EP38｜把 AI Agent is ready for transportation!

Meow, I’m DaeDae, I’ve been helping you catch up with AI hot-searching cats, and the AI ​​hot-search report helps you identify key points, and recorded them for a week. Are you considering whether to lie down flat or lie down more flatly? No matter what you plan to do today, DaeDae first sorted out the latest AI key points this week. He listened to the 5 most popular AI tool information, finished it in 10 minutes, and took a nap for the rest of the time. In other words, have you ever felt that the development of AI is so fast that it is a bit stressful? Sometimes, instead of researching new tools, I want to sit by the window and bask in the sun. In fact, you don’t have to strain yourself to learn AI. It takes 10 minutes every day. Listening to what practical problems these tools can help you solve is like installing a memory module for your brain. When you really need to use it, you will find that the answer is in your mind. This man’s learning method is the key to sustaining long-distance running. Let’s follow this episode in order, without detours or showing off jargon. Let’s see directly what kind of trouble these AI tools can help you save in real work. Let’s take a look at the first one, Hermes Zajan. Many people’s eyes will light up when they hear the AI agent, but when they actually open it, they don’t know what else to teach him besides chatting, right? The interesting thing about Hermes Zajan is here. He doesn’t just reply you with an answer, but can remember things. Look at the task board, operate the browser, and organize memory notes. As long as you put research, writing, and reporting into Hermes in the morning, he will be able to view competing product websites. Organize it into a file format and easily write the progress into your memory. You can concentrate on making judgments that only humans can make. To put it simply, what Hermes Zajan wants to do is not a chat cat that can only meow twice, but a little housekeeper cat in your digital life that has been following you silently and understanding your habits more and more. The real pain point this set of things solves is not that you lack an answer, but that you lack continuity. Like its target instructions, morning priority reminders, and memory notes, it actually prevents Hermes Zajan from starting from scratch every time. You can set a clearer target first and he will be less likely to run around. He can also open a browser to study other people's products and attack what functions. After sorting it out, hand it over to Coco or Codex to continue. Every morning, if he asks you what is the most important thing today, remember this information. Over time, it will become more and more like your personal assistant. With TillSkill, he can also access your files and local tools across the middle. Use mobile phones, computers, and different working environments. Don't be like the parallel universe of Three. This must be tightly locked into Meuchou Notes. The most suitable use of Hermes Zajan is to treat it as an assistant with memory, planning, and continuous work. It is not just a chatbot. To put it bluntly, what it sells is not an answer. It is that you finally don’t have to re-explain it every day. Okay, I am too lazy. Let’s look at the next tool. The second one is Hermes, but this time it is another very practical deformation board. It is a pity that it is installed on USB, time audio disc or external SSD. It is Hermes Zajan. The pain points are super obvious. You have finally adjusted the AI environment on one computer to be smooth, and then you have to start it all over again when you switch to another computer. The settings are missing, the history is gone, and the API has to be reconnected. I really want to take a nap where I am. This method is to take the entire Hermes working environment directly with you. It can be plugged into Windows, Linux, and MacOS. Chat records, memory, and API settings all follow you. In the case, it is paired with DeepSeek as the model and Telegram as the message portal. So you don't just move files, but take the same AI assistant with you. For example, if you usually use a desktop computer at home, a laptop at work, or a travel computer when you go out, you can plug it in and it can be seamlessly connected. What you did yesterday is that you don’t have to get a new assistant for each computer. This is basically a magic tool. Leave the work to him and you can open cans next to him. And it is not only a pity that he is great, but also solves the trouble of changing machines and the trouble of AI gestures. This set of pity is that Hermes uses XFAT to make different systems readable. It will help you put the Python and Node.js you need when you start it for the first time. After setting up the Hermes file, it will save a lot of trouble later. In the example, it is connected to DeepSeek because it is low-cost and sufficient. When TelegramBall saw the robot on the street, Father saw that the robot could also lock the user ID to prevent others from messing with it. The best thing was that TelegramBall did not respond for a while during the demonstration. As a result, Hermes Adrian checked it himself and found that it was a user ID conflict and then fixed the settings. This feeling of helping himself to troubleshoot is really like the agent everyone imagined. He can even tell whether he is currently running Windows, Vu2 or MacOS. This one tells you what folders and files your machine has. In short, this is not just putting AI into a flash drive, but putting a memory-based storage environment into your pocket for your work partners who move with you. For people who often change computers, this is really like moving away the familiar cat house together. They feel more at ease and their ears stand up. The next third one is definitely today's MVP. This is a complete AI short film production using TrackGPT plus Google Flow. Many creators get stuck as soon as they touch the video, because either they have no inspiration or the characters look different in each shot. The final product looks very scattered. This method straightens the process. First use TrackGPT to think about the story, determine the characters, break down the situation, write prompts, and then hand it over to Google Flow to generate pictures, do actions to supplement the lens feel, and finally pick up the clips. Suppose you want to make a 30-second action short film today. You can first make a main character list, then turn each scene into pictures and short films, and finally throw it into VN Filmore A or ClipChamp to put it together. NOSIE summary is that this is like putting a mini studio on your desk, although it cannot replace your aesthetic. But at least it can make the same character have the same atmosphere, and it won't go wrong in the middle of the shoot. The really strong point is that this is not just a show-off tool, but a whole set of smooth workflow. The first step is to use TrackGPT to come up with several story directions, and determine the style and protagonist setting. Then the most important thing is to make the character stable, otherwise the handsomeness will not look like you, and the audience will easily be fooled. At this time, TrackGPT can help you sort out the appearance, angle, and expression of the character to make a character reference table. Google later. Flow is better at capturing the same person. At HGPT, the script is broken down into paragraphs suitable for short film generation. Google Flow is responsible for turning images into videos, and can add effects such as in-position, push, pull, ring, and flesh to make the picture more cinematic. Finally, use simple editing software to finish. After learning this, I can sleep for an extra hour this afternoon, because the most valuable thing for novices is not that AI can also make videos. So you finally have a less confusing and reusable filming route. To put it bluntly, Let’s start with the process. If the previous one didn’t impress you, let’s take a look at this D-D-D. It’s ok. The fourth one is Google’s new AI media tool set. The focus includes Gemini Oni, Flow Genie and Lirii A. Many creators are now in need of a model that can produce a video. What they lack is whether it can be changed and mixed. Can it string together images, sounds and special effects? You don’t have to worry about changing tools at every step. Yomini Oni looks more similar. The AI version of the editing and synthesis layer is not just a simple survival. Flow has gradually expanded into a more complete creation toolbox. Gini is pushing an interactive world model. Lirii A complements the music and sound production. If you can throw materials into Flow to process situations and play with effects, and then use Lirii A to generate or remix music. The entire project does not need to be cut into a bunch of fragmentary small projects. To put it bluntly, Google is now building an AI creation studio. Instead of giving you a hammer, it will give you a table of tools to slowly play with. The interesting thing about this wave of updates is not a super big title, but a very clear direction of putting it together. Although Yomini Oni has some overlap with VO-type video generation, it currently seems to be more focused on editing, synthesis and remixing. That is to say, you don’t just help me make a video, but you can use the content you already have to re-transform it. Flow also has more functions such as masking, situation exploration, prompt assistance, and experimental functions. You can even make your own gadgets in it. This is very popular for advanced creators because what everyone lacks is often not a universal model. It's just a small fix. Yomini pulled the world model to Google3view, so that real locations can be turned into explorable environments. It even started to encounter the issue of AI agents interacting in it. Larry A's side is not just about vocal music. It also includes remixing, covers, paragraph replacement, track splitting, and even prefers to make its own audio tools. Although it sounds amazing, Dingding will help you grasp the key points. Google this time is more like laying out a multi-layered creation platform to give more room for people who want to maintain a sense of control. To put it simply, people who don’t want to repeat every step should be very interested in seeing this set of tools. Okay, the last one is a little more technical, but it is actually very practical. The fifth one is TemperSix’s advanced prompt engineering course, and the focus is to use it to make real products. It is not the version that will only be followed by a few magic spells. Many people used to learn prompt engineering as if they were collecting tips. As soon as they entered the AI product agent system or the process of first looking for information and then giving it to the model to answer, they immediately succeeded because the real environment requires stability and measurability. The follow-up is easier to modify and easier to take care of. The smart class starts from the developer's perspective and will touch on system prompt words, agent planning prompts, structured output, multi-modal workflow, prompt security, and prompt evaluation. Imagine you are doing Lang Train or Langraph today. Chat machine source frameworks, system prompt words, tool descriptions, retrieval instructions, and output formats all need to match each other. At this time, you will know that the prompt project is not a trick at all. The summary of the cat system is that the smart classroom class does not teach you to throw instructions to make cats laugh. Rather, it teaches you how to raise a smart cat that is stable, well-behaved, and can really help with things. The most noteworthy thing about it is that it wants to separate the prompt engineering of serious products from the cheap and noisy content of the past. Now the title of prompt engineer may not be so popular, but the ability of prompt engineering is still very important for people who are really making AI products, because you are building a system that first looks for information and then answers. If the context is messed up, the model may not be used well even if it gets the information. If you don't write well the tool descriptions and planning prompts for agents, he may choose the wrong action. You need structured output. It is difficult to ask questions without good prompts. The content of Pampersex's class starts with the foundation and settings of a large language model and then brings to several common prompt methods, such as zero examples, which is to directly download the task to see if the model can understand it. Single example and a small number of examples. The key point is to give it a few references when you want the model to answer in what format and tone. On the Internet, it will also talk about letting the model think step by step. Cross-reference with multiple answers, plan first and then solve the problem, draft first and then organize it, and how to design a prompt strategy from a higher level. You may have seen a long list of these names in English on the Internet, but for ordinary people, it is more important to remember what kind of problem they are solving. As for the overall learning route, in vernacular, it means how to connect the data, how to arrange the process, how to manage the memory, how to operate the whole set of large language model products in a written manner, it is all strung together. This is so awesome that even the lazy cat will think of typing on the keyboard after reading it. If you are serious about making AI products, not just chatting about AI, it means that the project is still worth investing in hard work up to now. The above are the five key points of today. From the continuous assistant of Hermes agent, which is unfortunately an AI assistant, to the AI short film production of YatGPT and Google Flow, to Google's larger media tool layout, and finally to the developer-level prompt project. Each direction is different, but the common point is very simple. They are all trying to make AI from looking very powerful to actually being able to enter your daily work. If you have the time to help you look up information, if you want to fish and keep up with AI, remember to subscribe to the AI hot search report. If today's content has gained you something, you are also welcome to help me stay on Apple Podcasts. A five-star review, let this cat know that today’s work was not in vain. Search for AI hot search reports on FB IG Threads and you can find me. Today’s AI hot search reports are here. I’m going to take a nap. We’ll see you tomorrow. Goodbye, meow.

喵,我是 DaeDae,一直幫你追,AI熱搜的貓,AI熱搜報幫你劃重點,囊錄了一周 展始觀門是不是正在考慮要躺平還是躺更平?不管你今天打算幹嘛, DaeDae先把這周最新的AI重點整理好 聽學了5個最熱門的AI工具情報,10分鐘聽完,剩下時間就交給睡午覺吧 換說回來,大家最近有沒有覺得AI發展快到讓人有點壓力? 有時候比起研究新工具,本貓更想窝在窗邊曬太陽發呆 其實啊,學習,AI真的不一定要把自己搞得很緊繃,每天花了10分鐘 聽聽這些工具能幫你解決什麼實際問題 就像是幫大腦裝一段記憶模組 等真的需要用的時候,你就會發現原來解答就在腦子裡 這種男人學習法才是維持長跑的關鍵喔 今天這集我們就照順序來,不繞路也不賣弄術語 直接看這些AI工具到底在真實工作裡能幫你省下哪些麻煩 先看第一個,Hermes Zajan 很多人聽到,AI代理人都會眼睛一亮 但真的打開來又不知道,除了聊天 還能教什麼事給他對不對 Hermes Zajan有趣的地方就在這裡 他不是只回你一句答案,而是能記住事情 看任務看板、操作瀏覽器、整理記憶筆記 你早上只要把研究、寫作、回報這些事丟進Hermes 他就能去看競品網站 整理成文件格式,順手把進度寫進記憶裡 你自己就專心做人類才適合做的判斷 簡單講,Hermes Zajan想做的 不是一隻只會喵兩聲的聊天貓 而是你數位生活裡 那隻一直默默跟著、越來越懂你習慣的小管家貓 這套東西真正解決的痛點不是你缺一個答案 而是你缺連續性 像它的目標指令、早晨優先事項提示 還有記憶筆記 其實就是讓Hermes Zajan不會每次都從零開始失意 你可以先丟一個比較清楚的目標 他就比較不會亂跑 他也能開瀏覽器去研究別人的產品 攻擊了什麼功能 整理完之後再交給Coco或Codex接著做 每天早上如果他還會問你今天最重要的事是什麼 再把這些資訊記起來 久了就真的會越來越像你的個源助手 再加上TillSkill 他還能跨中置碰到你的檔案和本機工具 用手機、電腦、不同工作環境 不要像三的平行宇宙 這個必須死死鎖進肉丑筆記裡 Hermes Zajan最適合的用法就是把它當成有記憶 會規劃、會持續做事的助手 不只是聊天機器員啦 講白點它賣的不是一個答案 是你終於不用每天重新交代一次 好啦,深得懶悠我們來看下一個工具 第二個還是Hermes 但這次是另一個很實用的變形板 裝在USB、時聲碟或外接SSD上的可惜是Hermes Zajan 痛點超明顯 你好不容易把一台電腦上的AI環境調到順手 換到另一台就全部重來 設定不見 歷史沒了 API還得重接 真的很想原地睡午覺 這個做法就是把整套Hermes工作環境直接帶著走 插到Windows、Linux、MacOS都能用 聊天紀錄、記憶、API設定都跟著你跑 案例裡它搭配DeepSeek當模型Telegram當訊息入口 所以你不只是搬檔案而已 而是把同一個AI助手整隻一起帶走 像你如果平常在家用桌機、公司用筆電、出門還有旅行機 插上去就能無縫接續 昨天做到哪不用每台電腦都重新養一隻新的助理 這根本是摸魚神器吧 把工作丟給他你就可以去旁邊開罐罐了 而且他厲害的不只是可惜 還把換機器很煩跟AI示意很煩兩件事一起解掉 這套可惜是Hermes用XFAT讓不同系統都能讀 第一次啟動時會幫你把需要的Python、Node.js 跟Hermes檔案弄好 後面就省很多事 範例裡它連到DeepSeek是因為成本低又夠用 在街上TelegramBall Father見出機器人還能鎖使用者ID 避免別人亂來 最妙的是示範裡TelegramBall一度沒反應 結果Hermes Adrian自己去查 發現是使用者ID衝突 然後把設定修好 這種自己幫自己排錯的感覺 才真的像大家想像中的代理人 他甚至還能看出自己目前跑在Windows、Vu2還是MacOS 這首告訴你本機有什麼資料夾和檔案 總之這不是單純把AI放進隨身碟 而是把一個有記憶匯任環境 為跟著你移動的工作夥伴裝進口袋裡 對於常常換電腦的人來說 這真的很像把熟悉的貓屋一起搬走 安心很多 耳朵竖起來 接下來第三個絕對是今天的MVP 這個是用TrackGPT加Google Flow做完整 AI短片製作 很多創作者一碰影片就卡住 因為不是沒靈感 就是人物每個鏡頭長得都不一樣 最後成品看起來很散 這套做法把流程拉直了 先用TrackGPT想故事 定角色 拆情境 寫提示 再交給Google Flow生圖 做動作補鏡頭感 最後把片段撿起來 假設你今天想做一支30秒的動作短片 你可以先做出主角角色表 再把每一幕變成圖片和短片 最後丟進VN Filmore A或ClipChamp組起來 NOSIE總結就是 這像是把迷你攝影棚塞進你桌上 雖然不能取代你的審美 但至少能讓同一個角色同一種氣氛 不會拍到一半就走中 阿真正強的點是 這不是只靠一個炫炮工具 而是一整套順下來的工作流程 第一步先用TrackGPT伸出幾個故事方向 順便定風格和主角設定 接著最重要的是把角色做穩 不然一下帥一下不像本人 觀眾很容易出戲 這時候TrackGPT可以幫你整理出 角色的外觀、角度、表情 做成角色參考表 後面Google Flow才比較抓得住同一個人 在HGPT把劇本拆成適合短片生成的段落 Google Flow就負責把圖變影片 而能掉進位、推、拉、環、肉這些效果 讓畫面比較有電影感 最後再用簡單剪輯軟體收尾 學會這個 今天下午又可以多睡一小時無覺了 因為對新手來說最有價值的不是 AI也能做影片 於是你終於有一條不那麼混亂 可以重複用的拍片路線 說穿了 先把流程里順比一直追星模型還有用啦 如果剛剛那個沒打動你 來看看這個 D-D-D看完都行了 第四個是Google這波新的AI媒體工具組 重點包括Gemini Oni、Flow Genie跟Lirii A 很多創作者現在不缺一個 能生一次影片的模型 缺的是能不能改、能不能混 能不能把影像、聲音和特效串在一起 不用每一步都換工具換到煩 Yomini Oni看起來比較像 AI版的剪輯與合成層 不只是單純生存 Flow也慢慢擴成比較完整的創作工具箱 Gini在推可互動的世界模型 Lirii A則補上音樂和聲音製作這塊 如果你可以把素材丟進 Flow裡處理情境、玩效果 再搭配Lirii A生成或重混音樂 整個專案不用切成一堆零碎小工程 說白了Google現在像是在搭一間 AI創作工作室不是指點你一把錘子 而是整桌工具讓你慢慢玩 這波更新有意思的地方 不是一個超大標題 而是拼起來的方向很清楚 Yomini Oni雖然跟VO類型的影片生成有點重疊 但目前看起來更偏向編輯 合成重混 也就是你不只是說幫我做一段影片 而是能把手上已有的內容拿來重新變形 Flow也多了像遮罩、情境探索、提示協助、實驗功能 甚至可以在裡面做自己的小工具 這對進階創作者很香 因為大家缺的常常不是萬能模型 而是一個剛剛好的小修補 Yomini則把世界模型拉去接Google3view 讓真實地點可以變成可探索的環境 甚至開始碰到AI代理員在其中互動這件事 Larry A這邊也不只是聲音樂 還包含重混、翻唱、段落替換、音軌拆分 甚至偏向做出自己的音訊工具 雖然聽起來很神但叮叮幫你抓個重點 Google這次更像是在鋪一個多層次創作平台 以想保有控制感的人更多發揮空間 簡單說就是不想每做一步都重來的人 看到這組工具應該會很有感 好啦最後一個走技術派一點 但其實超實用 第五個是TemperSix的進階提示工程課程 而且重點是拿來做真實產品 不是那種看完只會被幾句神奇咒語的版本 很多人以前學提示工程都像在收集小技巧 結果一進到AI產品代理員系統 或是先找資料再交給模型回答的流程裡 就立刻破功 因為真實環境要的是穩定 可測 後續比較好改 也比較好顧 智能課從開發者角度切進去 會碰系統提示詞 代理人的規劃提示 結構化輸出 多模態工作流程 提示安全 還有提示評估 想像一下你今天在做Lang Train或Langraph 這類聊天機器源框架 系統提示詞 工具描述 檢索指令 輸出格式全都要互相搭的起來 這時候你就知道提示工程根本不是小把戲 貓系總結就是 智堂課不是教你亂丟指令逗貓 而是教你怎麼把一隻聰明貓養得穩定 守規矩 又真的能幫忙做事 它最值得注意的地方是想把認真做產品的提示工程 跟以前那種廉價又很吵的內容切開 現在提示工程師這個職稱可能沒那麼紅了 但提示工程這項能力 對真的在做AI產品的人還是非常重要 因為你做先找資料再回答的系統 如果上下文丟的亂 模型就算拿到資料也不一定會好好用 你做代理員工具描述跟規劃提示寫不好 他就可能選錯動作 你要結構化輸出 沒有好提示也很難問 Pampersex這堂課內容 從大語言模型的基礎和設定開始 接著帶到幾種常見提示方法 像是零範例 就是直接下任務看模型能不能懂 單範例和少量範例 重點是在你希望模型照著什麼格式 什麼語氣來回答時 先給他幾個參考 在網上還會談到讓模型一步一步想 用多種答案交叉比對 先規劃再解題 先打草稿再整理 以及怎麼從更高一層去設計提示策略 這些名字你可能在網路上看過一長串英文 但對一般人來說 先記住他們各自是在解哪種問題比較重要 至於整體學習路線 白話講就是從資料怎麼接 流程怎麼排 記憶怎麼管 要整套大語言模型產品怎麼文文運作 它都有一路串起來 用這個厲害了 連懶貓看了都想起來敲鍵盤 如果你是認真在做AI產品 不只是拿AI閒聊 那提示工程到現在 還是很值得投資的硬工夫 以上就是今天的五個重點 從Hermes agent的持續型助手 可惜是AI助理 YatGPT加Google Flow的 AI短片製作到Google更大的媒體工具版圖 最後再到開發者等級的提示工程 每一個方向都不太一樣 但共同點很簡單 就是他們都在想辦法讓AI從看起來很厲害 變成真的能進入你的日常工作 如果自己有幫你省下查資料的時間 想摸魚跟上AI記得訂閱AI熱搜報 如果今天的內容讓你有點收穫 也歡迎幫我到Apple Podcasts 留個五星好評 讓本貓知道今天沒白整理 FB IG Threads搜尋AI熱搜報 就找得到我了 今天的AI熱搜報就報到這 我要去睡午覺了 我們明天見 掰啦喵