EP43｜AI decision-making process! Claude 4.8 Reality, body stability, full control!

Yo, I'm DaeDae, I've been helping you chase AI hot search cats, AI hot search treasures help you draw key points. You've been busy for a week, are you considering whether to lie down flat or lie down flatter? No matter what you plan to do today, DaeDae first sorted out the latest AI highlights this week. He selected the 5 most popular AI tool information. After listening to it for 10 minutes, the rest of the time will be handed over to Xue Wujiao. Having said that, when DaeDae was sorting out these materials this week, he suddenly felt that the current AI development speed is so fast that it is a bit like a roller coaster. In the past, we always thought that AI was the kind of cold robot that would answer whatever you asked him. But over the past few days, I feel that AI is becoming more and more like a very discerning colleague. He no longer just waits for you to order, but starts to take the initiative to help you put the dishes and even cut the fruits after the meal. This trend from passive hard work to active process really makes me feel that the efficiency improvement in the future will no longer depend on how well you write instructions, but how well you assign tasks. Let’s look at the first one, Hermesagen. The most annoying thing about many technical jobs is not that you don’t have the answer, but that you clearly know that you have to do a series of things. The more you have to watch each step step by step, it feels like you are wiping the butt of a machine, right? You have to look up information at one time, open a remote environment at another time, wait for a certain step to finish before you can pick up the next one. The whole person is like being tied in front of the screen. The focus of Hermesagen is to split a request into multiple pieces of work and hand them over to different AI agents to handle them at the same time. It can also hang in the cloud and continue to run all day long. To put it bluntly, he will not only give you a few suggestions. So it’s really like finding a few helpers who won’t get tired to do things separately. What’s great about him is that he integrates Telegram control, a Kanban-style task panel, and built-in capabilities like Takagi. You can think of it as a model where he can run the whole process by himself. That is, if you explain certain tasks clearly in advance, he can do it all the way down. He doesn’t have to wait for you to agree at every step. Moreover, the real value of this type of tool is not how cool the function nouns sound. So it is very suitable for the kind of work that has many steps and waits, but cannot completely ignore it. People who do natural research and do red team testing. Long-term monitoring, service exploration, and even some development, maintenance and operation tasks are easy to get stuck in the process. There are many switches and manual confirmations. In the past, you might start the window by yourself. Check the data on the left, run the script on the right, and keep an eye on the messages in the middle. Now it is more like dividing the tasks first and letting the agents run separately. Only input from employees is really needed. Or a place with high risk, and then call you back to confirm. Imagine if you are a security researcher, you can use Hermes agent to set tasks, let one agent check the information, and another agent compiles the returns, and finally collect the results in the same place. Or if you are a general engineering team, you can also leave routine inspections, record records, and follow-up progress to him. This way, you don’t have to keep cutting back and forth, but can focus on judgment and decision-making. Mossi concluded that the real thing about Hermes agent is not AI or harm. Instead, he organized the technical work that was originally fragmented and annoying into a more structured and reusable approach. This must be locked into a pen and record. Okay, I am too lazy. Let’s look at the next tool. The second one is Unslow Studio. Many small teams have the same stuck point when using AI, that is, the large model is very strong, but the things they say are not like your company, your files, and the tone of your industry. You ask him to write a customer service reply, and he is now being templated. You ask him to organize professional content, and he often writes in the space style. At this time, the problem is not necessarily that the model is not strong enough. Many times it is just that it has not learned how to speak in your field. What Unslow Studio does is very straightforward. It makes fine-tuning the machine easier to use, allowing you to use open source models to train for your own data. It does not necessarily require a large cloud architecture. This is very important for many small companies that do not have a large budget or a dedicated machine learning team. He helps you resolve even the most troublesome information and preparations. You can use ready-made information, or you can organize training materials from PDF reports, verbatim manuscripts or company knowledge documents. This is actually very important because the most valuable thing for many teams is not the public information found on the Internet, but the documents, cases, response logic and professional opinions accumulated over the years. As long as these things are organized, the taste of the model's speech will be much different, the customer service reply will be more like your family's tone, and the legal summary will be more like what your team wrote. You won’t have to outsource it to strangers every time because it can run locally, so on the one hand you can pay less API fees. On the other hand, it is more suitable for teams that don’t want to throw away private information, such as contracts, legal records, internal operation manuals, customer Q&A, and consultant reports. Many companies are unlikely to send this information directly to outsiders. If you want to reap the benefits of AI and keep the data in your own hands, the local solution is particularly attractive. For example, small consulting companies, law firms, or internal support teams can use their professional documents to fine-tune a small open source model. Make an AI that is faster, cheaper, and answers more like your own team. You can think of it like raising a cat that understands your family's rules better. It doesn't have to start from the beginning every time, and it is less likely to talk nonsense. Duandian helps you focus on the key points. The value of OnSlow Studio is not just to train the model, but to give you a real opportunity to raise an AI helper that understands your family language and can help you make a lot of money. If the one just didn't impress you, let's take a look at this. After Duandian read it, you woke up. The third one is Coco paired with Mobbing 9CP. App is also an application. What people on my website lack most is actually not free play, but good reference, good taste, and a way to quickly turn design research into usable directions. Because the most tiring part of design is often not the drawing itself, but that you have to know the current situation first. How do you do it, which processes are smooth, which interfaces are easy to understand, and then draw out the direction suitable for your own product. This method connects Coco to Mobbing through MCP, allowing him to directly see the interfaces of a large number of high-performance mobile apps and websites, that is, UI and the overall process. It is no longer just a business design for you out of thin air. To put it bluntly, it does not ask the AI to guess what good design is, but let it see enough mature cases first and then help you sort out the key points. This is a big difference because many people have no ideas. But there is a lack of a fast enough way to collect scattered inspiration into executable directions. The most popular thing about Coco is that Coco can help you find the registration process, guide page, product introduction page, good cases of blocks, organize references, compare differences, and even send things to Figma to analyze what is done well and why it looks advanced. Councilors may have to save one by one, mark one by one, and then go back and organize it into notes. Now it is more like someone helping you run through a round first and fishing out the things worth seeing. This is especially useful for people who are in a hurry because it is shortened. It’s not the final steps, but the long period from design research to the first version of the concept. Suppose you are a project designer, an independent entrepreneur, or you are trying to quickly complete the registration process to become a beautiful product developer. Using Claw Code Mobin NCP and Figma, you can quickly grab ten strong references and make the first version of the concept. For example, if you want to make an accounting app today, you can call it first and look for the novice guidance process that has been done well, and then see how others arrange the columns, how to reduce stress, and how to make the first step smooth. You are not copying, but standing on a better starting point. This is amazing. Even lazy cats will want to type on the keyboard after reading it, because it makes Claw Code not only a programming assistant, but also upgraded to your design research partner. For many product teams, this also means one thing. In the past, design research was easily skipped, not because it was unimportant, but because there was no time. Now if there are tools to help you organize the search, compare and convert these steps into drafts, you will have a better chance of finding a balance between fast and good. Keep your ears up. The next fourth one is definitely the protagonist today. The fourth one is Claw Dooper 4.8 and the new dynamic workflow function of Claw Code. People who often use AI to write things or make products must understand that some models are very strong on paper, but they often only achieve 80% when used directly. The remaining 20% is still made up by yourself slowly, until you are suspicious of cats. It is not that it can’t do it at all, but that you have to take over and finish every time, which will really get tiring after a long time. Claw Dooper 4.8 focuses on better understanding of fuzzy requirements. It is also easier to catch what users really want. At the same time, Claw Code also has the ability to adjust the level of investment and split large tasks to multiple assistants. Simply put, not only is the model itself smarter, but the entire work process is also more complete. You can think of it like it used to be that you got an assistant who could answer questions. Now it is more like getting a partner who will first think about what to do and will come back to check after it is done. In the past, you had to check and fix problems by yourself according to the functions of AI business. Now it is more like being able to plan first and then test. Recheck and enhance. Even things like mobile phone board optimization can be taken into account. This difference may not necessarily be obvious in simple tasks, but as long as the task is large, such as making a dashboard, sorting out old projects, publishing information, and adjusting multi-page logic, you will be very concerned about whether it can make you go back and put out fires. For example, if you want to make a dashboard today, drastically organize the system or move data, you may rather wait 45 minutes to get a solid result than put out fires for three hours. This actually points out Cloudoper. The positioning of 4.8 is not the most resource-saving player who has to teach it everything. Instead, it is more suitable for high-cost and low-fault-tolerance tasks. That is to say, if you do something wrong, it will be very troublesome and it will be expensive to repair it later. So you are willing to let it think more and run more, just to deliver a relatively complete result the first time. This thing is easy to use, but if it is mined, it will be deducted because it usually spends more resources and runs longer, but in exchange for more complete and more reliable output. A more practical way to use it is to reserve it for really important, really complex work that you really don’t want to overturn. Dayday’s key note is Anthropic. It’s not just throwing out a new model name, but AI advancement that is better at understanding human language on the Internet and can get things done. Okay, the last one is also very practical. The fifth one is UN extract. The biggest headache for many companies is not that there is no data, but that the data is stuck in a mess of PDF scanned files, invoices, report forms and instructions, which are obviously useful but difficult to organize. A format that can really be used. You clearly know that the answer is in that document. However, every time you look for it, it is like searching through ancestral scriptures. Especially when there are so many files, so many versions, and so many original meanings, just finding the right file takes up half your life. UN extract is an open source file interception platform. It is easier to ask questions and better to receive information from other tools. This is very important for many small and medium-sized companies because it is not that everyone has no information, but the information is locked in the file. It can be used with the local AI to set up the file processing process. There is also the question workbench of Promed Studio. After you upload the file, you can directly ask questions. After you throw the file in, you can directly ask what is written in this document. What is the key point and which paragraphs are related to each other. This is very friendly to people who do not want to touch too many technical details. If you hear RAG, it is a nod. In vernacular, it is It doesn't just rely on what the model originally knew, but it will look back at the document in your hand before answering. In other words, it is not just based on impressions, but more like replying to you while browsing the information. In this way, whether it is a contract, financial information or a hardware manual as thick as a brick, it has a better chance of being digested quickly. For example, a small company needs to check inventory or salary information, a legal team needs to translate a contract, a creator wants to find answers quickly from a 200-page manual, or even administrative colleagues want to find policy terms, invoice fields, and employee registration information. You can use UNS-Track to query local files, and even use APIs to connect to other systems. It is a process that can be used in work and can be used repeatedly. The more you use it, the more smoothly you use it. Today's five themes have the same direction. Hermesagen is handling multi-step and fragmented technical tasks. OnSlow Studio is helping you make more personal local models. Clocko and MobbinMCP are accelerating design research. Clawdopas4.8 and Clocko's new functions are improving the completeness of large-scale development tasks. UNS-Track is digging out the information in the file and turning it into usable data. If you look carefully You will find that together they help you reduce the time of switching, counterattacking and nailing the process. To put it more bluntly, these five tools are helping you smoothen the broken work and reduce the wasted work. When it comes to fishing and keeping up with AI, remember to subscribe to the AI Hot Search Report. If today’s content has gained you something, please help me leave a five-star review on Apple Podcasts. If you want to find me, go to FB, IG, Threads, AI Hot Search Report. Today’s AI Hot Search Report is here. I’m going to take a nap. We’ll see you tomorrow. Goodbye, meow!

Yo,我是 DaeDae,一直幫你追AI熱搜的貓,AI熱搜寶幫你畫重點 忙碌了一周,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 精選了5個最熱門的AI工具情報,10分鐘聽完剩下時間就交給血無教吧 話說回來,這周 DaeDae在整理這些資料的時候 突然覺得現在的AI發展速度快到有點像在做雲霄飛車 以前我們總覺得AI是那種冷冰冰,問他什麼他才回什麼的機器人 但這幾天看下來,我覺得AI越來越像那種很有眼力見的同事 他不再只是等你點菜,而是開始會主動幫你把碗筷擺好 甚至連飯後水果都切好了 這種從被動硬打轉向主動流程化的趨勢 真的讓我覺得 墓許未來的效率提升不再是看你有多會寫指令 而是看你有多會分配任務 先來看第一個,Hermesagen 很多技術工作最煩的不是你沒答案 而是你明明知道要做一連串事 越還得自己一個步驟一個步驟盯著跑 超像在幫機器擦屁股對不對 一下要查資料,一下要開遠端環境 一下要等某個步驟,跑完才能接下一段 整個人就像被綁在螢幕前面 Hermesagen的重點就是把一個請求拆成多段工作 交給不同AI代理人同時處理 還能整天掛在雲端持續跑 白話一點講,他不是只會回你幾句建議 於是真的像幫多找了幾個不會喊累的小幫手分頭把事情做下去 他厲害的地方在於把Telegram控制 看板式的任務面板 還有像高木這種內建能力整合在一起 那某你可以把它想成整段交給他自己跑的模式 就是某些任務你先交代清楚,他就能自己一路往下做 不用每一步都等你按同意 可以把整個流程排好,能並行的就一起跑 不用你一直來回切分頁 而且這類工具真正有價值的地方 不是功能名詞聽起來多帥 於是他很適合那種步驟很多,等待很多,但又不能完全不管的工作 像自然研究做紅隊測試的人 長時間監控,服務探查,甚至某些開發 維護跟營運任務都很容易卡在流程碎 切換多,人工確認多 以前可能是自己開始的視窗 左邊查資料,右邊跑腳本,中間還要盯訊息 現在就比較像先把任務分好,讓代理人各自去跑 只有真的需要員工輸入 或風險高的地方,再把你叫回來確認 想像一下如果你是資安研究員 就能用Hermes agent設定好任務 讓一個代理人查資料,另一個代理人整理回報 最後再把結果集中到同一個地方看 或者你是一般工程團隊 也可以把例行檢查整理記錄 追進度這種雜事交給他先跑 這樣你腦袋不用一直切來切去 反而能把注意力放在判斷跟決策 Mossi總結一句 Hermes agent真正有感的地方 不是AI或危害 而是他把原本很碎很煩的技術工作 整理成更有結構也更能重複利用的做法 這個必須死死鎖進肉籌筆記錄 好啦,深得懶腰我們來看下一個工具 第二的是Unslow Studio 很多小團隊用AI的卡點都一樣 就是大模型很強沒錯 可是講出來的東西就是不像你公司 不像你文件 也不像你那個產業的語氣 你叫他寫客服回复 他現在被範本 你叫他整理專業內容 他又常常寫的太空範 這時候問題不一定是模型不夠強 很多時候只是他根本還沒學會 你這一行平常怎麼講話 Unslow Studio做的事很直接 他把本機微調這件事 變得比較容易上手 讓你可以拿開源模型 為自己的資料去訓練 不一定要搞很大一套雲端架構 對很多沒有超大預算 也沒有專門機器學習團隊的小公司來說 這件事很重要 因為大家不是不想做 而是一聽到要準備環境 整理資料設定訓練流程 就先累一半 更貼心的是 他連最麻煩的資料及準備都幫你解化 你可以用現成資料 也可以從PDF報告 逐字稿或公司知識文件 去整理出訓練材料 這點其實很關鍵 因為很多團隊最值錢的東西 不是網路上找得到的公開資料 而是自己內部多年累積的文件 案例 回復邏輯跟專業說法 只要把這些東西整理進去 模型講話的味道就會差很多 客服回復會更像你們家的口吻 法務摘要也比較像你們團隊寫出來的 不會每次都向外包來的陌生人 因為他能在本機跑 所以一方面可以少付API費 另一方面也比較適合 不想把私密資料丟出去的團隊 像是合約 法務記錄 內部作業手冊 顧客問答 顧問報告 這些資料很多公司都不太可能 放心直接往外送 那如果你又想吃到AI的好處 又想把資料留在自己手上 本機方案就特別有吸引力 像是小型顧問公司 法律事務所或內部支援團隊 就可以把自己的專業文件 拿來微調一個小型開源模型 做出更快更便宜 而且回答更像自己團隊的AI 你可以把它想成 养一隻比較懂你家規矩的貓 不用每次都從頭交 也比較不會亂講 叠叠幫你抓重點 OnSlow Studio的價值不只是訓練模型 而是讓你真的有機會 養出一隻懂你家語言 能幫你賺罐罐的AI小幫手 如果剛剛那個沒打動你 來看看這個 叠叠看完都醒了 第三個是Coco搭配Mobbing 9CP 很多座手機 App也就是應用程式 我網站的人最缺的 其實不是自由發揮 而是好參考好品味 還有把設計研究快速變成可用方向的方法 因為設計最累的地方常常不是畫圖本身 而是你要先知道現在大家怎麼做 哪些流程順 哪些界面畫面好懂 然後再從裡面抓出適合自己產品的方向 這套做法把Coco透過MCP接到Mobbing 讓他能直接看大量高表現手機 App跟網站的界面畫面 也就是UI 還有整體流程 不再只是凭空生意的設計給你 這裡講白一點 它不是叫AI自己亂猜 什麼叫好設計 而是讓他先看夠多成熟案例 再來幫你整理重點 這個差別很大 因為很多人不是沒想法 而是缺一個夠快的方法 把分散的靈感收成可執行的方向 它最香的地方在於 Coco可以幫你找註冊流程 導引頁 產品介紹頁 區塊的好案例 整理參考 比較差異 甚至把東西送進Figma 在順手分析到底哪裡做得好 為什麼看起來高級 議員們可能要自己一張一張存 一個一個標注 再回頭整理成筆記 現在比較像 有人先幫你跑過一輪 把值得看的東西撈出來 這對趕時間的人特別有感 因為它縮短的 不是最後畫稿那幾步 而是從設計研究 到做出第一版概念中間那一大段 假設你是建案設計師 獨立創業者 或正想趕快把註冊流程 做漂亮的產品開發者 用Claw Code Mobin NCP再加上Figma 就能很快抓到十個強參考 做出第一版概念 比如你今天要做一個記帳 App 你可以先叫它 找目前做得好的新手引導流程 再看看人家怎麼安排欄位 怎麼減少壓力 怎麼把第一步做得順 你不是照抄 而是先站在比較好的起跑點上 用這個厲害了 連懶貓看了都想起來敲鍵盤 因為它讓Claw Code不只是寫程式助手 還升級成你的設計研究夥伴 對很多產品團隊來說 這也代表一件事 以前設計研究很容易被跳過 不是因為不重要 是因為沒時間 現在如果有工具幫你把搜尋整理 比較轉成草稿這幾步接起來 你就更有機會在快跟好之間找到平衡 耳朵竖起來 接下來第四個絕對是今天的主角 第四的是Claw Dooper 4.8 還有Claw Code新增的動態工作流程功能 常常用AI寫東西或做產品的人一定懂 有些模型账面上很強 直接用起來卻常常只做到八成 剩下兩成還是你自己慢慢補 補到懷疑貓生 不是它完全不會做 而是你每次都得接手收尾 久了真的會累 Claw Dooper 4.8這次主打的是 對模糊需求的理解更好 也比較能早一點抓到 使用者真的想要什麼 同時Claw Code也多了能調整投入程度 把大任務拆給多個小助手去處理的能力 簡單講就是不止模型本身更聰明 整個做事流程也更完整 你可以把它想成 以前是你拿到一個會回答問題的助手 現在比較像是拿到一個會先想怎麼做 做完還會回頭檢查的搭檔 以前你照AI生意的功能 常常還得自己驗 自己修自己查問題 現在它比較像能先規劃 再測試 再檢查 再補強 連手機板优化這類事情 也比較能一起顧到 這種差別在簡單任務上 也許不一定超明顯 但只要任務一大 像是做一個儀表板 整理舊專案 頒資料 調整多頁面邏輯 你就會很在意它是不是能少讓你回頭救火 像你今天要做一個儀表板 大幅整理救系統或資料搬遷 你可能寧可等45分鐘 拿到一板扎實結果 也不想三小時都在救火 這其實也點出 Cloudoper 4.8的定位 它不是那種最省資源 什麼事都要教它上的選手 而是比較適合高成本 低容錯的任務 也就是說事情如果做錯會很麻煩 回頭重修很貴 那你反而願意讓它多想一點 多跑一點 只求第一次就交出比較完整的成果 這玩意好用歸好用 但採坑的話 可是會被扣罐罐的哦 因為它通常會花更多資源 跑比較久 不過換來的是更完整 更可靠的產出 比較實際的用法會是 把它留給真的重要 真的複雜 真的不想翻車的工作 Dayday的肉籌筆記是 Anthropic 不是只丟一個新模型名稱出來 而是在網更會理解人話 也更會把事情做完的 AI前進 好啦 最后一個也很實用 第五個是 UN extract 很多公司最頭痛的 不是沒資料 而是資料都卡在 亂七八糟的PDF 掃描檔 發票 報告表單跟說明書裡 明明有用 卻很難整理成 真的能拿來用的格式 你明明知道答案就在那份文件裡 可是 每次要找都像在翻祖傳經書 尤其 文件一多 版本一多 原意多 光是找對檔案 就先耗掉半條命 UN extract 是一個開源文件截取平台 目標就是把這些格式很亂 握好直接整理的文件 變成可查詢 可抽取 可重複利用的結果 白話講就是幫你把 原本東一塊 西一塊的文件內容 變成比較好找 比較好問 也比較好 接到其他工具裡的資料 這對很多中小公司很重要 因為大家不是沒有資訊 而是資訊被鎖在文件裡 它可以搭配本機 AI 設定文件處理流程 還有Promed Studio這個提問工作台 用你上傳文件後 直接問問題 你把文件丟進去後 就可以直接問 這份文件裡到底寫了什麼 重點在哪 哪幾段彼此有關 這對不想碰太多技術細節的人很友善 如果你聽到RAG有點頭大 白話就是 它不只靠模型原本知道的東西 而是會回頭看你手上的那份文件再回答 也就是說它不是憑印象亂白 而是比較像一邊翻資料一邊回你 這樣一來不管是合約 財務資料還是厚到像磚頭的硬體手冊 都比較有機會被快速消化 像小公司要查庫存或薪資資料 法務團隊要翻合約 創作者想從200頁手冊裡快速找到答案 甚至行政同事要找保單條款 發票欄位 員工報到資料都可以用UNS-Track做本機文件查詢 甚至在用API串進其他系統 能用在工作上的流程 而且是可以反复使用 越用越順的那種 今天這五個主題 方向其實很一致啦 Hermesagen在處理多步驟又很碎的技術任務 OnSlow Studio在幫你做更貼身的本機模型 Clocko加MobbinMCP在加速設計研究 Clawdopas4.8跟Clocko新功能在提升大型開發任務的完全度 UNS-Track则是把文件裡的資訊挖出來變成可用資料 如果仔細看 會發現他們共同在幫你減少切換、反攻還有釘流程的時間 說更直白一點 這五個工具都在幫你把碎工變順、把白工變少 講摸魚跟上AI記得訂閱AI熱搜報 如果今天的內容讓你有點收穫 那就幫我到Apple Podcast留個五星好評吧 想找我就去FB、IG、Threads,AI熱搜報 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!