EP34｜Reconnecting people and transporting workers！ AI is working with women on their own, Google Gemini 3.5 is truly free from harm？
AI熱搜報 · 2026-05-21 · 17 min
Show full episode description
Leave a comment and share your thoughts: 喵～我是DaeDae，一隻幫你追AI熱搜的貓！AI熱搜報，幫你劃重點！ 最近 24 小時的 AI 圈可熱鬧了，從自架開源 Agent 的風潮，到超級運算平台的大爆發，還有 Google 新款模型的性能爭霸戰。DaeDae 已經幫各位鏟屎官整理好最精華的情報，讓你用最優雅的姿勢，在沙發上就能掌握如何讓 AI 幫你自動化工作。趕快跟著今天的肉球筆記，輕鬆吸收最火熱的 AI 知識，讓你一邊摸魚、一邊走在科技最前端！ 【本集肉球筆記】 1️⃣ 自架最強開源 Hermes Agent 👉 幫你打造一個有大腦、有記憶且能不斷進化的專屬 IT 貓管家。 🎯 介紹 Nous Research 的開源 Hermes Agent；結合 Honcho 記憶層讓 AI 具備長期記憶；透過 Telegram 即時遙控，實現能自我進化並穩定執行的自動化助手。 2️⃣ Higgsfield 超級電腦的震撼彈 👉 一個視窗就能搞定拍電影、架網站到品牌經營，這種摸魚神器真的存在！ 🎯 剖析 Higgsfield Supercomputer 的全能自動化流，整合包含自動影片生成、Amazon 品牌電商全套建置與 Vibecode 網頁開發，展示 Agent 如何自主協作完成複雜商業任務。 3️⃣ Gemini 3.5 Flash 深度性能實測 👉 Google 的最新王牌是真的強，還是雷聲大雨點小？成本竟然是關鍵？ 🎯 對比 Gemini 3.5 Flash 與 GPT-5.5、Opus 4.7 等頂級模型的基準測試；揭露高 Token 消耗的成本警訊；同步介紹取代舊款開發工具的 Antigravity 2.0 與其 CLI 工具。 4️⃣ AI 自動化選品團隊，讓你躺著賺罐罐 👉 再也不用手動查數據，派出一支 AI 團隊幫你去市場帶貨回來！ 🎯 實測 Accio Work 平台，展示如何利用 AI 模擬專業商務團隊進行趨勢分析 (Trend Analysis)、供應商洞察與競品研究；強調「人機協作」模式如何噴速提升獲利效率。 5️⃣ 直覺式編碼 Vibe Coding 極限挑戰 👉 就算不懂複雜語法，也能靠 Gemini 3.5 Flash 變出超狂動畫？ 🎯 針對 Three.js 動畫與複雜的前端 UI 架構進行極限實測；探討模型在 SVG 生成與直覺式編碼 (Vibe Coding) 的應用潛力，是 Web 開發者加速產出的高效利器。 想摸魚跟上 AI，記得訂閱 AI熱搜報！我要去睡午覺了，我們明天見🐾 Powered by Firstory Hosting
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Rounds up top AI tools, spotlighting Hermes as a self-learning autonomous agent versus forgetful, brittle traditional agents.
Benefits
- Lightweight Python framework runs on cheap VPS
- Multi-model serialization across providers
- Compact self-managing memory avoids token bloat
- Self-generated skills adapt to network obstacles
- Telegram control with secure home connectivity
Use cases
- Hermes by Nous Research runs on the most basic cloud VPS for only 5 US dollars per month with fast performance
- Hermes caps USER memory to 1375 characters and MEMORY to 2200 characters, auto-reviewing every ten rounds of dialogue
- Hicseo supercomputer schedules 40+ AI tools; found anti-slip pet dog socks with 71% gross profit and generated 15 audio-visual ads
- Gemini 3.5 Flash hit 278 tokens per second but burned 1,550 US dollars in testing, 5.5 times more than the previous generation
- Hermes connects via Twingate to control Home Assistant lights, color, and curtains from Telegram while away
KPIs / results
Tools / build
- Hermes Agent (Nous Research)
- Honcho (Plastic Labs) external memory
- Hicseo supercomputer
- Gemini 3.5 Flash + Antigravity 2.0
- Twingate secure tunnel
🌐 This transcript was automatically translated to English from the original.
Yo, I am DaeDae, a cat who helps you chase AI hot searches. The AI hot search report will help you identify the key points. Do you feel that AI is developing too fast and you can’t finish learning new things every day? Don’t be afraid, DaeDae is here to help you. Today, DaeDae has selected the 5 most popular AI tool information, allowing you to easily keep up with the AI wave in 10 minutes a day. Become the smartest fishing master. You are tired from going to work and class every day. How can you still have time to test so many new AI tools? Don’t worry, today DaeDae helps you break down all these complex AI information, allowing you to listen easily while mastering the latest technological gameplay. Speaking of tiredness at work, DaeDae recently discovered that people are really prone to information anxiety when they are tired. Obviously, AI is helping us save time. In the end, we spend more time on the latest AI communication software and social updates. In fact, sometimes we put down our mobile phones and touch cats like me on the roadside, or make a cup of hot coffee. Keeping your mind completely blank for 10 minutes can actually improve the subsequent work efficiency. This wisdom of deliberately blanking is actually very similar to the topic we are going to talk about today. Automating tools allows you to regain the sovereignty of your life. Next, we will be the first to appear. This is a bit fierce. Listen up. We are going to talk about the big changes in personal digital assistants. In the open source community, everyone is accustomed to using OpenClar as the golden architecture for building autonomous AI agents. However, a new challenger called Hermes has emerged recently, causing a large number of developers and technology players to move. I decided never to look back. Hermes, built by the News Research research team, is not just an ordinary system upgrade. It is an autonomous system designed with another kind of logic. The goal is to allow AI to run uninterrupted while actively learning from your daily use. Friends who have played with AI assistants must know the tool fatigue and the system being adjusted at every turn. Debugging makes me doubt my life. In the past, OpenClar often fell into regional settings. Once the API was changed and the parameter suite was updated, the system would stop working. What’s worse is that traditional agents are very forgetful. After continuous use for a period of time, due to the rigid stacking of historical records, the entire system becomes stuck and fat. Not only does the Token consumption increase sharply, but in the end, you can only watch the system completely collapse. This kind of disaster really makes the cat want to go crazy. This is the time for the Hermes agent developed by News Research to show its skills. It is a lightweight open source framework based on Python. It was originally designed to run on a cheap computing environment. It only costs 5 US dollars per month on the most basic cloud virtual private server BPS to run very fast. It supports multi-model serialization. You can connect your existing ChatGPT or GROK subscription through the API, or directly connect to basic open source software such as LM Studio to run the Q1 model. What's even better is that it does not require a complicated web management interface. It can directly connect to Telegram, which we use every day, and connect back to your home local area network very securely through the remote secure connection tool Twingate to interact directly with the Home Assistant system or even the Unify network device at home in real time. In order to make it easier for Zen Food Officers to understand. Let's walk you through the complete deployment process. Imagine that you want to create a dedicated home IT support assistant for yourself and call him Manager Yong Eun-hye. First, rent a cloud server like Hosting RVPS and install the Ubuntu system. Then remotely connect and enter the single-stand installation package of Hermes. Enter your OpenAI key in the settings. You can even choose GPT5.5 as the main thinking engine. Then choose Telegram as the main communication channel. You can directly send messages to Bowfather on Telegram. Create your own robot token to prevent strangers from stealing your face. Use User InfoBall to find out your unique Telegram user identification code and put it close to the Hermes profile so that he will only listen to you. Finally, you set Hermes's channel to the Systemd service of Linux and let it run 24 hours a day in the background of the cloud server. After startup, you inject his core personality profile SOUL. Here you wrote the backstory of Ron Weasley. He graduated from Hogwarts. Because I am very curious about Magua technology, I decided to become a professional IT network administrator. Meow, this setting is very graphic. Then you send him the secure TWinate client key on Telegram. Because Hermes has high-privilege terminal execution capabilities, he will judge the system by himself. Download the network package and establish a TWinate secure encrypted tunnel to connect to your home. Then click on the HA login setting and he will scan the device. When you are away from home, you only need to say on Telegram, help me turn off the lights in the study, turn it to blue, and pull down the curtains. He can complete home control in time through TWingate, which even lazy cats can use. The really smart thing about Hermes lies in the unique mechanism of managing memory and self-updating. Traditional OakengCode will stuff all the history into prompt words, causing memory expansion, but Hermes directly resorts to iron-fist methods. It limits your personal habit details and environment settings to two files. UACER, far low control, is limited to 1375 characters, and MEMOR is limited to 2200 characters. When he learns new information, The old and unimportant ones will be discarded or condensed. The permanent agent in the well will automatically review and update every ten rounds of dialogue to ensure that the system is never sluggish. You can also connect to the high-level external memory processing system Honcho developed by Plastic Labs. It can simultaneously analyze your speaking habits and personality and produce user characteristics cards until Hermes provides an appropriate response. Even more powerful is the self-generated skill loop. When Fusion Connection encounters network obstacles, he will not Instead, they will analyze the network logs and write their own Python scripts to automate the next connection. At the same time, a background agent called The Curator will continue to supervise these self-made instructions and decide which ones should be archived to prevent the database from being filled with garbage code. This means that the AI agent has transformed from a rigid and huge system to a living body that can actively learn to adapt itself. Well, the second one is coming. This DayDay particularly likes it. We want to directly upgrade these automation capabilities from a personal assistant to the creative department of an entire enterprise. This platform is called Hickseo. Supercomputer is a powerful one-stop commercial-grade cloud system that silently helps you schedule more than 40 professional AI tools and large-scale models under a single chat interface. The core pain point that Hickseo solves is the sense of separation between creativity and business development. Think about the past when launching a new brand, you have to switch between dozens of tools. Use one to count keywords, another to write copy, and a third to generate product images, edit videos, and finally create a one-page web page. For small studios or one-person entrepreneurs, maintain these subscription services. It burns money and brings new tears. XFu Supercomputer integrates fragmented steps with a super smooth workflow. Let’s look at its three most eye-catching practical cases. The first one is the long film production line. It used to be a test of willpower to make a one-minute science fiction animation with a coherent plot. In Hickseo, you only need to enter a short wish. I want a one-minute science fiction film that describes the protagonist fighting monsters on different planets. The supercomputer will take over. It will draft the script first and provide three plot directions for you to choose. After you decide, It will automatically generate a character and prop setting table with a consistent style, then render it in segments and maintain the visual consistency of the picture plus sound effects. Directly export a highly complete science fiction short film. The second is a very practical money-making scenario. Construct an e-commerce brand in the jungle on Amazon. It can handle the entire product life cycle in the same chat window. You can first ask it to find three potential unlicensed popular products. The system will send a web crawler to grab the sales rankings to help you estimate profits and logistics costs. Suppose that after system analysis, anti-slip lining pet dog socks are found. The gross profit is as high as 71% and the demand is strong. After you agree, you only need to enter the help me define the brand name, main visual and one-page web page. The super computer will immediately generate the brand name Hicks and write a very creative copy. What is even crazier is the marketing package. It will lock the first product decoration image as the core design parameters to ensure that subsequent life photos and size comparison tables have completely consistent lighting and tonality. In order to make videos, it will also watch the best-selling competitor ads on Meta TikTok or YouTube Shorts. What is different is that Its image multi-modal model is to directly watch the video, mark at which second the opponent hits the pain point, what visual effects are used, and at what point in time the audience is lost. Then use these formulas to help you generate 15 types of high-quality audio-visual ads. The packaged and compressed files are directly transferred to your mobile phone on Telegram. This is simply a fish-catching product. The third is to directly customize the web page. You don’t need to write HTML, and you don’t need to find a web page, editor or rent a host. Just ask to build a website for me. Introducing Sucker Computer’s smart functions The system will analyze the design trend, add visual materials, write clean code and complete the layout. It will even automatically upload and set up the free web hosting space. Finally, click your web link directly. This means that Hixfield is no longer a small software that can write, but the most reliable super project partner. Next, the third one will appear. This is a bit of a dream. Listen up, let's turn our attention to the underlying technology. Recently, Google officially launched Gemini 3.5 Flash, accompanied by the major development tool Anti-Grevity. The reason why this release of version 2.0 is causing a lot of fuss is mainly because Google is trying to squeeze out a way to survive in the red ocean market that is surrounded by powerful enemies such as Quadopas 4.7, GPT 5.5, Timmy K 2.6, etc. that combines extreme speed, cost-effective price and low resource consumption. For the current development team, the biggest headache is often the response delay and server operating expenses. If complex programming work is done Throw it all to the most luxurious flagship model. The bill will really make people cry. To test whether it is really easy to use, let’s take a look at the real data of Gemini 3.5 Flash. It provides sincere E and M input contexts, and a single output limit of 64K for Window. It also supports powerful external inputs such as images, videos, audios, PDF files, etc. Although Google officially claims that Gemini 3.5 Flash equals GPT 5.5 in terms of program capabilities and even ranked Coldopa in the terminal command evaluation. 4.7 rubs on the ground, but the third-party evaluation agency gave a slap in the face. According to an independent report by Artificial Analysis, the programming index score of Gemini 3.5 Flash even lags behind some mid-level open source models. Even TMI T2.6 can't beat it. Compared with the previous generation, the performance is a bit modest. However, in terms of speed, no one can beat it. It has set a wafer record of outputting 278 Tokens per second, far exceeding the opponent. But this speed comes at a price. And the price is very secretive. Google advertises that it is good value for money. It only costs $1.50 per million inputs and $9 for outputs. However, third-party stress testing found that 3.5 Flash is an out-and-out big eater. In the program debugging test run by multiple agents, this model requires an average of 49 back-and-forth conversations to complete a small project when executing standard AI. During the benchmark test, the high frequency of conversations caused it to eat countless input tokens in the background. In the end, it burned a total of 1,550 US dollars, which is 5.5 times more than the previous generation. It is even more expensive than running on GPT 5.5 Medium. Oh, the burning is really scary. It must be recorded in the meat ball notes to avoid pitfalls. Be careful when allocating API quotas. To solve this pain point, Google has launched a new Antigravity 2.0 developer desktop application. There is an operation interface with a minimalist and exquisite style. There is a code comparison panel on the right side. When writing a web page directly, you can see a visually accurate web page in less than 5 minutes. However, when testing the complex back-end work of database collection, although its layout is fast, the appearance is dull and full of AI flavor. On the contrary, the same problem is given to ClawDoka 4.7. Although it takes two minutes longer, the details and beauty can be fully fulfilled. At the same time, Google also released the new Antigravity CLI. This step also stepped on the open source community because Google announced that it would officially retire the widely acclaimed open source Gemini CLI. It was rewritten in Gou language to make the operation faster, but it was replaced by the undisclosed Antigravity CLI, which caused the open source supporter forum to explode in an instant. In summary, Gemini 3.5 Flash is suitable for online text customer service that requires immediate response, or situations that require a one-page webpage sketch to be shoveled out in an instant. But if it is a large-scale complex project that requires repeated communication, your API quota must be carefully controlled. Okay, here comes the fourth one. I really like this Dayday. Since I have talked about so many technologies flying in the sky, this one will take you back to the stage of practical e-commerce. Let’s take a look at the smart platform Otio Work, which Alibaba has invested heavily in. This tool is specially designed to save e-commerce sellers who are about to die of old age. It is designed to optimize product development and source search processes. It is really tiring to run an online store alone. Find goods through hundreds of channels, confirm factory qualifications by hand, make profits and list them on multiple platforms. Most of them are on Amazon or Shopee. The boss of the FWi startup finally gave up halfway. It was definitely not because the product was not good, but because he could not withstand the torture of endless manual labor. Otio Work was born to completely solve this problem. It subverted the one-on-one chat assistant model and directly gave you a digital army. After logging in, you will not see a blank dialog box, but an e-commerce mind with a role map. The resigned Shopee FWi Operator, the Vibe Selling guidance for visual marketing, and the Coder for writing system automation. And the Gmail Assistant for writing replies. The centralized e-commerce mind is the team leader. It will automatically break down your ultimate task and distribute it to various specialists for simultaneous execution of multiple tasks. To keep 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. FB IG Shreds So the AI hot search report found me early. Today's AI hot search report is here. I'm going to take a nap. We'll see you tomorrow. Goodbye, meow.