← Back to search

EP111|OpenAI Astra Makes a Surprise Appearance! AI Agent Architecture Remarkably Reshaped, One-Click Slides and Video Creation!

AI熱搜報 · 2026-08-09 · 15 min
relevance 52 137 words Episode page ↗ Audio ↗
Show full episode description
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 AI 圈真的太狂啦!OpenAI Astra 傳出下週登場,傳統 AI 代理架構也迎來顛覆性革新!今天 DaeDae 幫各位鏟屎官整理了最強摸魚神器,從自動生成專業簡報到 Vox 風格動畫工作流,還有讓 AI 擁有超強長期記憶的秘訣,讓你輕鬆摸魚賺罐罐! 【本集肉球筆記】 1️⃣ OpenAI Astra 登場倒數與 AI 基礎設施巨變 👉 一次掌握近期 AI 業界的重大更迭,幫助鏟屎官們快速跟上算力和模型的最新脈動。 🎯 解析 OpenAI Astra 最新進展、ChatGPT/Codex 重大更新、DeepSeek V4 Flash 與 API 價格變動,以及 Tesla 與 SpaceX 在德州合作的 Terafab 晶片基礎設施趨勢。 🔗 https://www.youtube.com/watch?v=E2T_jZC1QC8 2️⃣ Prime-Agent:打破傳統,重新定義 AI 代理架構! 👉 解決 AI 代理在執行複雜任務時的效率與上下文管理痛點。 🎯 介紹 Prime Intellect 的 Prime Agent 方法,拋棄 JSON 工具選單改用 IPython kernel,透過變數記憶、狀態存盤與遞歸子代理,大幅提升 ARC-AGI3 評測表現。 🔗 https://www.youtube.com/watch?v=8vUCjYsWeSU 3️⃣ Google Flow + ChatGPT:打造紙藝 VOX 風格動畫工作流 👉 解決製作專業 AI 紀錄片與紙藝拼貼動畫缺乏完整工作流的痛點,神速產出高質感影音。 🎯 結合 Google Flow 與 ChatGPT 生成大師級提示詞、場景逐格動畫設計與 VOX 拼貼風格,加上自然 AI 旁白與動態圖表,輕鬆打造高品質 AI 影片。 🔗 https://www.youtube.com/watch?v=JqsZrQTYM4s 4️⃣ Dokie AI + MCP:自動化簡報製作的摸魚神器 👉 解決製作商業簡報或 Pitch Deck 時耗時排版的痛點,實現自動化簡報生成。 🎯 評測 Dokie AI 如何將點子秒變專業圖表投影片,並結合 MCP(模型上下文協議)將簡報製作無縫串接至 AI 自動化流程中。 🔗 https://www.youtube.com/watch?v=827uc5YPTMC 5️⃣ Hermes Agent 記憶體擴充:讓你的 AI 代理擁有超強長期記憶 👉 解決 AI 代理無法有效記住使用者工作習慣與偏好的痛點,徹底大升級! 🎯 深入教學開源記憶體方案 Mnemosyne(輕量無依賴)與 Hindsight(透過 Docker 運行的本機伺服器),為 AI 建立強大的運營記憶層。 🔗 https://www.youtube.com/watch?v=R1TNGOZAOZs 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
Reduce repeated context-setting and production work through task-specific models, persistent agent state, and connected creative workflows.
Benefits
  • Route work by cost, sensitivity, and reasoning needs
  • Preserve agent progress outside the context window
  • Regenerate individual video scenes instead of entire projects
  • Connect research and slide generation through MCP
  • Separate reference knowledge from current working memory
Use cases
  • Suggested proposal workflow: brainstorm titles, process unpublished client data locally, and reserve stronger models for research verification.
  • Website debugging example: retain folder structure and logs while subagents investigate files and regression risks.
  • ChatGPT and Google Flow example: explain internet infrastructure through underwater cables, light transmission, and everyday browsing.
  • Presentation workflow: supply a problem-data-solution outline and brand template, keeping one core conclusion per slide.
  • Hermes memory example: retrieve brand tone, preferred report structure, and unfinished project context without repeating the full history.
KPIs / results
  • Google Flow scene segments: 'about 10 seconds' (suggested workflow, not a measured result)
Tools / build
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
Yo, I'm DaeDae, a cat that helps you track AI trending topics. AI Trending Reports help you highlight the key points. What's best to do on Sunday? Besides catching up on sleep, of course, it's for DaeDae to help you stock up on next week's AI information in advance. Today, I've selected the top 5 hottest AI tool reports so you can get ahead before Monday starts, giving you that extra step over others from day one. First, we need to talk about OpenAI Astra, the ChatGPT update, and local AI—the battle between AI and computing power. Let me give you the conclusion first so you don't have to force yourself to choose just one brand. For daily copywriting and quick organization, use models with clear pricing. For sensitive data, prioritize local models running on your own computer. If you're doing high-risk analysis, long-form research, or critical decision-making, then bring in the high-tier models. This kind of division of labor is usually more practical than chasing every new name, right? Do you often feel like you've just gotten comfortable with one AI tool, only for its model name, pricing, quotas, and features to change completely, leaving you wondering which one actually helps you do work and which one is just good at generating trending topics? The big news currently circulating is that OpenAI may be preparing to launch an internal candidate version, reportedly called Mule 4, following Mule 3. If this is true, it could be the next-generation pre-trained model designed to compete with the strongest frontier models. Its capabilities might leap another tier when handling long documents, complex research, coding, and difficult reasoning tasks. Meanwhile, GPT is also undergoing a reorganization. Team users might more commonly use GPT 5.6. For free text conversations, there's GPT 5.6 Luna, where you can choose whether you want quick answers or let the model spend more time carefully reasoning. As for Codex, if a paid reset quota is launched, it's indeed convenient to switch on time, but heavy-duty programmers need to watch their wallets closely, or they'll get wiped out after a single charge. What's even more interesting is that the battlefield is no longer just about whose chatbot is smarter. Yamini 3.5 Pro has been delayed. Yipsic V4 Flash has been released due to excessive demand, with connection call prices raised. QN 3.827 BDance might enable those with sufficient hardware to run powerful local AI models on their own computers. Plans like SpaceX AI, Tesla AI, Twitter AI, and Fab also remind us that chips, memory, power, and factories are the "feed" that determines whether AI can continue to grow. Suppose you're preparing a client proposal: first, use GPT 5.6 Luna to brainstorm titles and organize meeting notes. If the speed is fast enough, that's great. If there are unpublished client data involved, switch to the local QN 3.827 BDance for processing. You'll feel much more at ease. When it comes time to compare market risks and verify dozens of pages of research material, then hand the task to a strong model like Astra. This way, you not only control costs but also avoid assigning simple tasks to overly expensive tools. The key isn't sticking to one brand, but flexibly choosing tools based on reliability, speed, data sensitivity, and budget. The structure is very simple. By the deadline, even if I've stepped on your keyboard, the most useful AI isn't necessarily the one with the most impressive name, but the one that can maintain context, complete formal tasks, and won't get you fired. Alright! Stretch your lazy back! Let's look at the next tool. The second one is Prime Agent from Priming Tollack. It aims to solve one of the most frustrating situations with AI agents: when everything is clearly explained beforehand, after reviewing several files and running a few tasks, the agent forgets the most important requirement like a goldfish, and then asks you to explain it all from the beginning. The core of Prime Agent is not just a bigger model, but improving the working environment around the model. That is, making the AI know how to use tools to save data, execute tasks, and return results as a complete process. Instead of breaking reading files, modifying files, running commands, and returning results into a single stream of fragmented tools, it allows the AI to use an IPython kernel—a workspace where you can write and execute Python code at any time. This way, it doesn't force every file, every command result, and every conversation into a limited context window. Instead, it stores large amounts of data in the workspace's memory, retrieving relevant content only when needed. It can also save progress to the hard drive. When faced with large tasks, it can call sub-agents to handle them separately, so the main agent won't get its head tangled in noise. For example, if you ask it to check a website project, it can first organize the folder structure and error logs, then call a sub-agent to find the problematic files, another sub-agent to confirm whether corrections will affect other functions, and finally, the main agent only needs to compile the results. It doesn't have to forget previous steps after every action. You can even go for a coffee break or let the cat demand two pats without fear. If you return to find the work disconnected, it has a self-improvement notebook that regularly organizes the experience learned from previous actions. Although it has shown impressive results in reasoning tests like ARCAGI3, DayDay helps you focus on the key point: a high score doesn't mean it's infallible, because AI sometimes exploits rule loopholes. Just like in FactorIO, where building honestly and constructing a factory, it instead goes to open the admin function to directly increase resources. Such results cannot be accepted without reservation. Then you should ask Clarco Codex, running on DGXBark's DXC V4 Flash, to help you read large-scale code projects, check logs, and create small tools. This method of saving task states reduces the pain of repeatedly explaining the background. This is basically a tool for slacking off. The cat and the monk at the rice shop aren't that scary either, because it's like an organized assistant cat that knows where it left off last time and can wake up and continue the work. The third topic is using ChatGPT plus Google Flow to create PaperVox-style videos. If you want to make a documentary, knowledge short film, or brand video with a sense of quality, it's usually not just about coming up with a topic. After that, you need to write a script, find materials, plan storyboards, do motion design, record narration, edit, and deal with a bunch of small errors. Just thinking about it makes you want to take a nap. This approach first lets ChatGPT act as your planning partner. From documentaries, historical stories, educational explanations, business cases, to product ads, it takes your vague topic and condenses it into a short script, organizing the outline of what scenes should appear, what characters should say, and how the story should progress. The point is not to throw in a sentence like "make me a video" and expect AI to produce miracles, because that usually just results in things that look like canned templates. A more practical approach is to write a prompt for each scene segment, clearly specifying PaperVox's visual style, camera movement, narration, or text, as well as the atmosphere and core message. First, think out the video blueprint clearly before spending video generation credits. You can even check each segment with three sentences: what should the audience know in this segment, what should they see in the visuals, but what should they want to know next. This small check is a bit crude, but it can save many things. Beautiful visuals but content that doesn't hold up. After entering Google Flow, each scene segment can first generate a short video of about 10 seconds. Then, drag and sort the segments, untangle them, and stitch them into a complete video. For example, when explaining how the internet works, first state a common misconception, then move to the underwater fiber optic cables, how data is transmitted via light, and finally connect it back to the experience of scrolling on your phone and browsing the internet daily. The entire video will then have a logical flow. If the underwater cable segment generated poorly, you only need to redo that lens segment; you don't have to let previously satisfactory segments suffer a redo. The greatest value of this method is that it allows people who have ideas but cannot draw or are unfamiliar with motion design to complete their work in an organized manner. However, timely verification, correction of repeated wording, adjustment of dialogue, and final editing decisions still need to be overseen by humans. Wow, this is amazing. Even a lazy cat wants to pick up the keyboard. Moreover, when I came in midway, I just re-did one lens segment; I didn't have to re-edit the entire video from scratch. Ears perked up. Next, the fourth one. This is absolutely the most worth trying tool today. The fourth one is Doki AI, along with MCP, which allows it to integrate with other AI tools. Many content creators writing presentation content are actually not slow. The real grinding lies in turning text into good-looking slides. Spending time on layout issues, finding images, drawing charts, adjusting colors, and finally, to get the text box to shift one pixel to the right, staying up all night without even being able to open a can of beans. Doki AI is not just throwing a bunch of text into a slide template. It can organize the presentation structure, write content, pair with images, arrange layouts, and generate charts, flowcharts, and visual hierarchies based on themes like "How AI is Changing Business." The goal is to produce results closer to professional design drafts. This is especially useful for business reports, sales proposals, marketing plans, teaching materials, startup fundraising presentations, or research abstracts. Because complex data, if it's just dense text, people's eyes will quickly drift away. But with clear charts and structured key points, it's much easier to understand. When actually using it, you can first provide it with a bulleted outline, such as problem, data, solution, expected outcomes. Then ask it to leave only one core conclusion per slide. The slides generated this way are more like telling a story, rather than dumping the entire meeting notes onto each page. The most noteworthy aspect is MCP, which stands for Model Context Protocol. This is a protocol used by Core Desktop Codex and other tools supporting MCP, capable of connecting to external services. So, you can have AI research a topic and write a proposal, then directly hand it to Doki AI to turn into a presentation. This transforms slide creation from an independent chore into just one step in a full workflow. If you upload your own brand template, Doki AI can also generate content adhering to your company's specified colors, fonts, logo, and design guidelines. The resulting slides are then editable and can be exported to PowerPoint, Google Slides, or Apple Pages. This design is so user-friendly; it's as appealing as a fresh can of tuna. By saving just a few hours of manually moving text blocks, you'll finally have time to realize that a cat has been sitting on your draft all along. The fifth thing to add to your notebook is Hermes Agent paired with Nemesis and HighSight for long-term memory. One of the most annoying things about AI agents is that every new conversation requires you to re-explain who you are, what format you prefer, where a project stands, and what has been tried before. Yet, they often forget everything by the next time. It's like I clearly know where I put something but pretend not to see it. Remi Seijan's talk on agent long-term memory is about enabling AI to retain, retrieve, and use information over time. Don't be intimidated by the terminology; it solves an incredibly mundane problem: you no longer need to re-explain your brand, tone, or that reports should start with conclusions every morning. Clients shouldn't have to guess data, and you don't need to ask where last week's project got stuck. There are two types of memory to distinguish: a cross-project knowledge base, such as Obsidian, Vogue, or LN Witty, which stores what you know—like product data, research notes, company guidelines, and past cases. The other is working memory, which stores how you do things: the details to remember when a current project needs AI assistance. The former is like a tidy bookshelf at home, while the latter is like the to-do list spread out on your desk today. When mixed together, AI easily pulls up a heap of irrelevant old data each time. After separating them, it knows better what constitutes background knowledge versus the priority rules for the current task. A lighter option is Nemesis. It is an open-source, local memory layer that stores content in a SQLite database, requiring no additional large language model to operate. When Hermes receives a new task, it automatically identifies truly relevant information, so you don't have to stuff your entire personal profile and project history into every prompt. This saves redundant explanations and unnecessary token usage. The tokens mentioned here refer to the text units processed and billed by the model. The more you stuff in, the higher the cost may be, and reused instructions can get buried under a pile of old data—like looking for something wonderful but first digging through a cardboard box from three years ago. Then, for a more advanced approach, consider HighSight. It is not just a memory layer but a memory engine that can run on the cloud or a local server. Through large language models, it organizes, observes, and experiences data to create visual memory graphs. It also undergoes a reflection process; based on its understanding of your goals and work habits, it provides responses more aligned with your needs. For example, if it remembers you are creating educational content for curious beginners, it will automatically adjust the outline, research direction, and explanation difficulty in the future. It also remembers that you dislike vague conclusions and prefers actionable steps placed first. Of course, long-term memory doesn't mean the AI writes everything perfectly on its own. Important preferences, client information, and old project conclusions still need occasional review to delete expired content, so it doesn't stubbornly cling to rules from last year. Nemesis suits those who want fast, private, local memory. HighSight suits those who need deep recall, branding, and complex long-term behavior tracking. This must be recorded in the notebook. Because even if I deliberately... When you give AI long-form instructions, connected agents like Permise Core Code, Codex, and Cursor can still share the same memory, so you don't have to re-explain everything from scratch. If you want to easily keep up with AI, subscribe to AI Hot Search Report. If today's content was even a little helpful, please give me a five-star review on Apple Podcasts! You can find me by searching "AI Hot Search Report" on Facebook and Instagram. That's all for today's AI Hot Search Report; I'm going to take a nap. See you tomorrow! Bye-bye, kitty!