Пост Андрея Карпаты: главная ошибка при создании AI Агентов и RAG систем
The speaker discusses Andrey Karpathy's post about a major problem in AI agent and RAG system design: the difficulty of searching and structuring knowledge bases. The core issue is that notes and knowledge fragments are stored as disconnected elements, making retrieval dependent on remembering exact keywords used during storage.
Summary
The speaker opens by describing a relatable problem many people face: maintaining personal knowledge bases in tools like Telegram or Evernote, where ideas, notes, and inspirations are stored in folders. The fundamental challenge is that retrieving specific notes requires remembering the exact keywords used when the note was created — a task that becomes increasingly difficult as the knowledge base grows over time.
The speaker explains that this problem is compounded by the fact that notes are stored as isolated, disconnected elements with no relationships between them. When browsing or searching, items surface individually without context or connection to related ideas.
This personal knowledge management challenge serves as the entry point into discussing Andrey Karpathy's post, which identifies this same structural problem as the major flaw in AI agent and RAG (Retrieval-Augmented Generation) system design. The problem is not just personal — it scales into enterprise and AI system contexts.
The speaker references practical experience building various AI agents using the N8N service, including: support service agents, HR agents for employee training, agents embedded in Telegram chats with connected knowledge bases for accurate Q&A, and sales agents in Telegram designed to sell products. These real-world implementations highlight the practical importance of solving the knowledge structuring and retrieval problem that Karpathy identifies.
Key Insights
- The speaker argues that the core failure of personal knowledge bases is that notes surface as entirely disconnected elements with no relationships between them, making contextual retrieval nearly impossible.
- The speaker claims that Andrey Karpathy identifies searching knowledge bases as the major problem in AI agent and RAG system design — not model quality or agent logic.
- The speaker observes that retrieval failure stems from a mismatch between the keywords a user remembers at query time and the keywords originally used to formulate and store a note.
- The speaker describes building multiple practical AI agents — including support, HR training, Telegram Q&A, and sales agents — all relying on connected knowledge bases via N8N, implying that RAG architecture is already widely deployed in real workflows.
- The speaker frames the knowledge structuring problem as one that compounds over time: the more knowledge is accumulated, the harder it becomes to structure and retrieve it meaningfully.
Topics
Transcript
[0:00] Well, or, you know, these bookmarks, yes, each of us probably has some kind of folder. Some people store it in Telegram, others store it, store it, like I do in Evernote. I have folders or notebooks or notes with various ideas that I would like to implement, which I found somewhere on the Internet, saw somewhere, heard somewhere. And, accordingly, I maintain such and such a knowledge base and notes between us. In order to find some topic, some idea, some note that I need, I have to [0:30] enter some keywords there . It is very common that I don’t remember which keywords I used to formulate this note. And, accordingly, it’s just that when I raise…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Игорь Зуевич
Сравнение Claude Code, Gemini и Codex — тесты на контент, анализ и изображения
The speaker is comparing AI coding assistants — Claude Code, Gemini, and Codex — by testing their ability to create image-generation agents. Each tool is being evaluated on how well it writes prompts and structures sub-agents for text-to-image and image-to-image tasks.
Экономия токенов и контекстного окна: настройка памяти нейросети – Compact сжатие контекстного окна
The transcript explains the 'Compact' command in AI agent sessions, which compresses the entire conversation history to reduce context window usage without switching sessions. It also covers related commands for full history clearing and automatic triggers for compact execution. The speaker recommends configuring automatic notifications or auto-execution when the context window reaches a certain threshold.
HR-агент на Claude Code: создаёт других сотрудников сам по запросу
The video demonstrates setting up a Telegram HR bot using Claude Code, walking through the process of creating a bot via BotFather, obtaining a Telegram user ID, and letting the code automatically generate the HR agent. The presenter shows how the system initializes itself by installing dependencies and launching the bot with minimal manual configuration.
Живые скиллы для Claude Code, Gemini и Codex | Как настроить super skills для ИИ агентов
The video explains how to create 'living' or 'super' skills for AI agents like Claude Code, Gemini, and Codex that self-update based on user feedback. The presenter demonstrates how these skills collect feedback after each task, store it in log files, and periodically incorporate that feedback into the skill's core instructions. This approach reduces token consumption and ensures skills continuously improve over time.
Планировщик для ИИ агента: 50 идей и готовые файлы для Claude, Codex и Gemini
The speaker describes their use of a scheduling/routing tool called Routins, which they compare to N8N, to automate daily delivery of curated YouTube videos on topics like Gemini and Claude. The system runs autonomously, collecting videos and generating concise summaries of their content without manual intervention.