AI agenti od nuly: Tohle potřebujete vědět

🔐Exclusive content on Patreon: Patreon: JirkaHernik ⚙️Best VPS for your agents: https://aox.cz/vps Most people use AI as an advisor, not as a worker. They ask a question, get an answer - and then still have to go do the clicking themselves. This is the first episode of the AI pro lidi (AI for Humans) series, where I start completely from scratch and explain how to get out of that mode. I was recently at a friend's place who pays $200 a month for ChatGPT and only ever uses the chat. That means he's leaving 95% of what he's paying for sitting unused. And he's not the only one. So I made an episode that'll leave you actually knowing what's going on inside these tools. We'll untangle three things that keep getting mixed up: an AI model, an AI chat, and an AI agent. Why a model is even called a model. What training via reinforcement learning looks like - I'll show it with a little car learning to drive in Trackmania, and with OpenAI's agents playing hide-and-seek and inventing their own strategies. Then how a pile of data turns into a large language model through pre-training and post-training, what RLHF actually is, and why people don't really train the language model itself - they train a reward model that then coaches it. Then we'll look inside the chat: the system prompt, the conversation history, your prompt - and why all of that together is called context. I'll open up leaked system prompts from ChatGPT and Claude side by side, because that's exactly where you can see what separates an agent from a chatbot. Plus inference, the context window, and why the model itself remembers absolutely nothing on its own. And finally, the whole magic trick behind an agent: a loop where the model reaches for tools on its own, reads a file, runs a command, and keeps going without you until the task is done. Plus a practical part - how to tell when a plain script is enough and when you actually need an agent, why you should commit to three tasks a day, and above all, which agent to install right now based on what you're already paying for. I tried to make this hold up a few years from now too. The tools will change, the principles won't. We'll go deeper in future episodes. LINKS FROM THE VIDEO Wikipedia - Agentic AI: https://cs.wikipedia.org/wiki/Agentn%... Model / chat / agent diagram: https://excalidraw.com/#room=b2dabf25... Reinforcement learning, RLHF, and script vs. agent diagram: https://excalidraw.com/#room=11ed6385... Trackmania Nations - AI learns to drive (reinforcement learning): AI Learns to Drive From Scratch in Trackmania OpenAI - Multi-Agent Hide and Seek: Multi-Agent Hide and Seek Database of leaked system prompts: https://github.com/YeeKal/leaked-syst... ChatGPT-4o system prompt: https://github.com/YeeKal/leaked-syst... Claude Opus system prompt: https://github.com/YeeKal/leaked-syst... WHERE TO START - AGENTS TO INSTALL Codex app (OpenAI, included with your ChatGPT subscription): https://openai.com/cs-CZ/index/introd... Claude Desktop (Anthropic): https://support.claude.com/en/article... Antigravity (Google, limited free tier): https://antigravity.google/download Grok Build (xAI, included with an X subscription): https://grok.com/build Cursor (has a usable free tier): https://cursor.com/pricing WHERE TO FIND ME Discord: Discord: discord Patreon: Patreon: JirkaHernik CHAPTERS 00:00 Intro 01:04 The definition you can leave on 01:45 Paying $200, using 5% 03:40 Why the work doesn't do itself 04:57 Why it's called a model 06:34 How AI learns to drive 09:16 When AI invents a trick 11:44 What happens inside the chat 15:30 What the model reads before it answers 19:01 Why AI remembers nothing 21:02 The whole magic of agents is a loop 24:29 Letting the loop run longer 26:29 Script or agent? 29:43 Three tasks a day 30:45 Where exactly to start 34:16 Where to go next #AI #AIAgent #ArtificialIntelligence #ChatGPT #Claude #LLM #AIproLidi #automation #agents #productivity This video has an English audio track spoken by an AI voice clone of the author. Audio created with Boson AI's Higgs Audio - https://www.boson.ai/higgs-audio