AI Stack Explained Podcast: LLM, Tokens, Context, Function Calling, MCP and Agents (22 min)

The audio-first version of the AI Stack thesis: 22 minutes on LLM, tokens, context, prompts, function calling, MCP, agents and skills, and why they are all one pattern. Shorter than the 48-minute deep dive, longer and more conversational than the 15-minute video.

The one idea

The model only ever outputs text. The program around it reads that text, runs the tool, and hands the result back. LLM talks, program walks.

What you will be able to explain after listening

• Why a token is not a word (3:01), and why that matters for cost and for odd failures.
• How a chat “remembers” you (5:00): it doesn’t. The program resends the history every turn, inside a context window with a hard size limit.
• What RAG actually does (7:07): embed the question, find the nearest text chunks in a vector store, and put only those in front of the model.
• The difference between user and system prompts (9:00).
• Function calling end to end (10:03): the model predicts JSON instead of English, the program executes it, the model writes the final answer from the result.
• Why MCP exists (13:03): before it, every tool needed its own schema for every model vendor. An MCP server exposes tools once, in one format.
• What an agent is (15:04): a while loop around function calling that runs until the task is done.
• What a skill is (18:04): a markdown procedure the program loads only when the task matches, so a thousand of them don’t crowd the context.

Related

• Blog post with diagrams and code: LLM talks, program walks
• Original 15-min video: The AI Stack Explained
• Full 48-min deep dive: The Complete AI Architecture
• Runnable demo: function calling from scratch