Companion videos

How AI generates answers: the mental model that makes prompting click — companion videos

The article gives you the practical mental model — once you see an LLM as a probability-weighted text completer, prompting stops being magic and starts being engineering. These two videos are the canonical, beginner-friendly long-form versions of the same model, delivered by the person who helped build a lot of this stuff.

Primary pick

59:48

[1hr Talk] Intro to Large Language Models

Andrej Karpathy

A founding OpenAI engineer giving the cleanest hour-long explanation of what an LLM actually is, in plain English. The "psychology of an LLM" section is exactly the mental model the article is trying to install — why these models hallucinate, why they're good at vibes and bad at counting, why prompting is closer to coaching than coding.

What you should get from this: Build a non-mystical mental model of LLMs as prediction systems with context windows, training data, sampling and failure modes.

Watch or know first: None; the talk is long but beginner-friendly.

AI Expert note: This remains one of the strongest conceptual explanations. Avoid copying anthropomorphic language too literally: models do not think or know; they generate likely continuations under constraints.

Open video page

Also worth watching

2:11:12

How I use LLMs

Andrej Karpathy

A follow-up where Karpathy actually sits in front of ChatGPT, Claude, Gemini and Perplexity and uses them like a power user. Watch this after the article and his "Intro" — you'll see the mental model put into practice in real time, including the small choices (which model, which mode, when to paste vs upload) that separate frustrated beginners from people who get useful answers on the first try.

What you should get from this: See how an experienced user switches tools, modes and context strategies while keeping verification in the loop.

Watch or know first: Watch or read a basic LLM explanation first so the live workflow does not feel like magic.

AI Expert note: The workflow habits are stronger than the exact product choices. Re-check current model defaults, file behavior and privacy controls before copying any upload-heavy pattern into client or workplace use.

Open video page