ynarwal (Independent) · 2026-04-24 · notable
How LLMs Work — Interactive Visual Deep Dive (179 HN Points)
Scroll-based interactive walkthrough of the full LLM pipeline — tokenization, pre-training, RLHF, RAG, hallucination, and LLM psychology — built by feeding Karpathy's lecture transcript to Claude Code. 179 HN points today.
A scroll-through visual guide covering the complete LLM pipeline from raw data to RLHF — built with Claude Code from Karpathy's lecture.
Key specs
| Hn points | 179 |
|---|---|
| Hn comments | 46 |
| Git hub repo | ynarwal/how-llms-work |
What is it?
An open-source interactive website that walks through the complete LLM lifecycle: internet data collection, Byte Pair Encoding tokenization, neural network pre-training, supervised fine-tuning, RLHF, hallucination, memory types, tool use, knowledge cutoffs, and retrieval-augmented generation. Transparently created by feeding Andrej Karpathy's lecture transcript to Claude Code, then iterated based on community feedback.
How does it work?
Scroll-based single-page layout with interactive code snippets, visualizations, and explanatory text. The creator used Claude Code to generate the site from the lecture transcript and subsequently corrected errors flagged in the HN thread — including a claim that 44TB fits on a single hard drive — and added attention mechanism content. Works in browsers without any dependencies.
Why does it matter?
High-quality visual LLM explainers remain rare. The 179-point HN score shows substantial practitioner interest despite mixed feedback about AI-generated content quality. The transparency about the Claude Code workflow and active community iteration make it a real-world showcase of AI-assisted educational content creation.
Who is it for?
Developers and students new to LLMs seeking an accessible visual foundation
Try it
ynarwal.github.io/how-llms-work — no install needed, works in any browser