If you’re curious about how large language models actually work, you’ll likely find that most resources are either too academic and dense, or too superficial, focusing only on high-level concepts. It’s rare to find content that clearly explains the entire process from start to finish.

Coincidentally, I came across the how-llms-work project, which presents the complete lifecycle of large models as a visual, interactive web page. The content is based on Andrej Karpathy’s classic lectures. From web data collection, tokenization, and neural network training, to inference, generation, and post-training alignment, each stage is illustrated with intuitive diagrams and explanations, allowing you to grasp the entire workflow in a single view.

GitHub: http://github.com/ynarwal/how-llms-work

It also covers advanced topics such as hallucinations, context windows, and Retrieval-Augmented Generation (RAG), helping us understand the model’s capability boundaries and practical application methods. If you want to understand the entire process of large models—from training to conversational interaction—without wading through academic papers, this tutorial is worth spending ten minutes on.

ynarwal/how-llms-work

Source: https://github.com/ynarwal/how-llms-work

How LLMs Actually Work

A visual, interactive guide to how large language models are built — from raw internet text to a conversational assistant.

Live site: https://ynarwal.github.io/how-llms-work/

Based on Andrej Karpathy’s Intro to Large Language Models (https://www.youtube.com/watch?v=zjkBMFhNj_g) lecture.

What’s inside

• Data Collection — how the web is scraped and filtered into training data (Common Crawl, FineWeb)
• Tokenization — how text is broken into subword tokens via Byte Pair Encoding (BPE)
• Neural Network Training — the loss function, gradient descent, and what a forward pass looks like
• Inference & Sampling — how the model generates text token by token, and how temperature works
• The Base Model — what a model knows after pre-training and what it can’t do yet
• Post-Training — RLHF, instruction tuning, and how a base model becomes an assistant
• LLM Psychology — hallucinations, context windows, and how to think about what models “know”
• RAG — retrieval-augmented generation: embeddings, vector search, and context injection
• Full Pipeline Summary — end-to-end visual of every stage

Files

HN discussion

Posted to Hacker News (https://news.ycombinator.com/item?id=47886517) and generated heated debate, mostly about it being LLM-generated. Fair point — but the content isn’t the AI’s. Every claim, figure, and framing is traced directly to Karpathy’s lecture, not hallucinated by a model.

Vibe check

The code and content in this repo is mostly LLM-generated (Claude via Claude Code). The ideas, direction, and editorial decisions are mine — the implementation was largely written by AI. The council fact-checker exists precisely because of this: automated content warrants automated verification.