How LLMs Work vs MacMind

"How LLMs Work" is a free, browser-based interactive guide that walks through the complete pipeline for building large language models — from raw web scraping to RLHF-trained conversational assistant. Created by Yash Narwal and based on Andrej Karpathy's technical deep-dive lecture, it's been getting significant traction on Hacker News (214+ points) for turning dense ML theory into something genuinely accessible. The site covers data collection and deduplication, Byte Pair Encoding tokenization with a live demo, pre-training and next-token prediction, inference with a probability sampling simulator, post-training with RLHF, and RAG. Each section uses animated visualizations, clickable pipeline diagrams, and canvas-based graphics — not static explainer images. The progressive narrative structure follows a single piece of text through every stage of the pipeline, making abstract concepts concrete. In an era where everyone uses LLMs but few understand how they work, this kind of high-quality educational resource matters for a different reason than tools: it raises the baseline competency of the entire developer ecosystem. Better-informed builders ask better questions, make better design decisions, and push vendors toward more transparency. This is the kind of project the HN community rewards — and deserves the signal boost.

MacMind is a complete single-layer transformer — attention, positional encoding, backpropagation, and weight updates — implemented entirely in HyperTalk, the scripting language built into Apple HyperCard, running on a Mac SE/30 with an 8 MHz processor and 4 MB of RAM. It trains to learn the bit-reversal permutation fundamental to the Fast Fourier Transform, and in doing so, the attention mechanism independently discovers the Cooley-Tukey butterfly routing pattern — not because it was designed in, but because the gradient descent finds it. Every operation is visible and editable in HyperCard's stack interface. Weights persist between sessions in card fields. The project is a deliberate demonstration that the mathematical operations underlying modern AI — matrix multiplication, softmax, cross-entropy, backprop — are substrate-independent: they work identically on hardware from 1989 as on an H100 cluster today, just much slower. The HN thread was warmly received as a genuine educational artifact: seeing attention, positional encoding, and gradient descent laid bare in HyperTalk's English-like syntax strips away 35 years of abstraction and reveals what transformers actually are. For educators, students, and curious engineers, MacMind is an unusually effective explanation tool.