One book a week, turned into a month of content.

Vizuara Books just published a new book: "Build an LLM from Scratch". Every step, from raw text to a working GPT, is a hand-drawn figure. books.vizuara.ai/book/build-llm-from-scratch This one took a long time. It grew out of my Building LLMs from Scratch video series, and it ended up with 115 figures. I redrew the attention chapter a few times before the matrix shapes lined up cleanly with the code. 𝐖𝐡𝐚𝐭 𝐭𝐡𝐞 𝐛𝐨𝐨𝐤 𝐜𝐨𝐯𝐞𝐫𝐬: - Tokenization and embeddings, drawn as matrices you can trace - Attention, worked out shape by shape: a 512 by 64 query meets a 64 by 512 key transpose - The full transformer block: six matrix multiplications, assembled into GPT-2 - Pretraining, then finetuning for classification and for instructions Right now the whole Vizuara Books shelf is free for a month, no card. This book and 41 others. If this helps someone learning LLMs, a repost goes a long way. books.vizuara.ai/book/build-llm-from-scratch

Here is one chapter of "Build an LLM from Scratch", flipped through in a few seconds. Six figures, from the bird's-eye view of a GPT down to how the transformer block is put together. Every book on Vizuara Books is drawn this way. books.vizuara.ai/book/build-llm-from-scratch 𝐈𝐧𝐬𝐢𝐝𝐞 𝐭𝐡𝐢𝐬 𝐜𝐡𝐚𝐩𝐭𝐞𝐫: - Why generation is memory bound, not compute bound - The causal mask, drawn before any code - Shape in equals shape out, which is why blocks stack The whole shelf is free for a month right now, no card. If you know someone learning this, a repost helps them find it. books.vizuara.ai/book/build-llm-from-scratch