What a Language Model Actually Is
From Illustrated LLM Engineering
Who it is for: Developers who can call an API and want to build more dependable LLM applications.
Working manuscript · 2027–2028 Edition. The full book is forthcoming; text and illustrations may change.
Otto will draft support replies, retrieve documents, and request tools. This chapter isolates one operation used during text generation: given the available context, a model assigns scores to possible next tokens. A decoding procedure selects a token, appends it, and repeats.
That account is useful but deliberately incomplete. Deployed models may accept images or audio, perform hidden or exposed reasoning, and use provider-specific processing around generation. The application still receives an autoregressively generated token sequence, but the visible reply is not necessarily the model's only computation.
Read the chapter
The preview below shows the first 3 pages. The PDF contains the complete 9-page chapter, with selectable text and the original illustrations.
Download sample chapter (PDF) ↓Open PDF in a new tab ↗
Keep exploring Illustrated LLM Engineering
Build a support assistant that can look up information and use tools. Learn how tokens, embeddings, and retrieval work, and test the answers it gives.
See the book outline →Questions or corrections? support@hackerbrewery.com