Book I From Models to Custom Silicon
How Modern AI Models Work
This book gives a capable software engineer the mental models and practical skills needed to inspect, trace, compare, and explain modern AI model artifacts.
Every chapter can be read or watched as short lessons, and each way teaches the whole chapter. 3 of 23 chapters are published; the rest are coming.
Part I Build the model vocabulary
- B1C011What a Model Actually Is
When you send a prompt, a tokenizer turns your text into token IDs, the model gives every possible next token a score, softmax turns the scores into chances, one token is picked and appended, and that repeats until a stop token.
- B1C022Tensors, Weights, Layers, and State
Every number in a running model sits in an array; some arrays the model learned and only reads, some are made during one call and dropped, and one kind is kept from call to call while a reply is written.
- B1C033Linear Algebra Inside a Model
One piece of arithmetic runs over and over inside the model, multiplying two lists of numbers pair by pair and adding up the products, which scores how well they match, and Paris at 95 percent is the best of 65,536 such scores.
- B1C044How Model Numbers Are Representedcoming
- B1C055Scores, Probabilities, Losses, and Uncertaintycoming
- B1C066How Training Changes Weightscoming
Part II Follow information through the model
- B1C077From Input Text to Model Vectorscoming
- B1C088A Transformer Block from First Principlescoming
- B1C099How Autoregressive Generation Actually Runscoming
- B1C1010Beyond the Standard Transformercoming
Part III Inspect and compare real artifacts
- B1C1111How to Read a Real Model Artifactcoming
- B1C1212Language Heads, Decision Heads, and Typed Outputscoming
- B1C1313Model Cost: Compute, Memory, State, and Critical Pathcoming
- B1C1414How to Compare Models Without Fooling Yourselfcoming
Part IV Eight model deep dives
- B1C1515Needle3: A Model Designed Around Edge Constraintscoming
- B1C1616The LFM2.5 Family: Hybrid Models Across Sizes and Taskscoming
- B1C1717Qwen3.8-27B: Anatomy of a Large Multimodal Parentcoming
- B1C1818Ternary Bonsai 2 27B: What Changed from Qwen3.8-27B?coming
- B1C1919Laya: Decision Models Are Not Small Chatbotscoming
- B1C2020Qwen Image 2.1: Inside an Image-Generation and Editing Pipelinecoming
- B1C2121DiffusionGemma: Generation Without the Ordinary Left-to-Right Loopcoming
- B1C2222DeepSeek V4.1 Flash: Inspecting a Large Modern Serving Systemcoming
Part V Synthesis
- B1C2323Eight Models, One Inspection Methodcoming