Working draft
Frontier LLMs, From the Inside
How the largest language models are built, trained, aligned, served, and used. For engineers.
The map
Every chapter points back here. Read left to right: the machine, then how a frontier lab makes one, then how it becomes an assistant, then how it is served and built upon, then what lies beyond text.
1 · The machine
ch 1–7
next-token game vectors, matrices attention, block learning, tokens train vs infer
2 · Pretraining
ch 8–13
scaling laws corpus, hardware parallelism architecture the run
3 · Post-training
ch 14–18
SFT preferences, RL reasoning alignment, safety evaluation
4 · Serving
ch 19–24
inference at scale prompting, tools retrieval, agents app evaluation production
5 · Beyond
ch 25–27
multimodal interpretability big to small
Lab spine · Beacon
one imagined frontier model, followed from raw data to deployed API; reasoned about, never run
Builder spine · Dispatch
an on-call assistant for Postbox, grown on the frontier API, one layer per chapter
Where does each chapter sit, and what runs through all of them? Five parts in reading order, and two threads that never stop: the model the Lab builds, and the product you build on it.
Contents
Part 1 · The machine: from numbers to next token
The next-token game Vectors and the functions that move them Attention The transformer block Learning Tokenization Training vs inference
Part 2 · Making a frontier model: pretraining at scale
Why bigger works The corpus The hardware Training across thousands of GPUs Frontier architecture choices The run Part 3 · From base model to assistant
From base model to assistant Learning from preferences Reinforcement learning and reasoning Alignment and safety Evaluating a model Part 4 · Serving and building on frontier models
Inference at scale Prompting and context as engineering Retrieval Agents Evaluating an application Operating in production Part 5 · The wider frontier
Multimodal models Looking inside Big to small Closing and appendices