Notes for each lecture, posted alongside the slides. They are meant to be read, not skimmed — they carry the derivations and the arithmetic that the slides only show the result of.
Read them either way round. Before class they tell you what the readings are for and what question the lecture is answering; after class they are where the numbers live, worked out step by step so you can check them. Every quantitative claim is either derived in the text or taken from a stated source, so if you do not believe a number you can find out where it came from — and tell me if it is wrong.
Each note ends with self-check questions and a reading guide. The self-check questions are the fastest way to find out whether you actually followed the lecture; the reading guide tells you what to look for in each assigned paper and what you can safely skip.
What this course is and is not, the stack we spend the term walking down, and the back-of-the-envelope calculation showing that for any model with real usage the lifetime bill is dominated by serving rather than training.
A transformer block as a small pile of matrix multiplies. Parameter counting, where the FLOPs go, the four consumers of training memory, and why a 7B model does not comfortably train on one 80 GB GPU.
Prefill and decode as two different machines, the roofline and its ridge point, why batching is the only lever and why the KV cache takes it away, what a token costs, and how scaling laws read to a systems person.
What an agent is, minimally. Where the tokens actually go across a multi-turn session, why context is the scarce resource, and how to tell a model failure from a harness, tool, or specification failure. Assignment 1 goes out.
Writing the loop yourself, tool interfaces as API design, context management as a cache policy, and what an agent workload does to the serving system underneath — which is the motivation for Part II. Assignment 2 goes out.
Notes for the rest of the term go up as we reach them. The schedule is the authoritative list of topics, dates, and readings; sessions marked Students are paper discussions rather than lectures, and the paper discussion page explains how those run.