CourseTransformers: From Attention to Trained, Served and Interpreted Models · Module 10: Efficiency and Long Context · part 72 of 111
Part 72 · Module 10: Efficiency and Long Context

Lesson 2: Linear attention, low rank, and what they cannot learn

28 min read·13 Oct 2026
Module 10 · Lesson 2 · About 50 minutes · Needs: Lesson 1 (what attention costs as context grows); Module 2 (attention by hand); Module 6 (the base recipe)

The situation

Sol 431 at Aurora Station. Tariq Haddad, the Comms Officer, brings a relayed assignment sheet: six crew, their places and systems, then "Question: which system does Greta Mensah look after?" It is 442 characters long. The base model reads 128 at a time, so the link software splits it, and Greta's line ends up in a different window from the question.

Tariq has read that some attention variants cost only linear time, and asks whether one would let base read the whole sheet at once. Kofi Mensah adds his condition: measure the quality cost, with seeds.

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.