CourseTransformers: From Attention to Trained, Served and Interpreted Models · Module 10: Efficiency and Long Context · part 74 of 111
Part 74 · Module 10: Efficiency and Long Context

Lesson 4: Long context, what it really buys you

25 min read·13 Oct 2026
Module 10 · Lesson 4 · About 45 minutes · Needs: Module 4, Lesson 3 (context extension, loss by position, needle tests); Module 6's base recipe

The situation

Sol 431 at Aurora Station. Tariq Haddad, the Comms Officer, reports that long transmissions from Earth are being split: a bulletin of 2,048 characters arrives in pieces, and base reads only 128 characters at a time. His proposal: "Teach the model to read a whole bulletin at once." Kofi Mensah agrees to the experiment but sets the bar: "A long window is a claim. Show me the model uses what is far back, on more than one test."

The running example is transmission 0: the first 512-character validation window, the start of a split bulletin. This lesson measures what a longer read really gives base, and what retrieval gives instead.

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.