CourseTransformers: From Attention to Trained, Served and Interpreted Models · Module 7: Architecture Families and Design Choices · part 50 of 111
Part 50 · Module 7: Architecture Families and Design Choices

Lesson 1: Three shapes of transformer, one block

24 min read·13 Oct 2026
Module 7 · Lesson 1 · About 55 minutes · Needs: Module 6, Lesson 1 (the recipe, lr_factor, the proxy); Module 2, Lessons 4 and 5 (masks, cross-attention)

The situation

Sol 418 at Aurora Station. Maya Lindqvist is planning the log cards: a log in, a four-field card out. "Do we need a model that reads the whole log first and then writes? Or is base enough?" Kofi Mensah wants seeds; Hana Aoki wants the cost per answer on the station's CPUs.

This lesson builds the three shapes of transformer from the package's Block, trains them on one shared task and measures quality and cost. The running example is one held-out row of the reverse task.

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.