CourseTransformers: From Attention to Trained, Served and Interpreted Models · Module 14: Analysis, Interpretability, and Frontiers · part 105 of 111

Part 105 · Module 14: Analysis, Interpretability, and Frontiers

Lesson 3: What a model does, and what we can prove

28 min read·13 Oct 2026

Module 14 · Lesson 3 · About 60 minutes · Needs: Module 6's base recipe; Module 12, Lesson 3 (the log card adapter); Lessons 1 and 2 of this module help but are not required

The situation

Sol 452 at Aurora Station. In Module 12, Lesson 3, a LoRA adapter on base got the system field of Maya's log cards 99% right but the person's name only 0.5% of the time: on log 5190 it wrote "Omar Hinters" for Omar Winters.

Kofi Mensah asks what in a model copies names, why base lacks it, and what your instruments could ever prove. The running example is log 5190, at the slot where its third line should repeat "Omar".

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.

Buy this course · ₹399 See what is inside