Write the evaluation harness before the model
Teams that build the test suite first ship slower for two weeks and faster for the following year. Here is the order we teach.
Mara Okonjo Journal
Notes from the teaching floor.
Written by the people who mark the assignments. No announcements, no roadmap posts.
A long context window is not memory
Larger windows changed what fits in a prompt. They did not give the system a memory, and conflating the two produces a specific class of bug.
Write the model documentation before you need it
Every team writes documentation after their first incident. The work is the same either way; only the circumstances get worse.