Week 8

Measure, then climb

Wednesday, November 18, 11:00–11:50 a.m., Slichter 3853

Joining remotely? The Zoom link is on the home page.

A week of concepts, with no code to write. How does anyone know whether an AI is any good? We start with the AI eval slides we didn't reach in week 6: how a model is trained, and why the check it trains against is the same object as the test that measures it. Then what a benchmark is and how it wears out, what happens when people push a score up by trial and error, and what it means for an AI to do that pushing itself.

In class

Five ideas, each built on the one before. Each card opens the slides at the start of that part.

  1. Part 1RL environmentsWhere a model's ability comes from: pre-training, post-training, and practice against an automatic check in a sandbox.
  2. Part 2AI evalThe same task with the weights frozen: five ways to measure a model, who does the measuring, error bars, and how to read a score.
  3. Part 3BenchmarkingA fixed eval that everyone runs the same way, how benchmarks wear out, and your own tasks as one.
  4. Part 4Hill climbingChange, measure, keep: the loop behind training, harness tuning and leaderboards, and the ways it goes wrong.
  5. Part 5Self-improvementRecursive self-improvement: when the AI does the climbing, what is real so far, and what limits it.

Further reading

The details and sources behind the slides, including the AI eval reading that moved here from week 6.

Homework

Homework 6, make your task hold, is due before this class. Open your pull request and we run GPT-5.6 Sol on it. You're encouraged to take your task on to Terminal-Bench-Science by December 15.