Course II — Training as Data and Experimental Engineering

Central question

How does a team turn an uncertain research idea into a measurable, repeatable sequence of training decisions?

Module 1 — Building Training Data

1. What a Training Corpus Looks Like
2. Collection and Extraction Pipelines
3. Quality, Filtering, and Deduplication
4. Data Mixtures and Sampling

Module 2 — Designing Training Experiments

1. A Training Run as an Experiment
2. Evaluation Before Generation
3. Ablations, Ladders, and Scaling Laws
4. Reading Training Curves

Module 3 — The Training Lifecycle

1. Pre-training and Mid-training
2. Supervised Fine-tuning
3. Preference and Reinforcement Learning
4. Distillation and Self-distillation

Module 4 — Case Study: The Hill-Climbing Machine

1. The MAI-Thinking-1 Development Loop
2. Rank Non-invariance of Data Mixtures
3. Scientific Claims and Evidence Boundaries

Course synthesis

Review one model-development decision as an experiment, including its hypothesis, controls, measurements, scaling risk, and evidence quality.