Why it matters
Reasoning models are a leading driver of runaway output-token bills right now — understanding the RL training that produces extended "thinking" behavior explains why reasoning mode costs more per call, not just that it does.
The tokenmaxxing angle
GRPO is the reinforcement-learning mechanism that teaches a model to spend more output tokens reasoning before answering — this session digs into the training internals behind the exact behavior that inflates reasoning-model API bills.
From the organizers
Group works through Sebastian Raschka's "Build a Reasoning Model (from Scratch)," covering Chapter 7, "Improving GRPO for reinforcement learning"; code-along uses github.com/rasbt/reasoning-from-scratch.