Issue 22, Oct 11, 2026
The Synthetic Bug Trap
Senior AI Engineer interview at Google DeepMind, and the interviewer asks:
“You need 50,000 runnable environments to RL-train a repo-level bug-fixing agent. Mining real pull requests gets you 3,000. Injecting bugs into healthy repos gets you unlimited. What’s the catch, and how do you combine the two?”
Don’t say: “Synthetic data is lower quality, so mix in some real data. Maybe 80/20.”
Why training RL coding agents on mutated code quietly creates shortcut learning, and how to calibrate synthetic environments so real-world resolve rates actually scale.
Share this trapShare on LinkedIn
The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.