The Synthetic Planner Trap

Senior AI Engineer interview at Google DeepMind, and the interviewer asks:

“You want to fine-tune a dedicated planner model, but all you have are raw agent action logs. Nobody ever wrote down plans. How do you build the training set, and what’s the most dangerous flaw in the naive version?”

Don’t say: “Feed each log to GPT, have it summarize the trajectory into a plan, then fine-tune on task → plan.”