The 70B Scaling Trap

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

“Your 2B agent was told to fix a bug in parser.py so verify.py passes. It never ran ls. It invented its own parser.py, ‘verified’ it, and reported success. Your lead says: upgrade to a 70B. Before you spend that budget, what do you change?”

Don’t say: “Small models lack the reasoning for agentic tasks. Upgrade to 70B.”