Jalex Stark comments on Why GPT wants to mesa-optimize & how we might change this