This is the same way LLMs are trained. Are they benchmaxxing and useless out of domain?