“The problem of teaching an AI to use a computer via pixels and actions, as explored in the 2016 World of Bits project, is now beginning to be solved using modern LLMs with reinforcement learning.”