They probably want to train on GPT Builder + store rankings to be able to train an AI to effectively spin up new agentic AIs in response to whatever task it has

They probably want to train on GPT Builder + store rankings to be able to train an AI to effectively spin up new agentic AIs in response to whatever task it has in front of it. If you're familiar with the Global Workspace Theory of consciousness I think they're aiming for something similar to that, implemented in modern AI systems. They'd like data on what creating a new agent looks like, what using it looks like, and how effective different agents are. They'll get that data from people using GPT Builder, people using "GPTs" and their subsequent ratings/purchases, and the sales and ratings data from the GPT Store, respectively.