The one idea I will share with you is that extractive part — the very energy consuming part — you can think of its data as oil. This kind of extraction goes to some large refinery somewhere, but the local way to train the small models — we can think of the data as soil. The local community tends to it together, fine-tune it and continuously train it so that whenever there’s a bias or an error, the course correction is immediate instead of waiting for the energy-consuming run that would take half a year or something.