Approximately 30 to 50 percent of a lab's compute currently goes to inference, and this inference compute is not playing any productive role in helping improve the model, which represents a huge waste.