“And what deep seeded is called mixture of experts or MOOC designs, where if you have a 100 billion parameter model, when you run the inference, you actually run it. You don't need to access all 100 bi...”