"The way I interpret it is you're getting in the way of the model's scaling by not giving them the proper input and output affordances for them to be able to do their thing. And so at both training and inference time, you're getting in the model's way. If you're not properly harnessing them, right, like they don't have access to the web, then even the smartest model is not going to be able to retrieve things from the internet, right? The way you think about it is like inserting new data into the model or giving the model more ingredients."