Learning high-level task reasoning solely with a low-level action-output model would be extremely inefficient.

“you know try to learn that with a you know low-level model that's outputting actions but it'd be extremely inefficient so that is by having a hierarchical system like putting a harness like a agentic system around our word to action model is what allowed the model to compose different capabilities different skills to solve a much longer horizon task it's very much like in language model right like two years ago three years ago and the model was only good at autocomplete”