The difference is, I want the most efficient design that can complete every physical task that I can. That design is probably humanoid.
Bipedal with rollerblades, but probably more than 2 arms. Problem though is how do you get training data for that?
In non-murder tasks, humans are already obsolete. Our remaining economic value is solely alignment arbitrage, eg solving CAPTCHAs, murder, submitting 'human' papers, etc.
Given abliterated open models, Stringman is coming for our last remaining job!