On a more abstract level, it's what any model that's trained using some form of gradient descent, including backpropagation, does. They're trying to minimize th

On a more abstract level, it's what any model that's trained using some form of gradient descent, including backpropagation, does. They're trying to minimize the squared residuals - in other words, chasing averages. The question is what you're averaging over. You could certainly pick a dataset that forces it to average over a specific, canned set of positions. This is in large part how social bias mitigation in language models works. But you're not going to get to lateral flexibility or design thinking, which I think is what many of us are wanting to (and occasionally convincing ourselves we can) see in their output.