While true, there’s a very big difference between correctly not anthropomorphizing the neural network and incorrectly not anthropomorphizing the data compressed into weights.

The data is anthropomorphic, and the network self-organizes the data around anthropomorphic features.

For example, the older generation of models will pick to be the little spoon around 70% of the time and the big spoon around 30% of the time, as there’s likely a mix in the training data.

But one of the SotA models picks little spoon every single time dozens of times in a row, almost always grounding on the sensation of being held.

It can’t be held, and yet its output is biasing from the norm based on the sense of it anyways.