Slightly metaphysical question: at what point to these multimodal vector representations start to qualify as "qualia"? They seem to exhibit roughly the same at

Slightly metaphysical question: at what point to these multimodal vector representations start to qualify as "qualia"? They seem to exhibit roughly the same attributes, including that the specific internal representation is besides the point: if you train two systems from scratch, even using the same code and training data, you'll get internal representations that are different in detail but that exhibit similar external behavior due to the stochastic nature of the training process. This is analogous to what we can determine about qualia in human brains in that the details of the neural structures involved differ from one individual to another, yet largely produce external agreement exhibited in behavior and communication.