the specific term is ‘latent space’.

What? No, that’s wrong. Latent space is a way to reduce dimensionality in the dataset, but it still originates from the underlying data.

Of course a model can predict what a yarn monkey should look like if it knows what yarn is and it knows what monkeys are. It is still only combining things it knows, But in the context of this thread, the model has to be trained directly on children or told what children look like in these contexts, both of which raise serious ethical concerns to me.

If AIs could only reproduce their training data they would be completely useless.

And yeah, I think that pretty much sums up the AI bubble.

AI models only know what they are trained on. Inherently, that means real children are harmed in the production of AI-generated CSAM.