Glad to see someone got it hahah. Yeah, that's something I thought about as well. It's a general "law" for any machine that the more specialized it is, the better it is at that task, and the more general it is, the worse it will be at any one given task, all else being equal.

I think the integration cost depends a lot on how interconnected the different capabilities must be. For example, it wouldn't be a huge leap to combine ChatGPT with Stable Diffusion or an image classifier since they use a common interface (language). But I don't know if that will be sufficient for AGI/ASI. I agree that it may turn out to be harder than expected and the performance would suffer as a consequence, good input!