“Right one” is a relative concept. I would like to recreate my grandfather, but my knowledge of him would probably only fill a page of text or so. Thus it is much much easier to recreate a being that passes my personal grandfather turing test than it would be to create a being that satisfies my internal model of my father (of whom I know much more).
The recreations only have to be accurate to the point of complete consistency with surviving evidence.
On the neural net issue, the exact weights certainly don’t matter so much. There’s massive redundancy at multiple levels. Every one has V1 layers which are specifically unique, but they all functionally do almost exactly the same thing.
“Right one” is a relative concept. I would like to recreate my grandfather, but my knowledge of him would probably only fill a page of text or so. Thus it is much much easier to recreate a being that passes my personal grandfather turing test than it would be to create a being that satisfies my internal model of my father (of whom I know much more).
The recreations only have to be accurate to the point of complete consistency with surviving evidence.
On the neural net issue, the exact weights certainly don’t matter so much. There’s massive redundancy at multiple levels. Every one has V1 layers which are specifically unique, but they all functionally do almost exactly the same thing.