Hacker News Reader

I believe the point is that the neural net will still have preconceptions -- but the random wiring just means you've chosen to not see what they are.

I think the point is that you can't remove external constraints from a situation simply by willing it (make the people in the room disappear, make your AI independent of assumptions).

For example the type of neural net chosen, the reinforcement and other parameters, the rules of the game chosen by the creator are all constraints which are not randomly chosen, so randomly choosing the starting state is not necessarily going to remove preconceptions built in to the model, and it is still choosing some starting state, just not an arbitrary one chosen by the creator.

I'm not sure it really applies to this article though, which is about a different problem - saving the time spent worrying about unimportant decisions or those where you don't have enough inputs and choosing to take a random path.

I think it means that:

(A) There are always "preconceptions" present, even if the programmer isn't working on a level of abstraction that lets them consciously realize it.

(B) Programming a system to be too naive can backfire.