Why do we assume only stationary computers - or even useful computers - would be the most likely to independently develop A.I.?