Re: Thinking about the future...

Vv&!&   Fx VzivzvxvvvvP pf B7D F| V~ Puiv 04 Sep 95

Robin Hanson wrote:
>But realistically, any one AI probably won't be too far
>ahead of any other AI, so they can police each other.

The issues are complex. To begin with, the claim that it is
improbable that one AI would be far ahead of any other AI
can be challenged. Suppose there are possible breakthroughs
to be made in computer technology. Once an AI became
sufficiently intelligent, it could think out a radical
improvement to its design; which would make it more
intelligent, allowing it to accelerate further; and so on.
(I think this is Dan Clemmensen's view.) For example, the AI
might be the first to make efficient use of nanotechnology.
If nanotech has such potentials as Drexler thinks, access to
a nanotech laboratory would be all the AI would need in
order to take off. The contest would be over before anyone
except the AI had realised it had begun.

In the slower scenario I depicted in the last letter,
however, it is unlikely that the AI would be alone. Its
discoveries would be diffused and employed to build other

But even in this case it is dubious that these other AIs
would be an effective protection against possible malicious
intentions of the first AI. They would, in effect, be its
offspring, and could therefore be expected to inherit some
of its basic properties, even values. If the first >AI is
bad, it might influence the development of computers so that
among subsequent machines a specific value set would be
prevalent which would allow them to form a tacit consensus
concerning the desirability of finally getting rid of the
human pest.

It is also worth considering what would make the grandpa >AI
bad in the first place:

1) Accident, misprogramming.

2) Constructed by a bad group of humans, for military or
commercial purposes. This group is presumably very powerful
if they are the first to build an >AI. The success of this
enterprise will make them even more powerful. Thus the
values present in the group (community, company, state,
culture) that makes the first >AI will not unlikely be the
value set which is programmed into subsequent >AIs as well.

3) Moral convergence. Sufficiently intelligent beings might
tend to converge in what they value, possibly because values
are in some relevant sense "objective". They just see the
truth that to annihilate humans is for the best. (In this
case we should perhaps let it happen?)

Notice that in case (2) and (3) policing would not work even
in the slow scenario and even if the first >AI itself had
little influence on the construction of other >AIs.

Nicholas Bostrom