Re: Implementation

From: Hans Moravec <hpm.domain.name.hidden>
Date: Mon, 26 Jul 1999 17:35:00 -0400

Russell Standish <R.Standish.domain.name.hidden>:
> ...
> In this context, a Turing machine can never be concious, because it
> follows a preprogrammed path, without free will. Note this is not
> the same as saying comp is false, unless you strictly define
> computers to be Turing machines. My suspicion is that adding a
> genuine random number generator to the machine may be sufficient to
> endow the architecture with free will
> ...

I don't accept any of this line of argument, since it is my stong view
that consciousness is an attribution that can be mapped onto anything,
even a clock, a rock or a teddy bear.

But, there are probably practical issues here, about what kinds of
interaction you can easily have with the thing, to guide you to useful
mappings out of an astronomically large set of alternative
consciousnesses you could attribute. A teddy bear at least has a
shape and facial features to guide you a little. With a rock any
mapping is about as good as any other.


A deterministic Turing machine with I/O, on the other hand, could
guide your attribution about as well as a human being. A strictly
deterministic program with I/O can have encoded in it an appropriate
response for every possible input history. You don't need a random
number generator, because external input selects the state and
responses from among the many possible paths. A Turing-test passing
version (or a robotic Turing test passer, which deals in sensory
and motor as well as textual I/O) will interact as richly and
appropriately as a human being, and is thus an equally appropriate
target for the attribution of a particular consciousness.

Now consider deterministic machines without I/O, that contain both
Turing-test-passing AI programs and simulated bodies and simulated
worlds for them to inhabit. The simulations probably contain
pseudo-random number generators, either explicitly in the algorithms,
or implicitly because they model chaotic physical processes like
flowing fluids and rolling dice. Unless the AI can perfectly model
the entire simulation that contains it faster than real time (usually
not possible), the inputs it experiences from the deterministic world
simulation will be just as surprising to it as the physical world
inputs are to the Turing machine of the last paragraph, and its
reponses will be just as rich.

Most of us attribute consciousness to people we only see and hear
interacting with other people (say on TV), on the basis of how they
react to what happens to them. The same attribution would naturally
apply if we observed a machine passing a real-world Turing test. The
same attribution would also be natural if we observed the life and
times of a deterministic Turing-test-passing AI living in a
deterministic simulated world. Most likely, unless we look under the
hood, we couldn't tell if it was deterministic or probabilistic, since
any finite stretch of pseudo-random numbers could just as well have
been produced by a nondeterministic random process.

So, deterministic machines can have just as much free will as you or
I. The key is that they don't know everything that's going on,
outside themselves or in, so often don't know what will happen next,
or how they will respond to it. Many-worlds may provide an
interesting additional "source" of ignorance, but limitations on what
a finite process can model already provide sufficient ignorance for
free will even in a fully deterministic framework.
Received on Mon Jul 26 1999 - 14:49:10 PDT