On Sun, 24 Jan 1999, Wei Dai wrote:
> On Sat, Jan 23, 1999 at 03:19:59PM -0500, Jacques M Mallah wrote:
> > It's an interesting idea, very similar to one I had and rejected a
> > while ago. But I see several problems with it.
>
> Perhaps you were too hasty to reject the idea. I think all of the problems
> can be solved.
>
> > First of all, I don't think a string can give rise to
> > consciousness. It doesn't seem to have the right sort of structure. To
> > some degree it's a matter of personal intuition, and some people have the
> > same kind of problem with computationalism as dealing only with math and
> > not being able to give rise to consciousness. While I can convince myself
> > that a computation, with its decision making and counterfactuals, can do
> > the trick, I find it harder to believe that a mere string can.
>
> It's a central insight of computer science that any discrete structure can
> be encoded as a binary string. You haven't told me what your definition of
> "a computation" is yet, but if it is a discrete structure than it too can
> be encoded as a binary string. Therefore the real question is not whether
> a string can give rise to consciousness, but rather whether "a
> computation" (whatever that means in this context) can give rise to
> consciousness and whether a brain/AI state can give rise to consciousness.
> As far as I can see there is no reason why a brain/AI state cannot give
> rise to consciousness.
I basically have told you the definition, as well as referred you
to my web page for a more detailed discussion.
A computation involves a set of states, labeled by one or more
indices, with state transition rules between them. A trace of the
computation occurs when it starts out in one state, transits to the next
state indicated by the rules, and so on. A physical system implements
such a trace when there is a legitimate mapping from the physical states
to the formal states, such that the states mapped to perform the trace.
The unsolved problem is which mappings are legitimate and reflect the
information in the physical system, and which are illegitimate and
artificially introduce too much information.
What a string lacks is the state transition laws. There is no way
that what the concept of 'if ... then' refers to can be a property of a
string. A string just is what it is; counterfactuals have no meaning for
it. A physical system has not only functions like x(t) and y(x), it has
laws of nature such as x(t+1) = y(x(t)), and computations reflect that
very important aspect of nature, dynamics. Note that a law must be a true
equation, but a true equation is just a coincidence, a mere fact (even if
it holds for all x,y,t) unless it is a law. Laws are prescriptive, not
descriptive, unlike information; they make decisions.
> > Second, I am not at all convinced that P1 would have the structure
> > that you mentioned. The second part - finding the brain state - would
> > require a huge amount of code. It might be more efficient to create a
> > bunch of neural nets, let them interact and breed for a while, and print
> > out their state once they show some sign of having evolved an
> > intelligence, for example.
>
> I don't think that would work. The subroutine for recognizing intelligence
> must itself be intelligent, otherwise you will only get a bunch of
> neural nets that evolved/learned to fool the detection subroutine (for
> example by spitting out nonsense sentences that are grammatically correct
> or otherwise satisfies whatever criteria the detection subroutine is
> checking).
I doubt that. Remember, they are printed out at the end, but
having that criterion would confer no evolutionary advantage in the
meantime. Intelligence would. I think it would be very likely that some
criterion could be selected that would be both easy to recognize and
unlikely to occur before the evolution of intelligence.
> On the other hand the state extraction subroutine I mentioned is probably
> shorter than you think. The first part of this subroutine would
> contain the space-time coordinates of the computer running the AI, the
> second part would use some simple pattern recognition to group the
> subatomic particles in that space-time region into atoms, molecules, and
> memory elements, then the third part would "read" the memory elements and
> output the AI's state.
I think that would have to be a huge program. Huge. Each step -
Plank scale to low energy, wave functional to particles, types of atoms,
molecules, neurons and their connections, etc. - huge. I mean, you try
writing that algorithm. How long would it take you? If not long, then
show me the money.
> > I don't think that having larger memory elements would really make
> > the algorithm much shorter. More important would be to have a regular
> > arrangement of memory elements - with the disturbing implication that an
> > artificial digital intelligence would have much more measure than a
> > messy human brain. Perhaps our own experience then argues against the
> > proposal.
>
> What makes you think messiness would make such a big difference? The state
> extraction subroutine for a human brain should be very similar to the one
> for the AI. After grouping the subatomic particles into neurons, it would
> just output the neural-net structure and its current firing pattern.
Structure labeled how? By connections and interaction strengths
among the neurons? For the AI, the pattern of spatial locations of the
momory elements would be much of the needed info. Not so the human.
> > Third, in order for such a distribution of strings to arise, it
> > must be the case that all of those programs are really in operation. You
> > call it the 'a priori distribution', but in fact it is a nontrivial
> > distribution, and nature surely does not know all of that information - it
> > must be calculated. The true default distribution would be to just let
> > each bit have an equal chance of being 0 or 1. That being the case,
> > strings that arise during such calculations ought to also be able to give
> > rise to consciousness.
>
> One justification for ignoring the intermediate strings that arise during
> computation is that their distribution, unlike the distribution of the
> final outputs, is highly dependent on the model of computation. Another
> justification is to think of all of the computations as occuring in a
> universe that doesn't end. Then all of the intermediate strings last for a
> finite amount of time, whereas the final outputs last for eternity.
> Therefore the intermediate strings contribute zero measure.
That's just one choice of model among the many possibilities. And
what about the computations that never terminate?
- - - - - - -
Jacques Mallah (jqm1584.domain.name.hidden)
Graduate Student / Many Worlder / Devil's Advocate
"I know what no one else knows" - 'Runaway Train', Soul Asylum
My URL:
http://pages.nyu.edu/~jqm1584/
Received on Mon Jan 25 1999 - 14:10:10 PST