Genius dogs

Nicholas Bostrom (bostrom@mail.ndirect.co.uk)
Sun, 5 Oct 1997 22:22:26 +0000


I'm writing a paper on superintelligence. One of the things that I
want to do is to spell out in some detail exactly why most of thinks
that we might very well have superintelligence within 30 or 40 years
or sooner, even in the absense of nanotech. Then I want to address
the problem of whether we can make any good guesses about what a
superintelligence might do, whether we will be able to control it,
etc.

Here's one fun idea that might elicit some interesting comments:

If we take a human brain and simply speed it up enough, will it be a
superintelligence? Would a dog brain be?

Any human of normal intelligence could function as a universal
Turing machine, if augmented with enough scrap paper, time and
patience. According to Church's thesis, Turing computability equals
mechanical computability, so what the brain does is Turing
computable. (I assume that Penrose's argument to the contrary is
wrong, and we disregard possible exceptions that have to do with the
feasibility of supertasks or the unavailability of enough matter in
the universe.) What a (finite) superintelligence does would also be
Turing computable, so a human scrap paper, time & patience), if
speeded up, could be a superintelligence, provided she run an
appropriate program. We don't know that this program could be made
short enough to be stored in human long-term memory, (else it would
have to be provided as an input on the scrap paper), but I believe it
could. My guess is that a universal intelligence algorithm could be
made so simple that it would be an easy task for a human to memorize
it. Perhaps it would suffice to specify some simple architectural
properties of a neural network, together with some fairly simple
learning rule -- something like combination of Backprop, Infomax and
Hebb's rule perhaps. But whether or not it could be made so concise,
by using some scrap paper input, I think that even dogs could be made
to perform on a genius level. Here's how:

Recipe for Cani da Vinci, genius dogs

Ingredients: 20 poodles, universal intelligence algorithm,
nanotechnology

1. Take a twenty poodles.

2. First we need to put the dogs in training. We teach them the basic
operations of a Turing machine: the inputs are given to the dogs as
color-coded light signals, the output is a push with one of their
paws. Poodles are clever dogs, and they should be able to learn this
in a few months.

3. When the learning task is completed, the dog's brains are uploaded
onto a virtual computer that runs on a Jupiter brain.

4. They are presented with input from a virtual tape ("scrap paper")
by having their (virtual) nervus opticus stimulated. The output from
each dog is defined by measuring the activity in their (virtual) motor
nerves.

5. The instructions the dogs have to perform are written on one side
of the tape, so that they don't need to remember them. The dogs just
(virtually) run back and forth between different parts of the tape,
carrying out simple instructions.

6. Here is why we need more than one dog: eventually a dog will get
bored with the task and begin making mistakes. But all twenty dogs
will not start to get bored at exactly the same time. Now, the output
of the total dog house is defined as the output that most of the
individual dogs gave for a the input in question. Any dog that gives
an output different from the total output gets a (virtual) electric
shock, so that it returns to the learned scheme.

7. The Turing machine that is simulated runs the universal
intelligence algorithm (or alternatively simulates the evolution of
the brain states of an uploaded human genius). Amazing feats of genius
are performed by this dog house.

-- I anticipate that Anders will suggest that superintelligences will
make entertainment out of shaping human organizations into
entities that perform intelligent tasks, just as we have fun by
watching circus animals behave. :-)

Nicholas Bostrom
http://www.hedweb.com/nickb