Re: AI Risks, was Re: Rulers and Famine in Poor Countries (was Obesity)

From: Adrian Tymes (wingcat@pacbell.net)
Date: Sun Mar 09 2003 - 19:46:36 MST

  • Next message: Spudboy100@aol.com: "Cleanest Car?"

    --- "Michael M. Butler" <mmb@spies.com> wrote:
    > Adrian Tymes <wingcat@pacbell.net> wrote:
    > > You have a *chance* of dying. You could, totally
    > by
    > > accident, wind up doing the right thing anyway.
    > This
    > > is not the same thing as guaranteed death.
    >
    > And the combinatorics could wind up resulting in a
    > likelihood smaller than
    > the quantity 1 over the number of seconds that have
    > elapsed since the Big
    > Bang. That's not the same as guaranteed death, but
    > it's the way to bet if
    > those are the odds. So are odds of 1 in 100.
    >
    > Adrian, just for my couriosity's sake, how do you
    > figure
    > the odds? I'm not just asking that question
    > rhetorically.

    To be honest, I don't. I haven't gone through the
    effort to make a hard statistical analysis, largely
    because most of the data this analysis would be based
    on is so fuzzy as to make the analysis near-useless
    for
    convincing anyone.

    What I have done, however, is draw parallels to my own
    knowledge and study of intelligence and power, and
    make
    the assumption that AIs of any import to the world
    would live in the same reality as I have. Which means
    that similar solutions for granting other people power
    work, for much the same reasons, if you make the
    assumption (which I believe most people would admit)
    that AIs are, essentially, "people" with certain
    enhancements (for instance, the ability to calculate
    much much faster, the ability to hold more points of
    data - though still a finite number, however large -
    in
    their minds, and so forth; the exact list can be
    argued).

    Which means that an AI could be sociopathic (i.e.,
    uninterested in mere human survival, or even opposed
    to
    it as in so many Hollywood flicks), greedy,
    solipsistic
    (i.e., refusing to believe that we humans are
    sentient,
    even if it were proven that we created it), and so
    forth, like any person. If that is the case, then
    planning on the case where you make an AI without
    those
    flaws is, itself, flawed: while you could never
    succeed
    in that goal, you *could* be pressured to declare
    success, pretend you have succeeded (and thus, do not
    need to defend against those flaws), and move on to
    grant the AI power. Which is a scenario for disaster.

    On the other hand, if you account for that and treat
    the AI like a human being, with potentials for flaws
    imagined and unimagined, then you can still set up
    systems where, no matter how intelligent the AI is, it
    can not magically outthink you and gain freedom until
    you, or someone, trusts it. And to gain your trust,
    it
    would have to emulate friendly (dare I even say
    Friendly?) practices. The only real data point we
    have
    for such scenarios is human behavior, and the human
    data shows that, the vast majority (easily over 99%)
    of
    the time, a human being who fundamentally acts
    friendly
    really is friendly. This trend does not go down with
    increasing intelligence; if anything, it goes up -
    which would seem that the odds of a seemingly-friendly
    AI actually being friendly would be even higher.

    Besides, any plan for dealing with and developing AIs
    must take into account the abilities of the humans
    developing it, else the plan can not be put into
    practice (though, as described above, fatally flawed
    versions of the plan can be; sadly, this happens all
    the time in politics and business). Humans do not
    always have the ability to thoroughly review and
    understand programs, especially nontrivial programs.
    (For this, I cite my own professional knowledge.) But
    they do have the ability to evaluate human-like beings
    as if they were humans (cite anthropormorphism
    throughout the ages).

    > Against that, I think a case can be made that, for
    > issues as hairy as this,
    > Bayes isn't applicable in a vacuum; we don't get to
    > refine our estimates
    > over multiple trials that include catastrophe--maybe
    > the human race does
    > (if
    > we luck out), maybe not, but probably not we here
    > assembled. Conflict
    > levels
    > of 11 (that's 10^11 casualties--hundreds of millions
    > dead) /or more/ also
    > surely change the weight.

    And what are the odds of such a conflict if we do
    *not*
    go this route? Neither way prevents or guarantees it;
    it just makes the outcome more or less likely, over
    whatever time span you wish to designate significant.

    > >> Do you have any ideas for dealing with this
    > besides
    > >> building FAI first? Because as far as I can tell,
    > humanity is in
    > >> serious, serious trouble.
    > >
    > > Yes. Build an AI that you can trust, even if it
    > > quickly goes beyond your control.
    > ...
    > > perspective), but if you can trust someone else to
    > be
    > > gifted with the capabilities that an AI would
    > have...
    >
    > I agree with what you're saying but (and perhaps
    > this is also what you're
    > saying) the only way I can see to apply the level of
    > trust I give to a
    > person is to guarantee the physical limitations and
    > the
    > developmental psychological ground of an actual
    > human.

    That's certainly one way to do it. In fact, one
    scenario I think rather likely to work - and to
    increase our chances of this being done in a way that
    guarantees the survival of most presently human
    sentient beings - is to find a way to augment human
    intelligence into AIs. Certain elites will gain this
    capability before most people; that can not feasably
    be
    avoided. But if this power can be allowed to spread
    rapidly and freely (despite the inevitable
    restrictions
    some of these early adopters, or non-adopter elites
    afraid of the power this could bring to non-elites,
    will impose), then IMO the only significant risk of
    human extinction would be to the mere biological
    shells
    we presently inhabit. And even a good number of those
    would likely survive, at least for a time, by uploads
    with a sentimental attachment to the past. (If
    nothing
    else. Mmaybe the first uploads will be upgraded
    organic circuitry, using some of the best life-support
    equipment for such things we have available today:
    already grown, living bodies, like the ones the brains
    came from in the first place.)

    > I wonder how likely this is, given the fraction of
    > people with technical
    > chops that would dismiss this out of hand as
    > [quasi]paranoid.
    > We internet-enable hot tubs, for goodness' sake. So
    > the first aspect
    > (physical limitations) is likely to be ignored in
    > some way by even the
    > "embodied consciousness" folks.

    You also have to consider their arguments in light of
    their likelihood of sucess. Intelligence, or at least
    that which we are trying to replicate here (whatever
    its true nature, and whether or not it really has a
    name as yet), is defined as a uniquely human trait (at
    the moment). Biomimicry has proven successful in many
    other fields, and I would argue that AI is no
    exception
    - if you do not define Eliza and similar chatbots as
    "successful". The most robust, adaptable, and (again,
    I would argue) true intelligence to date has been the
    very close mimicing of bugs' control algorithms, and
    there is every reason to believe that this approach
    will scale up at least to human levels (existence
    proof: it's done so with organic evolved products, and
    the main limitation to doing it in artificial products
    is one which Moore's Law looks like it will beat down
    in due time - assuming it holds long enough, and
    despite the hue and cry from certain sectors about its
    imminent failure, most of the evidence suggests that
    it
    will).

    All of which is to say, when true AIs finally do
    arrive, the odds are that they will have something
    like
    a human upbringing. And given that the odds favor
    this
    outcome as it is, actions (including strenuous debate)
    to prevent any other approach are of diminished value
    in achieving their approach, while their costs
    (including, for some actions, retarding potential
    advances along any approach to AI, even the preferred
    one) remain unaffected. Therefore, it is best to
    allow
    even the "likely to produce monsters" paths to
    succeed,
    such that the "likely to produce what we want" paths
    may benefit from the results.

    All of which is a long way to say: Don't Panic.



    This archive was generated by hypermail 2.1.5 : Sun Mar 09 2003 - 19:51:51 MST