Re: Why Does Self-Discovery Require a Journey?

From: Eliezer S. Yudkowsky (sentience@pobox.com)
Date: Fri Jul 11 2003 - 00:50:41 MDT

  • Next message: Jeff Davis: "RE: HUMANS and low genetic complexity"

    Robin Hanson wrote:
    > At 10:02 PM 7/7/2003 -0400, Eliezer S. Yudkowsky wrote:
    >
    >> I am objecting to your phraseology here because it seems to
    >> preemptively settle the issue by identifying people's built-in
    >> emotional reinforcers as their real wants, while dismissing their
    >> cognitively held hopes and aspirations and personal philosophy as a
    >> foreign force interfering with their true selves. One could just as
    >> easily view the system from the opposite perspective.
    >
    > I do agree that this is a subtle question, whose answer is not
    > immediately obvious. The topic of self-deception can be a conceptual
    > morass, as our usual anchors are not as available. Nevertheless, I do
    > want to argue for the claim you find questionable.
    >
    > If people have contradictory beliefs, how can we say which ones are the
    > "real" beliefs? By reference to the basic schema of self-deception,
    > in which the real beliefs tend to determine less visible actions with
    > more fundamental consequences, and the false beliefs tend to determine
    > what we tell others and ourselves about ourselves, and the most
    > socially visible actions with the least fundamental consequences.

    When you deal with a human, you are dealing with at least three
    intertwined goal-ish systems.

    There's the hardware wiring that determines which thoughts and experiences
    are pleasurably reinforced, or painful. This wiring is by no means
    simple; it has more modes of action than just "pleasure" or "pain",
    despite that usual division. Still, it overall tends to shift people's
    thoughts in a goal-ish way.

    There are people's declaratively held purposes and goals.

    And there is evolution, which is also a goal-ish system, and which acts to
    select among {heriditary information in the initial conditions of the
    other two systems} based solely on the correlation between {the ultimate
    outcome of the interaction of thought and emotion} and the number of
    surviving grandchildren.

    If you go around determining real beliefs by the *outcomes* of people's
    actions, you run the risk of confusing evolutionary motives with cognitive
    ones, the classic mistake in evolutionary psychology. Every atom of
    complex hereditary information in humans was constructed by evolution
    based on the sole and only criterion of reproductive success, yet the
    explicit wish for children is only one among hundreds of explicitly
    represented cognitive reinforcers, woven by memes and philosophies into
    individually unique personal philosophies.

    > People may want to produce art to gain social approval, wealth, mates,
    > etc., but want to be thought of as doing it just for the art. People
    > may want to advocate positions that make them seem clever and
    > compassionate, and get them social accepted by the right folks, but
    > want to be thought of as wanting only to tell the truth. People may
    > want to be unfair when serving as a neutral judge, but want to thought
    > of as fair. These examples should be familiar to everyone; is there
    > anyone here to whom these are "news"?

    Aren't these instances of the classic error?

    People have emotional hardware and cognitive representations leading them
    to be devoted to art for its own sake, because people who, in the past,
    possessed the hardware for art, or fell into the philosophical attractor
    for personal philosophies exalting art as the result of their innate
    tendencies and biases, gained social approval, wealth, and mates. And
    even these are proximal goals, don't forget; all evolution cared about was
    the grandchildren.

    People have emotions leading them to honestly advocate positions that
    people applaud as clever and compassionate, because their ancestors who
    did the same were socially accepted by the right folks. They say they're
    just telling the truth, and they *are* just telling the truth, because in
    the evolutionary arms race of liars and lie-detectors, it's easier and
    more reliable to deceive the phenotype you're building than to have your
    phenotype deliberately deceive other genomes' phenotypes. People think
    they're honest and they are, but what think is the truth is output by
    biased reasoning hardware that was constructed according to the sole and
    only criterion of the number of surviving grandchildren to which that
    hardware's outputs led.

    Over and over, people seize power "for the good of the community". Are
    they being self-deceptive? No, they are being evolutionary deceived. The
    universality of this motif suggests to me that it is the result of
    selection over evolutionary time among imperfectly deceptive social
    organisms who argued linguistically about each other's motives in adaptive
    political contexts. There is a warp applied between the selection of
    goals and the selection of subgoals, because both operations are carried
    out by evolutionarily constructed, emotionally influenced brainware. I
    hypothesize (this is not a generally accepted truth in evolutionary
    psychology, as far as I know) the following evolutionarily constructed
    template of cognitive, emotional, and socioenvironmental interactions,
    which was statistically likely to lead to reproduction in the ancestral
    environment, and which today has become a famous motif in large-scale
    human events.

    1. People (emotional hardware) care about the community they live in.
    (See "Any Animal Whatever" by Flack and de Waal for an argument that
    "community concern can be observed in primates.)

    2. People (emotional hardware) react to perceived abuses of power by the
    tribal chief with righteous indignation. Note that the evolutionary force
    underlying the rise of this emotional perception rests on the possibility
    of installing a less abusive tribal chief *or* the possibility of
    achieving higher social status in the post-revolutionary order, *not* the
    factual degree of the tribal chief's abuses, nor the degree to which
    opposing the tribal chief benefits the *community*.

    3. People who are reacting to perceived abuses of power by the tribal
    chief with righteous indignation, and see the opportunity to personally
    overthrow the tribal chief or do so with a clique of friends, are biased
    to see this as a *good subgoal* of serving the community. This is the
    first appearance of an explicit rationalization warp in the hypothesis.

    So what's a rationalization warp? Okay, as all good decision theorists
    know, you get the desirability D(a) of an action A by summing over the
    utility times the probability, U(x)P(x|a), for all x of interest (i.e.,
    that the system can afford to compute). What I'm saying is that in this
    case, the adaptive bias is being applied to the computation p(x|a) rather
    than U(x). It's not that people "really want" to take over the tribe.
    They really want to promote the good of the community, and they really
    believe that they can do so by taking over the tribe. It would not even
    be accurate to say that the people are being deceived about their "real
    motives"; they are being deceived about which means correspond to which
    ends. In other words, the rationalization warp looks like this:

    Evolutionary end, i.e., subgoal of reproduction: Y. (Status, power...)
    Cognitively held end which is socially acceptable: X. (Good of the tribe.)
    So evolution is applying a bias to the computation of p(x|a) such that
    people find A to appear very plausible as a subgoal of X, given that it is
    *actually* a subgoal of Y. In other words, p(x|a) will be computed as
    higher than it should be, given that p(y|a) is *in fact* high.

    This is not being carried out by a deliberative process that evaluates
    p(y|a) before evaluating p(x|a), any more than people stop to think about
    their number of surviving grandchildren before eating tasty berries or
    tasty candy bars. There may be emotional ties from statistical correlates
    of p(y|a) to the evaluation of p(x|a), and so on, but there's not a
    deliberate deception involved. It's an emergent outcome of evolutionary
    selection on hereditary information that biases p(x|a) and correlates to
    reproductive success after *everything*, the whole deal, is finished.

    The end result of this is that acting on the evolutionary biased
    computation U(x)p(x|a) tends to actually maximize Y instead of X,
    providing that all goes as planned and you are in the ancestral
    environment. Stalin, as far as I know, did not have a spectacularly large
    number of surviving grandchildren - was no great success from an
    evolutionary standpoint. Had he just been in one tribe, though, less
    educated and literate, with less grand plans, he probably would have done
    better from a reproductive standpoint. The point is that this is the
    machinery that sent Stalin bad. And Robespierre. And... so on. See here
    the roots of the human species; it is still better than not caring at all.

    Step 4, of course, is another rationalization warp to believing that,
    having taken power, people are then emotionally and cognitively biased to
    find plausible the proposition that power must be concentrated into their
    own hands, again for the good of the community.

    That this template has been carried out successfully so frequently in our
    own, literate, suspicious times, to say nothing of a tribe of
    non-timebinding hunter-gatherers, suggests that it has been operating as a
    successful strategy over evolutionary time. Evolution constructs
    phenotypes to believe that they are acting for the good of the tribe,
    because that is what wins public support.

    But the point is that people *really do want* to help others, to create
    art, to be compassionate. It's the whole reason why we find the
    evolutionary puppet strings so horrifying once we become aware of them;
    evolution didn't plan for that, any more than it planned for the
    introduction of contraceptives. Ultimately DNA is just an
    information-theoretic history of *who did in fact* reproduce, not a plan
    to make organisms reproduce in the future.

    We genuinely *are* altruists. Tainted by shadow, perhaps, but true
    altruists nonetheless. Evolution, acting purely on the criterion of
    covariance of hereditary information in dynamic cognitive processes with
    reproductive success, managed to construct people who truly care about
    each other, as ends in themselves, and not as means.

    I have often thought that the rationalization warp is the greatest mixed
    blessing in the entire history of the universe.

    -- 
    Eliezer S. Yudkowsky                          http://singinst.org/
    Research Fellow, Singularity Institute for Artificial Intelligence
    


    This archive was generated by hypermail 2.1.5 : Fri Jul 11 2003 - 01:03:14 MDT