From: Dan Fabulich (dfabulich@warpmail.net)
Date: Mon Jun 16 2003 - 20:23:09 MDT
Robin Hanson wrote:
> I edited out more of the discussion above because it all comes down to
> one simple point. Yes, truth is instrumentally useful in various ways,
> but unless you assign it overwhelming primary importance, it is easy to
> find situations where you should be willing to give up the instrumental
> advantages of truth to obtain other things. Such as in marriage,
> promoting group loyalty, and much else. Can't you imagine any such
> situations?
I'd like to add another argument to Eliezer's "Just Be Rational" argument.
(This is apparently one of those rare occasions when intelligent people
remind me of Nancy Reagan.)
Robin, you suggest that you can think of a non-trivial number of
situations in which we should be willing to give up the instrumental
advantages of truth to obtain other things, perhaps even in day-to-day
life. I think that the sort of cases you might be considering fall into
three distinct classes: the wireheaded, the rare, and the heuristic.
The first category, the wireheaded, includes cases where you'll be happy
that your goal was met by simply believing that your goal was met, even
though it wasn't. I think this class has its defeat laid at its own feat:
if you are anything like rational, it is against your goals, whatever they
may be, to wirehead. Your goal is an objective aside from its own
satisfaction... to the extent that you do something else instead of
satisfying your goals, especially to the extent that you do something that
makes it substantially less likely that they would be satisfied, you
violate your own goals.
The second category, the rare, is the category Eliezer himself has
thoroughly maligned: it includes those cases which ethical philosophers
might consider but which would, in fact, be so rare that considering them
would have net negative utility. Of those, I think, no more need be said.
But another final category might include cases in which we act on
approximating models, where the fundamental objects, properties and
relations in those models may not actually exist in the real world, not
even as "compositions" of real lower level objects. We use heuristic
models like these all the time to do our jobs and cope with the world; we
often act on these models automatically or on reflex. What's more, it's
obvious how, given the actual structure of our minds, setting these
behaviors up as reflexes might sometimes be a good idea.
But in all of the cases of this kind, I take it that if a well-informed
person (that is, someone who knew of a particular more accurate but less
usable model) assumed one of these models, even acted on it by reflex,
that person should not, on the whole, really *believe* in the models as
fundamentally accurate or true. A well-informed person takes these false
approximations as tools. He or she need not make a leap of faith and
*agree* with these models in order to get the job done.
OK, so let's consider your actual objections now with these categories in
mind. You suggest, for example, that I might believe that my hypothetical
spouse is faithful, in order to maintain the happiness of my marriage.
(Of course, this is just a special case of group loyalty, where the group
is a small family.)
Well, there are really *two* ways in which that could be instrumentially
beneficial; these need to be teased apart in order to properly understand
the example. First, to the extent that you simply believe that your
spouse is faithful, you may believe yourself to have a successful
marriage, and may therefore be happy. Second, to the extent that you
*act* as if your spouse is faithful, you won't, for example, obviously
withhold trust from your spouse, which may (on the whole) benefit you and
your marriage.
The first instrumental "benefit" is just a special case of wireheading.
And it doesn't make any more sense here than it does in general. You
_want_ a successful marriage; you don't want to *believe* that you have a
successful marriage. So, on this dimension, which is separate and
orthogonal from the latter instrumental benefit of action, there is no
instrumental benefit in believing a falsehood.
But as we can clearly see by separating out the effects, the latter
benefit can be acquired simply by *acting as if* the spouse were faithful:
don't accuse, don't divorce, don't act any differently at all, just go
about your life knowing what you know. You can argue as to whether that
may or may not be the right thing to do, but what you can't argue is
whether you could *do* it without going off the deep end and actually
*believing a known falsehood*.
Or could you? Certainly it could be the case that, despite your attempts
to be on your best behavior, your spouse might see through your act,
realize that you know, and could then ruin the marriage somehow. Sure,
it's possible, but this is a stereotypical case of the "rare" class that
Eliezer has already said so much against: what are the odds that, in order
for your goals to be accomplished, you not only need to *act* on an
incorrect model, but *believe* in that incorrect model, to get that last
0.1% of instrumental value? What are the odds that you'd do this by
mistake and get yourself into even more serious trouble? That's what we
mean when we talk about net negative benefit.
This is just one example, but I assert that the same applies to any and
all of the counter-examples you had in mind, especially if they're special
cases of group loyalty. Either you can get the benefit by "acting-as-if",
or you've fallen for wireheading, or your case is so rare that it's not
worth considering.
Hence, with the use of approximating models, we can get 99% of the value
of believing a falsehood without actually doing so. This bolsters
Eliezer's argument in the rarity of those "rare" cases (in which going off
the deep end and actually believing a known falsehood would be
beneficial); furthermore, it supports the claim that the rare cases really
do have trivially low pay-offs, especially when compared to the odds that
you might make a mistake, as well as the penalties that you might incur.
-Dan
-unless you love someone-
-nothing else makes any sense-
e.e. cummings
This archive was generated by hypermail 2.1.5 : Mon Jun 16 2003 - 20:33:07 MDT