AI ALIGNMENT FORUM
AF

1591
Wikitags
Main
3
LW Wiki

LW Wiki

Edited by Ruby, Eliezer Yudkowsky, et al. last updated 30th Dec 2024

Complexity of Value is the thesis that human values have high Kolmogorov complexity; that our preferences, the things we care about, cannot be summed by a few simple rules, or compressed. Fragility of value is the thesis that losing even a small part of the rules that make up our values could lead to results that most of us would now consider as unacceptable (just like dialing nine out of ten phone digits correctly does not connect you to a person 90% similar to your friend). For example, all of our values except novelty might yield a future full of individuals replaying only one optimal experience through all eternity.

Related: Ethics & Metaethics, Fun Theory, Preference, Wireheading

Many human choices can be compressed, by representing them by simple rules - the desire to survive produces innumerable actions and subgoals as we fulfill that desire. But people don't just want to survive - although you can compress many human activities to that desire, you cannot compress all of human existence into it. The human equivalents of a utility function, our terminal values, contain many different elements that are not strictly reducible to one another. William Frankena offered this list of things which many cultures and people seem to value (for their own sake rather than strictly for their external consequences):

Life, consciousness, and activity; health and strength; pleasures and satisfactions of all or certain kinds; happiness, beatitude, contentment, etc.; truth; knowledge and true opinions of various kinds, understanding, wisdom; beauty, harmony, proportion in objects contemplated; aesthetic experience; morally good dispositions or virtues; mutual affection, love, friendship, cooperation; just distribution of goods and evils; harmony and proportion in one's own life; power and experiences of achievement; self-expression; freedom; peace, security; adventure and novelty; and good reputation, honor, esteem, etc.

The "etc." at the end is the tricky part, because there may be a great many values not included on this list.

One hypothesis is that natural selection reifies selection pressures as psychological drives, which then continue to execute independently of any consequentialist reasoning in the organism. This may also continue without that organism explicitly representing, let alone caring about, the original evolutionary context. Under this view, we have no reason to expect these terminal values to be reducible to any one thing, or each other.

Taken in conjunction with another LessWrong claim, that all values are morally relevant, this would suggest that those philosophers who seek to do so are mistaken in trying to find cognitively tractable overarching principles of ethics. However, it is coherent to suppose that not all values are morally relevant, and that the morally relevant ones form a tractable subset.

Complexity of value also runs into underappreciation in the presence of bad metaethics. The local flavor of metaethics could be characterized as cognitivist, without implying "thick" notions of instrumental rationality; in other words, moral discourse can be about a coherent subject matter, without all possible minds and agents necessarily finding truths about that subject matter to be psychologically compelling. An expected paperclip maximizer doesn't disagree with you about morality any more than you disagree with it about "which action leads to the greatest number of expected paperclips", it is just constructed to find the latter subject matter psychologically compelling but not the former. Failure to appreciate that "But it's just paperclips! What a dumb goal! No sufficiently intelligent agent would pick such a dumb goal!" is a judgment carried out on a local brain that evaluates paperclips as inherently low-in-the-preference-ordering means that someone will expect all moral judgments to be automatically reproduced in a sufficiently intelligent agent, since, after all, they would not lack the intelligence to see that paperclips are so obviously inherently-low-in-the-preference-ordering. This is a particularly subtle species of anthropomorphism and mind projection fallacy.

Because the human brain very often fails to grasp all these difficulties involving our values, we tend to think building an awesome future is much less problematic than it really is. Fragility of value is relevant for building Friendly AI, because an AGI which does not respect human values is likely to create a world that we would consider devoid of value - not necessarily full of explicit attempts to be evil, but perhaps just a dull, boring loss.

As values are orthogonal with intelligence, they can freely vary no matter how intelligent and efficient an AGI is [1]. Since human / humane values have high Kolmogorov complexity, a random AGI is highly unlikely to maximize human / humane values. The fragility of value thesis implies that a poorly constructed AGI might e.g. turn us into blobs of perpetual orgasm. Because of this relevance the complexity and fragility of value is a major theme of Eliezer Yudkowsky's writings.

Wrongly designing the future because we wrongly encoded human values is a serious and difficult to assess type of Existential risk. "Touch too hard in the wrong dimension, and the physical representation of those values will shatter - and not come back, for there will be nothing left to want to bring it back. And the referent of those values - a worthwhile universe - would no longer have any physical reason to come into being. Let go of the steering wheel, and the Future crashes." [2]

Complexity of Value and AI

Complexity of value poses a problem for AI alignment. If you can't easily compress what humans want into a simple function that can be fed into a computer, it isn't easy to make a powerful AI that does things humans want and doesn't do things humans don't want. Value Learning attempts to address this problem.

Major posts

  • The Fun Theory Sequence describes some of the many complex considerations that determine what sort of happiness we most prefer to have - given that many of us would decline to just have an electrode planted in our pleasure centers.
  • Thou Art Godshatter describes the evolutionary psychology behind the complexity of human values - how they got to be complex, and why, given that origin, there is no reason in hindsight to expect them to be simple. We certainly are not built to maximize genetic fitness.
  • Not for the Sake of Happiness (Alone) tackles the Hollywood Rationality trope that "rational" preferences must reduce to selfish hedonism - caring strictly about personally experienced pleasure. An ideal Bayesian agent - implementing strict Bayesian decision theory - can have a utility function that ranges over anything, not just internal subjective experiences.
  • Fake Utility Functions describes the seeming fascination that many have with trying to compress morality down to a single principle. The sequence leading up to this post tries to explain the cognitive twists whereby people smuggle all of their complicated other preferences into their choice of exactly which acts they try to justify using their single principle; but if they were really following only that single principle, they would choose other acts to justify.

Other posts

  • Value is Fragile
  • The Hidden Complexity of Wishes
  • Fake Morality
  • Welcome to Heaven
  • Complexity of Value ≠ Complexity of Outcome
  • Not for the Sake of Pleasure Alone
  • A Non-Comprehensive List of Human Values

See also

  • Complex Value Systems are Required to Realize Valuable Futures
  • Human universal
  • Fake simplicity
  • Metaethics sequence
  • Fun theory
  • Magical categories
  • Friendly Artificial Intelligence
  • Preference
  • Wireheading
  • The utility function is not up for grabs
Subscribe
Discussion
Subscribe
Discussion
Posts tagged Complexity of value
45Shard Theory: An Overview
David Udell
3y
2
28But exactly how complex and fragile?
KatjaGrace
6y
30
27Review of 'But exactly how complex and fragile?'
TurnTrout
5y
0
31Disentangling arguments for the importance of AI safety
Richard_Ngo
7y
15
26The two-layer model of human values, and problems with synthesizing preferences
Kaj_Sotala
6y
3
29Alignment allows "nonrobust" decision-influences and doesn't require robust grading
TurnTrout
3y
31
26General alignment properties
TurnTrout
3y
2
19Conflating value alignment and intent alignment is causing confusion
Seth Herd
1y
0
25The Pointer Resolution Problem
Jozdien
2y
0
11Would I think for ten thousand years?
Stuart_Armstrong
7y
7
14Reversible changes: consider a bucket of water
Stuart_Armstrong
6y
15
57Evaluating the historical value misspecification argument
Matthew Barnett
2y
77
34Two Neglected Problems in Human-AI Safety
Wei Dai
7y
21
20Siren worlds and the perils of over-optimised search
Stuart_Armstrong
11y
0
23Three AI Safety Related Ideas
Wei Dai
7y
30
Load More (15/23)
Add Posts