Why the Good Guys Will Usually Win

(In Real Life Too) ... a Sketch of an Argument

It has seemed to me for a very long time that there’s a clear conceptual argument why, on average across the whole multiverse, on the whole “good guys” are likely to win and open-mindedness is likely to prevail over closed-mindedness.

I even think this could be proven mathematically, though it would take a fair bit of work and I don’t have time to undertake such a project right now.   If this post inspires some clever mathematicians with more time on their hands for such things than I have, that would be wonderful.

What I’m going to do here is simply sketch out the conceptual argument as it’s been floating around in my mind over the decades.  Many of the points I’m going to make have been made by others in various different forms, but I haven’t seen anyone express these ideas in what I’ve felt was a sufficiently simple and clear way.  (Which may just mean that I have an eccentric idea about how such things should be be expressed, of course….)

From a math-y point of view, the arguments I have in mind are basically what would be called “counting arguments.”   Basically I’m going to explain why I think there are more ways for good guys to win than for bad guys to win, and more ways for open-mindedness to prevail than closed-mindedness.   Given the way multiverse theory works, this implies that there is a greater probability-mass of multiverse-branches in which good guys, and open-minded guys, win than in which the opposite happens.

While I’m using the language of “multiverse” here because it’s the way I think, actually any sufficiently large and diverse subset of a single universe can impersonate (im-universe-ate?) a multiverse for many purposes. The ideas presented here seem to apply perfectly well in this context, i.e. across a large and diverse subset of our universe, there seem to be more ways for good and open-minded guys to win than vice versa, so I would suspect that on the whole in this large and diverse subset we’d find more good than bad, just by natural selection. So anyway, if you don’t like “multiverse”, you can substitute “large diverse subset of universe” when you read the following and it will pretty much make just as much sense ;-)

Goodness and Openness Will Generally Prevail in our Specific Universe As Well

Now even if this is true, it doesn’t necessarily provide great reassurance about what will happen in our particular swath of the multiverse.   Maybe we live in an anomalously screwed-up region of the multiverse, which is primed for badness to prevail in a cosmically unusual way?  Probably most of us have felt that way sometimes….

Actually, though, I do think these cosmic multiverse-scale arguments also have some relevance in our everyday world.   A large heterogenous system like the Internet or the world economy winds up having some of the aspects of a multiverse — a lot of different possibilities occur with in it, gathering resources and rising and falling in a somewhat (far-from-equilibrium) statistical-mechanics-ish way.

The same reasoning that leads one to suspect good guys will generally prevail across the multiverse, leads one to believe that positive-minded collectives will generally have more influence across the Net than bands of self-centered jerks.  Both exist for sure, but the latter in the end will be less robust.   

For all the influence that the Facebooks and Googles and Tencent of the world have, in the end the Internet itself, the global scientific community and the open-source development community are all more robust and are even more influential.  

For all the military and nationalistic might that the US, Chinese and Russian governments (for example) display, the global network of trade and of intellectual and artistic and social collaboration has even more power and ultimately has proved more powerful than nationalistic urges, time and time again.

The differential long-term success that we see in these open, beneficially-oriented alternatives ultimately has the same root cause, I suggest, as the overall prevalence of goodness and openness across the multiverse: As the world throws one after another weird thing at you, the more robust ways to be are those that can cope successfully with the greatest number of potential situations (i.e. a “counting argument”).

Why the Good Guys Will Usually Win — Getting to the Meat of It

Having “told you what I’m gonna tell you”, let me now move on to actually telling you.   I.e. what’s my actual thrust of argument why the good guys will usually win?

(And it’s not just that I’ve watched too many Hollywood movies.  Actually my cinematic taste it way more perverse than that.   My favorite movie is probably “Brazil”, which has a somewhat less than favorable ending.  Nor Tarkovsky’s Solaris.  Though I did grow up as a “Lawrence of Arabia” fanatic, and in spite of Lawrence’s tragic personal end, his successes do somewhat illustrate the points I’m making here….)

My first (informal) lemma is: Given the way the physical universe works, powerful and capable systems are generally going to be “communities of agents” of some form.  

This is basically because of how space and time work.   There’s a limit to how much energy and information one can fit in a certain small region of space.  So powerful and capable systems ultimately will need to be spread out across space.  

However, Special Relativity limits the speed at which information can propagate across space.   It implies that highly powerful and capable systems will need to be distributed systems, with different parts doing their own things and then communicating them to each other.

My second informal lemma is: Being smarter will usually help you win (all else equal, on average, yadda yadda).   

This is basically Legg and Hutter’s definition of general intelligence — they conceive GI as the ability to maximize computable reward functions in computable environments, averaged across all such reward functions and environments.  However, the overall point is broader than their particular formalizations.

Sometimes being stronger helps you win, more so than being smarter.  Sometimes being dumber will even help you win, e.g. because it limits the ways in which you can be fooled.

But overall, averaged across a wide number of different situations, being smarter will generally let you figure out how to get by better than being dumber.  It can allow you to invent new tools to supplement your abilities or overcome your shortcomings, it can enable you to communicate with others in ways they can understand and rally them to you cause, etc.

Basically, in a rough and informal way, all I’m saying right here is that being smarter means being able to solve a wider class of problems effectively

My third informal lemma is: On the average, an agent system wherein the various agents largely trust each other will be smarter than one where the various agents largely distrust each other.

This gets at one simple aspect of being “good guys.”   A community of individual agents who are “prosocial” (i.e willing to prioritize community needs over individual needs fairly often) and who know each other are prosocial will, I believe, be able to effectively cope with a wider variety of situations than community of individuals who are mostly selfish and know each other are selfish. 

Of course I know there is way more to “goodness” than this. There are wild cultural variations in what is considered “good” even within the human species at one particular point in time. The core of “goodness” however is I believe compassion, and one thing that happens when you have a community of multiple agents who are mostly compassionate toward each other, is that they can see that they are living in a mostly compassionate community, and they begin to interact with each other in a spirit of mutual trust. This is the aspect of goodness I want to focus on here, because it’s one aspect of goodness that I believe helps a lot in terms of enabling communities of good guys to solve more problems better than communities of bad guys (who are not so much mutually trusting) and thus overall tend to prevail via natural selection dynamics.

My rough conclusion from putting these three lemmas together is: On the whole, prosocial, beneficially-oriented agent communities are going to do better and survive and flourish more than communities agents focused more selfishly on individual advantages or the advantages of small “tribal” groupings.

The Algorithmic Benefits of Prosociality

Why would this “third informal lemma” generally be true?  This is perhaps the subtlest pat of my sketched-out argument.

Basically: Any problem-solving process that a group of self-centered, relatively trustless individuals can carry out, can also be carried out by a prosocial group.  If the best way for the prosocial group to get certain things done is to each work independently for a while, they can do that.

On the other hand, there are some problem-solving processes that a prosocial community can do, but a community of selfish individuals cannot.   In a prosocial community, individuals can sometimes carry out processes that they don’t especially understand or see the purpose of, or that won’t bring them any near-term reward.   They will do this because they trust each other and trust that one another are mostly working toward the common good.

OTOH in a community of more selfish and trustless individuals, getting this sort of process carried out requires complex mechanisms of incentivization, reputation, validation, etc.   As we see in the modern money economy and various crypto-economic networks, this sort of thing entails a great deal of overhead.

Hofstadter’s notion of superrationality is one way of looking at some of the advantages achievable via a trusted group of collaborators — who trust each other, trust each other to trust each other, trust each other to trust each other to trust each other, and so forth. Hofstadter and others have shown that this sort of mutual trust can cut through various decision-theoretic dilemmas that otherwise occur in social systems, such as “tragedy of the commons” type situations. And the trust doesn’t need to be perfect to have this sort of effect — a strong enough probabilistically-confident trust can have similar effect in nearly all practical situations.

For some sorts of problem-solving processes, trustlessness doesn’t entail so much inefficiency.  For instance, if we assume NP doesn’t equal P, then there is a large class of problems for which it’s hard to find a solution but relatively easy to check if a proposed solution is correct or not.   In this case lack of trust isn’t such an impediment — mutually mistrustful individuals can go away and solve problems on their own, then present to each other and check each others’ solutions. 

However, some important classes of problems are not like this.   

For example: medium to long term processes of self-improvement or self-transformation.   Say, a methodology for adjusting one’s lifestyle and supplement and drug intake to maximize one’s lifespan.  Or a spiritual path, or scientific research program.  Or a generally winning strategy for a complex game.   These sorts of things are both hard to discover AND time-consuming and laborious to validate.  These sorts of things can spread much more successfully and rapidly among communities of mutually trusting agents.

Excessive trust can be maladaptive and lead to failures of robustness, or to excessive conformity.   However these pathologies can straightforwardly be overcome by pairing trust in the other agents’ intentions with rational assessment of their capabilities and knowledge. For example, operating a reputation system and then acting in a “trustless” manner only when dealing with low-reputation parties is still much less expensive than dealing in a trustless manner with everyone — and running the ML needed to ward off reputation-fraud reasonably well can be relatively inexpensive given the computational efficiency achievable via a distributed network of mutually trusting parties (identifiable e.g. by this same decentralized, collaboratively-maintained reputation system).

So the crux of the argument that good guys win has to be that: On the whole, a decent percentage of the problems that the world presents agent communities with, are problems that are most effectively solved via processes that are more efficiently runnable via groups pf mutually trusting agents.

The fact that processes of medium to long-term self-improvement fall into this category, implies that highly dynamic environments where the fundamental properties shift over time, will tend to be more effectively handled by groups of mutually trusting agents.  Because these agents will be able to guide each other regarding how to evolve and develop in the right way to keep up with the environment’s changes.

But getting back to counting arguments, there are generally going to be MORE highly dynamic environments than states environments … at least according to the most simple and intuitive (to humans) distributions over environment-space.

And this ties in with why open-mindedness will generally beat closed-mindedness.  Like “goodness”, open-mindedness is a broad concept that will be interpreted by various parties in various ways. Weaver’s thesis on Open-Ended Intelligence does a pretty good job of nailing the concept down insofar as it’s possible to nail it down. But what I need here from the concept is pretty simple: An open-minded system, relative to a closed-minded one, is more willing to change its way of doing things and even modify some of its core presuppositions in accordance with how its environment changes and what it learns about itself.

A closed-off mind that doesn’t like to evolve itself may do very well in a particular environment that matches its assumptions.  An open-ended intelligence that is oriented toward self-transformation as well as individuation will deal better with environments that constantly self-disrupt — but “how to be an open-ended mind” is the sort of knowledge that’s much more simply and conveniently exchanged among mutually trusting minds, because it’s not the sort of thing that comes with cheap and ready modes of objective validation.

The only potential loophole here would be if being prosocial intrinsically made each subsystem less computationally efficient / intelligent somehow.   But in fact the opposite seems to be true.  Being prosocial opens up each system to allowing the others to give it advice on how to improve itself, i.e. it allows cheap crowdsourcing of self-improvement and meta-learning.

The conclusion is, across the multiverse, in most universes where there has been enough stability for evolution to take place, the systems that have prevailed will have been systems of "good guys"

So the multiverse should mainly consist of: Chaotic universes, early-stage universes where prosocial and selfishness-driven systems are battling it out for resources, and more mature universes where prosocial systems have prevailed.

Are Bad Guys More Likely To Seed versus Morph Into Good Guys?

Another related question is whether transition from selfish to prosocial dynamics will be more often smooth or abrupt?  (Or if so, how abrupt?) ...   I.e. are most prosocial problem-solutions obtainable as improvements on less-prosocial problem-solutions?  Or are more of them quite different from any less-prosocial problem solution?

If we consider 0 as totally selfish/closed and 1 as totally selfless/prosocial/open, then the question re gradual transition is, roughly: Out of all good problem solutions available for degree-x prosocial/open creatures, how many of these are quite close to solutions available for degree-.9*x prosocial/open creatures (let's call this Class 1), versus how many can be *seeded* by degree-.9x prosocial/open creatures (in the way e.g. that a caterpillar seeds a butterfly, but is not all that super similar to a butterfly, and the butterfly probably doesn't remember being a caterpillar... or the sense in which humanity could seed very non-human-like AGIs, which still in the biggest picture would retain some inklings of humanity, just as a butterfly's torso and internal organs are a bit caterpillar-ish...) (let's call this Class 2)?   Versus how many are not especially findable by reference to degree-.9*x prosocial/open solutions (Class 3).   The latter category, Class 3, may not matter in terms of immediate post-Singularitly realities, they are things that might emerge only later.  

Of course these are really fuzzy classes.  But my intuition is that Class 2 is more numerous than Class 1... and that the Class 1 solutions are sorta clustered together in solution-space, in regions that aren't generally as good as the best Class 2 solutons.  Whether Class 2 or Class 3 solutions are more numerous I don't feel confident, though it does seem likely that Class 2 solutions are also clustered together in solution-space, with some beautiful Class-3 solutions that are distant-ish from them.

Less technically: Good and open-minded guys are going to prevail most of the time across the multiverse.  But if one has a universe or a region thereof in which less-good and more closed-minded guys prevail, the most likely path from this condition toward one in which goodness and openness are dominant, is probably one in which some of the less-good guys build some other sort of better intelligence that’s a fair bit different from them, rather than one in which  the less-good guys themselves gradually morph into much better guys.

Scenarios in which human dominance is followed by dominance of AGIs seeded by human creations, seem on the face of it likely to be more common than scenarios in which human dominance gradually cedes to AGI dominance by humans improving and extending themselves step by step.  Because the number of ways to create prosocial AGI minds de novo, based on human-compatible ideas, seems likely to be much more numerous than the number of ways to continuously morph humans into something far more prosocial by nature.

The Value of Revolution

If we view a complex social arena like the Internet or the global economy as a sort of “multiverse” of social sub-universes, then this latter point suggests that: In many contexts, communities that are more internally prosocial and helpful are going to deal more robustly with the wildly changing environment.   However, less prosocial communities are more likely to spawn more prosocial ones by small groups of members leaving to create something new and better, than via the old community gradually morphing itself into something far better.   

This becomes something like a sketch of a counting argument for the value of revolution (though not necessarily of the violent kind).

Squashing Paperwork Maximizers via Peer Pressure

This is an entirely different conceptual direction from the oft-expressed worry of Nick Bostrom, Eliezer Yudkowsky and others about “paperclip maximizers” — runaway AI agents that are created with simple goals like “create as may paperclips as possible” and then run amok doing things like turning the whole universe into paperclips.   

Paperclips are a silly example but consider the similar case of an AGI whose job it is to run a company selling people toys.   If its prime directive is to sell as many toys as possible, ultimately as it becomes super powerful its destiny must be to transform all matter in the universe into either toys, toy producers or toy consumers.

In a positive social situation, though, what would happen is an AGI system that seemed to be overreaching in this way would rapidly get deprived of resources by other systems in the community.

This brings us back to the rivalry between opposing conceptual forces in the Internet and the current world economy overall.   On the one hand, is the increased general robustness and success associated with prosociality.  On the other hand, are the “runaway feedback loop” effects that allow companies like Google and Facebook and Amazon to grow their user bases exponentially fast — or that allow rich people to get richer faster and faster the richer they get, via exploiting better “only for the rich” investment opportunities.

In science or hard-core engineering like Linux kernel development, or in the fine arts, the power of pro sociality seems to largely overwhelm the runaway feedback growth effects.   These are then domains that foster goodness and openness — they work more like the multiverse as a whole.   

In pop culture or business, feedback growth sometimes happens so fast that the advantages of prosociality don’t have time to establish themselves.  Instead the environment is captured by rapidly growing self-oriented entities, which then redirect culture toward problems that are fairly effectively solved via selfish agents.   This sort of domination dynamic degrades the freewheeling heterogeneity of the environment and chips away the conditions that make the relevant sectors of human society resemble a “micro-metaverse.”

It’s interesting to conjecture how the prevalence of this sort of runaway feedback growth depends on the underlying parameters of the network.   It seems that if long-distance communication is cheaper (say, in physics, the speed of light is higher relative to other physical parameters) then there is less advantage to being a social agent system — so perhaps less opportunity for the advantageous dynamics of pro sociality to squash runaway feedback dominance dynamics.  OTOH if the speed of light is too slow relative to other physical parameters, it will be too hard for distributed intelligence to really happen, and we will see things gravitate toward separate self-contained systems just for efficiency reasons.   There should be a speed-of-light sweet spot that balances communication efficiency with the ability of pro sociality to combat runaway domination dynamics.

The very tentative conclusion within the human social world then is: 

• If we can keep things within a generally heterogeneous, open, decentralized vein, then we have an environment in which good and open-minded agents are likely to prevail.   

• When selfish, closed-minded agents pop up somewhere for a while, there is the possibility they may improve themselves, but also the perhaps stronger possibility that they seed something better than but quite different from themselves

• Speed of communication/influence among people in different communities should be not too high, and not too low relative to other parameters ( like speed of human thought, and speed of human creative and practical production),in order to keep this heterogeneity and balance

• It’s harder to find confidence about what will happen in any particular world (such as ours) than about what is the most likely thing across the multiverse (or a large diverse subset of a single universe) overall

The arguments I’ve presented here are admittedly rather hand-wavy.   However I think it’s quite possible to make them fully rigorous, e.g. either via formalizing them mathematically or via running simulations.

…

P.S. Thanks are due to Zar Goertzel, Ruiting Lian, Dann Toliver , Greg Meredith and Ted Goertzel for feedback on early partial write-ups on some of these ideas. You guys are wonderfully prosocial in sharing your insights, providing nice evidence in direction of the key points in this post ;)