Dissolving the Problem Of Evil

Published: 23 Jul 2022

I’ve discussed in Modeling Technical Income how in a short project, best practices are a waste of time, but over a long project those best practices are a huge time savings. This isn’t unexpected in theory: if we have the same evaluation function and we apply it to two different datasets we quite obviously have the capacity to get different results. Yet sometimes, in practice rather than theory, people fail to recognize this triviality. For example, people often explain cognitive bias in terms of the errors they cause rather than the good they produce, but a lot of what people treat as fallacious biased thinking is really just good thinking given tight time constraints. The time saving optimizations that their mind is taking advantage of short-circuit the evaluation in situations where knowing one of the datasets is hard.

For a more contentious example of the same flawed thinking lets consider the infamous problem of evil. It states:

It is making an argument which is claiming that set x is guaranteed to be equal to subset(x). We can know there is a full x because omniscience is introduced. We know there is a subset(x) because we aren't omniscient.

It declares a guaranteed equality of evaluation function over all possible data subsets of an omniscience entity. Humans evaluate f(subset(x)). The omnscient evaluates f(x).

If f(subset(x)) = f(x) in a general way then subsets are equal to sets. If the problem of evil is a valid argument, then it must be the case that 0 = 1. It isn't a valid argument. See further below for a simulation demonstrating that limited information produces the observation of evil even when evil does not exist.

Less Informed Agent

Score:

Moves:

Average Score Per Move:

More Informed Agent

Score:

Moves:

Average Score Per Move:

I've hidden part of the board above which calls attention to the way more information lets you gain utility despite that utility not being visible. I'm also only showing one simulation, because it should draw attention to the statistical nature of the issue. Some will watch this and conclude from their sample a falsehood - that the two agents were of equal ability. They aren't.

A somewhat interesting caveat of great import only to certain personalitites is the that the average game length for the more observant agent is higher. This has radical implications for how an intelligent agent ought to be made.

P1. If an omnipotent, omnibenevolent and omniscient god exists, then evil does not.

P2. There is evil in the world.

C1. Therefore, an omnipotent, omnibenevolent and omniscient god does not exist.

Thinking through the implications of being omniscient is, when reasoning about the evil that we observe in the world, rather hard. A lot of strategic lazinesss has been applied to not thinking about what omniscience would actually imply. Obviously if we point to evil, it is evil. Case closed right?

Not so fast. Literally, don’t try to think fast. Your mind is good at being hasty, but that doesn’t mean you are right. It just means you skipped thinking.

Lets think about this more slowly. Below I’ve built a simulation. It has two agents. One agent can see one space in front of it. The other is ever so slightly expanded in terms of what it can see. It can see one additional unit. Observe that despite sharing the same function for valuing situations, the two agents can disagree as to what action is good to take. The one space agent will never explore areas in which there is a negative reward, but the two space agent will do things that the one space agent considers evil, because it could make better decisions on account of additional information.

What we’re seeing here is that when you evaluate the problem of evil from first priniciple, using the part of the mind that is comfortable with thinking about things slowly and in depth, rather than using the intuitive part of the mind which short-circuits evaluation, the argument reveals itself to be fundamentally broken. Not only is it not true that omniscience results in a lack of evil, it creates the conditions for us to observe someone else doing evil when in fact they are doing good.

This happens well before you get to omniscience. It happens the moment you add even one more unit of capacity to observe what is happening.

This should be obvious. If we have the same function, but we apply it to two different datasets, clearly the result can differ. Each additional unit of information just increases the likelihood of divergence.

I’ve taken to calling this error the horizon effect, because when I see it happen it appears to happen on account of people having an information frontier which they haven’t explored. They don’t recognize that the dataset they evaluate isn’t actually the only dataset present. They don’t extend their thinking beyond the horizon, but keep it fixed on what is near at hand.

It makes sense that people do this.

We have a limit to how much time we have. We have to use that time wisely. As a consequence, strategic laziness is essential. Real wisdom recommends diligence, but insight tells us that strategic laziness and strategic dilligence are the same thing: a choice about where to apply our time in order to be effective. So people do it and sometimes they make the inevitable mistakes.

We can’t explore every frontier. We can’t push every horizon outward. We simply don’t have the time to do so. If just one step of additional information is enough to cause differing evaluation functions, it is quite obvious that extending it an infinte amount opens up the opportunity to even more divergence.

This might feel like sloppy reasoning to someone who really likes the problem of evil. There are, after all, very evil things. Innocent children, for example, die horrible deaths.

As a Creator myself - though I understand this is horrible from the perspective of the time bound - I create metrics by which agents are evaluated then I very intentionally set out to find the areas in which those metrics are not properly maximized toward good with full intention to get them to be good. An algorithm that experiences itself temporally would, over small time periods, conclude that I am evil because I subject it to evil experience which point out the minimization of those metrics. Yet nevertheless the end result given a larger time bound creates the condition that the evaluation of what I did maximized, not minimized, but maximized the metric.

More knowledge can lead to optimal decisions which appear sub-optimal to the blind. As a consequence, prefer learning to judgement. This logic extends especially unto God who knows more then all. The fall out of this would seem to be that a person should prefer gaining more information in order to make better decisions. Also worth noting is that the two space agent who knows more can prefer longer paths and have worse then one-move only average short-terms despite being logically greater in the long-term.

The code to generate the visualization, imports not included: