by Sven Nilsen, 2017
This blog post is the start of a new research topic that I have done quite a bit of pre-research on, and it is approaching a state where I think it is justifiable to start some serious investigation. I think it fits the AdvancedResearch community because it deals with a concept breakthrough in system thinking.
Disclaimer: I can not explain all of it in one post, but I hope you will get something out of it. Some parts of this will probably make you more confused. :P
By “nature having intelligence” I mean by a very precise mathematical definition and explained in technical terms: Observers living in universe with a selection effect similar to an exponential reference class with a time symmetric assumption, which allows observers to predict some very computationally expensive aspects about their future.
In a such universe, by using some very few ingredients…
… it is possible to end up with some kind of “intelligence”. Obviously, it is not a human-like intelligence but merely an optimization algorithm (more on this later).
A first attempt to approximate encode this effect on the Traveling Salesman Problem. There is a layered inference of intepretation to understand exactly why you compute this factor and add them together for random routes, and why this is done for pairs of cities, but I hope to explain this better later on. Wish I had more mathematical proofs for the transition from the original concept to the algorithm, but I have not figured out to express it better than using words in natural language. I have done an Dyon implementation of this algorithm that will be added to the repo.
There are other proofs that I have done previously and would like to add, which is about:
I would also like to formalize an argument about expected probabilistic Kolmogorov complexity from uniform initial conditions. There is no existing technical term for this, so I call it “probabilistic Kolmogorov complexity” to borrow the idea of Kolmogorov complexity but with some smooth landscape around it.
When I say “merely an optimization algorithm”, this is not true in the everyday meaning of these words. It is true that a such algorithm will optimize for a single thing, but the problem is that this simple thing is highly interconnected with everything.
For example, human brains are extremely complex. It is the most complex machine we know about. Human brains were evolved by natural selection. One could think of natural selection as a “high level algorithm” running on physical laws.
The intelligence due to selection effects in an exponential reference class, is a “low level algorithm” that could be much closer to the description of physical laws. Therefore, since the observer in a such universe is a product of a high level algorithm that runs on top of a low level algorithm, those two effects combined can be highly complex and interacting and interfering with each other.
For example, an selection effect does not explain e.g. why human noses are of a particular size, unless you have a very powerful computer that could trace all mutations and compute an expected nose size in a flat reference class and compare it to an exponential reference class. On the other hand, a selection effect makes it easier to explain how an observer ends up with the most complex brain among all other animals evolved so far. However, there are nuances to this idea that shows how problematic it is to separate the effects:
The “merely an optimization algorithm” we are talking about, changes the perspective of how one looks at the universe, at humans and everything.
We have yet to come up with an immediate practically useful application of it. There are two ways we can benefit from this knowledge:
The first kind of applications is trivial: More knowledge that allows us to improve algorithms is nice. However, this would only let us exploit the power of the computers we are able to build ourselves, or by finding approximations to this behavior in naturally occuring systems. The second kind of applications lets you harvest the computational energy of the whole universe, without paying the electrical bill, possibly in a way to make ridiculously powerful predictions about the future.
For example, if alien civilizations are rarer because of the selection effect, then you can assign a lower risk of meeting hostile aliens on your space exploration among the stars. This assumes that you can sort out all the hair ball and spaghetti code problems you get by considering this effect’s influence on everything that makes up a civilization, and in the end you come to the conclusion “yeah, there will be fewer alien civilizations”.
Unfortunately, we humans have yet to explore the stars, so it could take hundreds of years before a such piece of knowledge becomes useful. It also does not say how likely you would expect to see alien civilizations in a flat reference class. Such civilizations could be rare anyway, and then a such prediction would not be that useful.
Another issue is whether by some strange coincidence, this optimization effects leads to some skewed probabilities on important moral decisions. For example, the effect of the risk nuclear war on the number of observers is highly non-linear. A high risk of nuclear war could lead to less war overall, but in the event it breaks out it could be catastrophic. It is imaginable that some “sweet spot” is located near a high risk of catastrophe, so observers get more likely to be located in a time where they are on the edge of extinction. This is also a studied phenomena in flat reference classes, called Antropic Shadow.
The obvious problem with observer selection effects is that they are subjective. One reason it is useful for humanity overall is, simply put, because this species share the same brain design as the observer, at least in our particular case.
In a flat reference class, there are no particular favorism of an observer, but there could be in an exponential reference class. This is not properly understood yet, in part because of quantum physics that I could elaborate on in a later post.
So, we got the weird situation where multiple people are interested in observer selection effects, but none of them can prove to the others that they are the focus point of the reference class. After all, a person experience themselves from the inside and all other people from the outside.
It gets worse: We can find animals in nearby surrounding that does not have as complex brains as we do. We could talk to some smart animals, or perhaps aliens, for example by using a telepathy technology based on infrared light with brain-to-brain communication, which is already being developed and rapidly improving. What if we convince them the observer selection effect was plausible, and they got interested in researching this field? They might be confronted with the view that e.g. their brains are unexpectedly simple, compared to what they should be in an exponential reference class. With other words, they could agree with us about the theory, while at the same time find it incredible implausible to actually happen from their point of view.
If we do not find suitable animals or aliens for this experiment, we might be able to create artificial minds where we control the complexity manually. A paradox is when you are able to build a more complex mind than your own, which does not seem that implausible given the current pace of technological development. Can we construct a mind that is a more plausible observer than our own mind?
When encountering such scenarios, which seem very real and plausible given more research in this direction, we have to take into account selection effects that interacts with the complex systems that make up observers, and also the effects on seemingly paradoxes that must be resolved in some way, due to existing of lower level systems described by physical laws. This can put us in a situation where we are unable to comprehend the complexity of the effects, since we can create arbitrary computationally hard problems. On the other hand, this could be exploited to narrow down the search for the correct physical theory of observers.