Indirect Mind Uploading:
Using AI to Avoid Staying Dead
Paul Almond
page 10 of 13
It is possible to state a criterion that the person-specific world and brain models must satisfy. Just as recorded data was obtained about the original person’s life, it would be possible, by running these models, and also using the other modelling systems mentioned above, to generate the recorded data that would have resulted if these models had accurately described the person. The criteria that needs to be satisfied is simple: both the original archives from which we are working (made by recording the original person’s biological life) and these archives obtained from running the model must be the same for a given combination of person-specific brain and world models to be valid.
To put the above idea more simply, if we have a possible model describing someone’s life and experiences and we run that model so that it relives the experiences of that person, then what that model describes the person doing and experiencing must match with the information we already have about what the original person did and experienced.
This criterion would allow a lot of possible models to be eliminated, but there are two problems:
- An exhaustive search of all possible models until we happened to find one that meets this criterion would be impractical. It would take too long.
- There will actually be an infinite number of possible models that meet this criterion: for any model that we can find that meets it we will always be able to produce a number of more complicated variations of that model that also meet it.
How do we deal with these problems? The first one must be dealt with by using a search method that does not involve generating and checking every possible model. The second problem needs a method of selecting the most plausible model from all of the possible models.
One way that this could be achieved is to start by expressing the model in very vague terms, by assigning only general qualities to it. If the degree of generalization is adequate, then there would be a limited number of such models available and all of them could be checked. There is no point checking each such model to see if it satisfies the criterion of consistency with the recorded data: it is an incomplete model and cannot be expected to do this. For each partial model, we should, however, make an estimate of the number of complete models that could be made, which have the same general characteristics of the partial model that we are testing and that would be consistent with the recorded data. When we have tested all the possible ways of setting up general characteristics of the model we would select the one that we estimate to have the largest number of complete models with the same characteristics that are consistent with the recorded data. We would then continue the process, by adding more characteristics to this partial model, testing at each stage for the number of complete, consistent models that are available with the characteristics that we are trying.
Some readers will recognize such a computation as equivalent to Occam’s razor, the philosophical idea that scientific models should contain as little as possible. It would be not be a trivial computation to perform. It would require more information than we currently have about model generation to do it really well and would merit an article in itself. However, human brains can construct sophisticated models of reality so we have reason to believe that such sophisticated modelling processes can be performed.
Image 6: William of Ockham (Occam) (c.1280-1347), famous 14th century intellectual of Occam's Razor fame.