You don’t need SI to find the exact truth if you have all information, but without all information, SI finds only the probable.
Yes of course. We are always engaged in a probabilistic search for truth.
But even then it’s based on Kolmogorov complexity which is not computable, so SI is not computable, so “exactness” isn’t ideal. Isn’t it just that other approaches do worse on average?
The ideal, if we had infinite computing power, would be to discover what is most probably true. But the ideal is uncomputable, so we seek out approximations of that ideal that are computable, especially ones that are computable within a certain domain of problems, like the Bayesian information criterion.
While you don’t have to mention Tarski, you do need to remember that some smart people will object to the claim that (in principle) there exists a recipe for truth.
As it turns out, there is a best recipe for finding truth from uncertainty (although it won’t outperform recipes that do things like call for the truth as an ingredient and say not to cook anything). And a recipe that’s almost as good for every problem was discovered in the 1960s. It was also discovered that no one could find a better one and know that they had discovered it.
To find a slightly better one and know you had would require you to follow one step after another for an infinite amount of time. That’s impossible, so the best you can do in theory is to follow almost as good recipe we have.
The problem is that you don’t have time to follow this recipe either. To find the truth to even a simple question using this recipe would require you to follow one step after another until long after the heat death of the universe. While this is less than an infinite amount of time, you can’t wait that long either.
It was meant as a draft of an alternative, based on my limited understanding. I see two thresholds where I think you see one. The recipe is uncomputable, so it would take longer than “long after the heat death of the universe” or any other finite amount of time to finish. Also, the computable functions most similar to it would take more steps than there is time to do.
It’s a good sign our understandings match, but consider that after simply reading the explanation I thought you meant something other than what you did.
Yes of course. We are always engaged in a probabilistic search for truth.
The ideal, if we had infinite computing power, would be to discover what is most probably true. But the ideal is uncomputable, so we seek out approximations of that ideal that are computable, especially ones that are computable within a certain domain of problems, like the Bayesian information criterion.
And a good thing, too, if I understand Tarski’s undefinability theorem correctly.
While you don’t have to mention Tarski, you do need to remember that some smart people will object to the claim that (in principle) there exists a recipe for truth.
I assume some of this, the parts I didn’t write, are not meant to be blockquoted?
It was meant as a draft of an alternative, based on my limited understanding. I see two thresholds where I think you see one. The recipe is uncomputable, so it would take longer than “long after the heat death of the universe” or any other finite amount of time to finish. Also, the computable functions most similar to it would take more steps than there is time to do.
Yes, approximations of Solomonoff Induction need to be not just computable but also tractable.
It’s a good sign our understandings match, but consider that after simply reading the explanation I thought you meant something other than what you did.