Less Wrong is a community blog devoted to refining the art of human rationality. Please visit our About page for more information.

buybuydandavis comments on The Useful Idea of Truth - Less Wrong

77 Post author: Eliezer_Yudkowsky 02 October 2012 06:16PM

You are viewing a comment permalink. View the original post to see all comments and the full post content.

Comments (515)

You are viewing a single comment's thread.

Comment author: buybuydandavis 03 October 2012 01:54:51AM 10 points [-]

I don't think EY has chosen the most useful way to proceed on a discussion of truth. He has started from an anecdote where the correspondence theory of truth is the most applicable, and charges ahead developing the correspondence theory.

We call some beliefs true, and some false. True and false are judgments we apply to beliefs - sorting them into two piles. I think the limited bandwidth of a binary split should already be a tip off that we're heading down the wrong path.

In practice, ideas will be more or less useful, with that usefulness varying depending on the specifics of the context of the application of those beliefs. Even taking "belief as predictive model" as given, it's not that a belief is either accurate or inaccurate, but it will be more or less accurate, and so more or less useful, as I've claimed is the general case of interest.

Going back to the instrumental versus epsitemic distinction, I want to win, and having a model that accurately predicts events is only one tool for winning among many. It's a wonderful simulation tool, but not the only thing I can do with beliefs.

If I'm going to sort beliefs into more and less useful, the first thing to do is identify the ways that a belief can be used. What can I do with a belief?

I can ruminate on it. Sometimes that will be enjoyable, sometimes not.

I can compare it to my other beliefs. That allows for some correction of inconsistent beliefs.

I can use it to take action. This is where the correspondence theory gets its main application. I can use a model in my head to make a prediction, and take action based on that prediction.

However, the prediction itself is mainly an intermediate good for selecting the best action. Well, one can skip the middle man and have a direct algorithmic rule If A, do(x) to get the job done. That rule can be useful without making any predictions. One can believe in such a rule, and rely on it, to take action as well. Beliefs directing action can be algorithmic instead of predictive, so that correspondence theory isn't the only option even in it's main domain of application.

Back to what I can do with a belief, I can tell it to my neighbor. That becomes a very complicated use because it now involves the interaction with another mind with other knowledge. I can inform my neighbor of something. I can lie to my neighbor. I can signal to my neighbor. There are quite a number of uses to communicating a belief to my neighbor. One interesting thing is that I can communicate things to my neighbor that I don't even understand.

What I would expect, in a population of evolved beings, is that there'd be some impulse to judge beliefs for all these uses, and to varying degrees for each usage across the population.

So charging off on the correspondence theory strikes me as going very deep into only one usage of beliefs that people are likely to find compelling, and probably the one that's already best analyzed, as that is the perspective that best allows for systematic analysis.

What I think is potentially much more useful is an analysis of all the other truth modalities from the correspondence theory perspective,

Just as Haidt finds multiple moral modalities, and subpopulations defined in their moral attitudes by their weighting of those different modalities, I suspect that a similar kind of thing is happening with respect to truth modalities. Further, I'd guess that political clustering occurs not just in moral modality space, but in the joint moral-truth modality space as well.