Angles of Attack

post by whpearson · 2011-02-03T18:53:23.127Z · LW · GW · Legacy · 4 comments

Contents

  Consider this article in collaborative draft
  Meta angles of attack
None
4 comments

For humans problems can seem intractable for a long time and then suddenly become easy. Forming a coherent chemistry was taking a very long time until Lavoisier thought to look at the mass of reactants and products. And then we had a periodic table 100 years after that. Compare this to little progress from having bounced around in alchemy for 2000 years or so.

So identifying and understanding angles of attacks is important for tackling the thorny problems that face us today.

Consider this article in collaborative draft

So why do we need angles of attacks? Getting an angle of attack can be seen as reducing the search space. Imagine in the pre-chemistry era, we want to make a substance that you can make a sword out of without all the bother of smelting metal. You have seen that things can be created by placing different things together, how do you select the things to make your new sword material? You know of a multitude of different materials: thousands of different types plant matter (100s of plants and the different parts of each), similar amounts of animal matter,various stones,  water and more. You know of different processes such as heating, bashing and freezing. This adds up to a huge search space, you can't afford to brute force it. You need an atomic theory of chemistry to understand what gives items strength and brittleness and even given that we still haven't found much easier to make materials than metal for the purpose of swords. But we know that mixing rhino horn, bones from a strong man and bull's liver and freezing it over the night of a full moon will not make sword material.

We currently lack angles of attack for intelligence and meta-ethics. Even worse, we don't have ways to tell when we do have an angle of attack, without implementing a full AI. The principle of conservation of mass reduced the search space for atomic chemistry and could be tested immediately.

To give an example from lesswrong. Eliezer takes the principle that it doesn't matter how an agent computes its actions [citation needed], I think that energy considerations might be important. His supposition would give us an angle of attack, but how would we know whether it is a good one or not? Evolutionarily it would seem important, but a decision theory where you have to consider the energetic costs of making a decision is a lot more complex. You have to specify a machine and the energetic costs and payoffs of the computation before you can say what the optimal decision is.

Meta angles of attack

There are angles of attack that have worked in the past for finding new angles of attack that might help for the fields we are stuck in.

Quantification: That is measuring some well defined stable property of an object or the world, and relating that number to other numbers.

This one seems to be running out of steam. It is not so hot in the field of economics as it was in physics. Measuring intelligence is uncomputable by Hutter and Leggs definition of it (one that ignores energy costs though). So to apply this angle of attack we need a different

Reductionism: I was reminded in the comments that reductionism is a good meta angle of attack. That is look for smaller simpler sub units. And then derive the high-level behaviour from your understanding of them.

Following paths: If you assume that things like energy and information do not pop out from nowhere, then if you identify them in the system that you are studying they must have come from somewhere. So in biology you know that the information to make a daughter look like her mother has to flow somehow so you would look inside the egg.

This I think is a somewhat promising for studying the brain, if we get sufficient processing power and information on the brain. For example we could do things similar to taint-checking to discover exactly what information was processed over an  individuals lifetime in order to make one action. Depending upon how far the influence of incoming information spreads, this may or may not be useful information.

However the obvious form of this angle of attack doesn't seem so useful when discussing meta-ethics. Knowing that evolution created our ethics does not tell us much about what they are.

Can anyone think of any others? How do people see these problems getting solved?

4 comments

Comments sorted by top scores.