AI ALIGNMENT FORUM
AF

Factored Cognition

Aug 30, 2020 by Rafael Harth

Factored Cognition describes a process in which difficult problems are recursively decomposed into simpler problems. This sequence summarizes my conclusions from thinking about this topic for a few months.

Before we get to the problem itself, first an introduction to the existing alignment proposals that make use of Factored Cognition:

28A guide to Iterated Amplification & Debate
Rafael Harth
5y
0

... and an attempt to characterize what makes something 'a subproblem' rather than 'a part of a large problem':

13Hiding Complexity
Rafael Harth
5y
0

A few more words about the sequence:

16Preface to the Sequence on Factored Cognition
Rafael Harth
5y
0

For our first angle of attack, we use pure mathematics. We define a formal model and explore it as much as possible:

18Idealized Factored Cognition
Rafael Harth
5y
0
9Traversing a Cognition Space
Rafael Harth
5y
0

Before switching gears, we briefly address some commonly misunderstood aspects about Factored Cognition:

11Clarifying Factored Cognition
Rafael Harth
5y
0

Finally, we take a look at the human component that exists in both IDA and Debate. Our goal is to evaluate whether the human's task remains feasible as the schemes are applied to progressively harder problems:

13Intuition
Rafael Harth
5y
0
10FC final: Can Factored Cognition schemes scale?
Rafael Harth
5y
0