Background

Currently, Metaculus AI looks like:

A collection of questions of various shapes and sizes, with some clusters here and there.

We want to build towards something that looks more like:

A structured worldview of AI, where making progress on each question usefully feeds into a larger whole.

To start this, we are building up a structured model of the drivers of AI progress. That is, key inputs that determine the pace and direction of progress. We invite discussion and suggestions from the community. For example:

• Adding further factorization and branches
• Suggesting questions that capture progress under a particular branch
• Modelling/analysing the relative importance of different branches
• ...or something else!

Model

The current model looks like this (if you prefer to read and comment in a Google doc, you can do so here):

• Compute
  • Economics (how much compute is there, and what does it cost?)
    • Supply
      • Regulation (e.g. export restrictions on compute, as suggested by Shahar Avin)
      • Technology (see this comment for more on this tripartite distinction)
        • Moore’s law chip improvements
        • Specialised hardware (e.g. TPUs)
        • Rewriting software to increase performance
    • Demand
      • Industry
        • Applications and actual use cases
        • ...
      • Government
        • Military
        • ...
      • Academia
        • ...
  • Algorithmic parallelizability (how efficiently can more compute be used to train deep learning models?)
    • Data-parallelism (see e.g. OpenAI’s work on optimal batch sizes)
    • Model-parallelism (see e.g. this paper)
• Algorithms
  • Talent(?)
    • Supply
      • Regulation (e.g. regarding immigration)
      • Education
      • Questions on conference attedance
      • ...
    • Demand
      • ...
  • Insight
    • Progress in neuroscience
    • Questions on trends
      • e.g. Will the current linear progress in conceptual insights be sustained for Y years?
    • Questions on unsolved benchmarks
      • When will an AI system match amateur human performance on Obstacle Tower with no domain-specific hardcoded knowledge?
• Data
  • Simulators (e.g. DeepMind PsychLab)
  • Data-sets (e.g. for supervised learning)
    • Regulation
    • …