resources

Example Of Using Cruxes For Ai Safety Uplift Organization

Last updated: 10/16/2026

Example of Using Cruxes for Ai Safety Uplift Organization

here is an example exercise of me doing cruxes. you can see below a statement that i want to investigate if is true and what cruxes need to be true in order for the statement to be true.

you can investigate and identify your cruxes by asking, "if cruxes A, B, and C are true, then statement X is true" and if that doesn't feel true, you have more cruxes. for example, if "the delta between existing commercial solutions and what I can build is large; there are a lot of actual productivity gains to be made; ... I can get funding for this organization" are true, then "An AI uplift consultancy is a great idea. there is nothing wrong with it. by doing this consultancy it will significantly counterfactually uplift ai safety." is true

The data structure is

  • Crux
    • Other cruxes or uncertainties

Cruxes

Statement: An AI uplift consultancy is a great idea. there is nothing wrong with it. by doing this consultancy it will significantly counterfactually uplift ai safety.

Cruxes:

  • The delta between existing commercial solutions and what I can build is large

  • There are a lot of actual productivity gains to be made by. General organization productivity coaching; AI-specific uplift

  • The reason this is not being done already is because of coordination failure, rather than it actually just not being effective.

  • Organizations have an appetite for "AI uplift"

  • Having a seperate organization is better than just coordinating a network of AI-enablement people

    • Is this true? what needs to be true for this to be true?
    • How do forward deployed engineers work? In what situations are they better than just the organization hiring someone to do the thing?
      • When there is a function that needs to be ongoing and deep organization knowledge is important
      • When there are some legal / IP concerns about the FDE
    • Centralizing tools allows making information tacit
  • Why should resolution not just do this?why shouldn't they centralize ai safety automation, as they already will be spending a shit time of effort doing this

  • I can get funding for this organization

    • I am 'good enough' to actually run this organization, and I am one of the best fits for it.