An AI safety plan is a description of proposed actions people could take in the world to make AI go well. This is NOT supposed to be a political agenda, and we hope it to avoid turning into one. Although it may be inevitable we have to take some political actions. This is kind of described in section 1 of https://www.bridgespan.org/getmedia/16a72306-0675-4abd-9439-fbf6c0373e9b/strong-field-framework.pdf [TODO: we haven’t described this super clearly - need to expand on this for clarity].

Some SAFE properties for this plan (the mnemonic is not too forced, I promise):

• Sufficient: if all the actions are carried out, we would consider the world in a good state.
  • NB: just solving the alignment problem, or tackling near-term risks, is not sufficient.
• Action-orientated: the plan is a set of actions that explain how people can contribute (and thus how our course could prepare people to contribute), and does not just describe a list of events that happen.
• Feasible: we think it’s reasonably plausible the plan can be executed. This generally excludes plans that require actors to take significant actions against their own interests.
• Encompassing: the plan covers all actions needed, not just some for a short time frame or one jurisdiction.

Why do we need criteria?

• To help clarify what it is we’re looking for
• To make sure what we produce will be sufficient to guide the field

Earlier notes

Criteria

• Must have
  • Gives us confidence that if all the actions are carried out, things would be good
  • Seems feasible. Generally this excludes plans that require:
    • actor(s) to take significant actions against their own interests
    • all governments to magically agree on something very complex, particularly where there are benefits for defectors and enforcement is hard
  • Explains what actions people can take to contribute to the plan (e.g. is a set of assignable/assigned actions, not just list of events to happen)
  • Are comprehensive, e.g. cover all the actions we need to be taken, not just those in one jurisdiction
• Should have
  • Detailed enough for people to evaluate whether they are contributing to the plan
  • Well-communicated / clear
  • We can track progress on the plan
    • Ideally from early on, so we can adjust it if we realise the plan won’t work. E.g. having falsifiable hypotheses
  • Robustness to some failure
• Could have
  • Beyond just making sure we don’t die, the plan explains how we can maximize the benefits of AI
  • Multiple viable paths to success

Notes on Good Strategy Bad Strategy:

• Also see Notes on Good Strategy/Bad Strategy
• Problem diagnosis, guiding policy, coherent actions
• Leverage, proximate objectives, chain-link systems, comparative advantage