My recent posts

post by paulfchristiano · 2016-11-29T18:51:09.000Z · LW · GW · 0 comments

Contents

  Terminology and concepts
None
No comments

Over at medium, I'm continuing to write about AI control; here's a roundup from the last month.

Many of these seem like interesting things to discuss here; would it be better to post each of these as a link when I write it?

#Strategy

• Prosaic AI control argues that AI control research should first consider the case where AI involves no "unknown unknowns."
• Handling destructive technology tries to explain the upside of AI control, if we live in a universe where we eventually need to build a singleton anyway.
• Hard-core subproblems explains a concept I find helpful for organizing research.

#Building blocks of ALBA

• Security amplification and reliability amplification are complements to capability amplification. Ensembling for reliability is now implemented in ALBA on github.
• Meta-execution is my current leading contender for security and capability amplification. It’s totally unclear how well it can work (some relevant speculation).
• Thoughts on reward engineering discusses a bunch of prosaic but important issues when designing reward functions.

Terminology and concepts

• Clarifying the distinction between safety, control and alignment.
• Benignity may be a useful invariant when designing aligned AI.

0 comments

Comments sorted by top scores.