Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

TL;DR: If you want to notice when powerful AIs are behaving badly, a reasonable strategy is to carefully audit a small fraction of AI actions. This kind of strategy has two kinds of failures which can be tackled somewhat independently: auditing failures, where the audits fail to recognize some bad behavior, and concentrated failures, where the AI caused an irreversible catastrophe in so few actions or such a short time frame that auditing didn’t have a chance to review the bad actions.

Many oversight techniques are so expensive that using them on every AI action would destroy most of the value of using AIs in the first place: you might want to have humans extensively review all AI-generated code line by line, but this reduces AI usefulness so much that you would often be better off having [...]

---

Outline:

(03:26) Sampling concentrated failure

(03:52) Latency concentrated failure

(05:14) Mitigations of concentrated failures

(06:01) More than one level of auditing

(08:22) Comparison with related decompositions

(08:43) Low-stakes vs high-stakes

(09:53) Minor issues with low-stakes vs high-stakes

(11:37) Outer alignment and inner alignment

(12:43) Conclusion

The original text contained 4 footnotes which were omitted from this narration.

---

First published:
December 11th, 2023

Source:
https://www.lesswrong.com/posts/hirhSqvEAq7pdnyPG/auditing-failures-vs-concentrated-failures

---

Narrated by TYPE III AUDIO.

“Auditing failures vs concentrated failures” by ryan_greenblatt, Fabien Roger

LessWrong (30+ Karma)