2023 Air Street Capital The State of AI Report 2023

Slide schematic 4/4

list 1/1 paragraph 1/1 table 1/1 title 1/1

Components 4

bullet paragraph data headline

Tools 1

table/data: Level 2 (≅ NP^NP) Humans cannot reliably evaluate what AI is doing. Humans can evaluate what AI is doing with AI assistance. Humans can evaluate this assistance.

Metrics 0

∅

No metrics

No metric insight is anchored to this slide.

Frameworks 1

recursive-reward-modeling 90%

Jan Leike, the leader of OpenAI's alignment team, frames this as a recursive reward modeling problem

Find anything

The slide uses a complexity-theory-inspired notation (NP^NP) to illustrate the recursive difficulty of evaluating AI assistants.