Incomplete contracts are an underrated concept for thinking about AI alignment.
Specifying the complete set of things we want is in important cases doomed to fail, as was writing down all the rules for behavior in expert systems decades ago.
Two sources of AI agent failure:
1) humans can't articulate what they want,
2) agents misread instructions or hallucinate
When writing this last year, the bottleneck was #2. Striking how fast it moved to #1.
(cc: @BenSManning @AndreyFradkin @johnjhorton)
nber.org/books-and-chap…
