Read Tyler Cowen’s Talent, and the later sections of Huang Zongxi’s waiting for the dawn on how to identify talent/hire good men. What questions reveal people in the face of incentives, what cannot be optimised against, and how could we operationalise them and put equivalents of these questions to AI systems to prevent goodharting?
Hiring someone with a weird background, someone who gets rejected at the first-pass HR filter at every other company, is a non-correlating bet. And portfolio theory tells us that even if your non-correlating bet doesn’t do great on its own, your whole portfolio will do better if you take those bets. (Byrne Hobart) The uses of randomness Legibility, markets, communities
bureaucratic hiring = avoid hiring bad people (errors of commission) VC/EA hiring = avoid not hiring good people (errors of omission)
Silicon valley VCs want to avoid errors of omission because the ROI on employees at early stage startups has a very heavy right tail. People in risky roles, like politics maybe, or cops and the military, really want to avoid errors of commission because there’s only a heavy left tail
Tyler also wants to avoid errors of omission in his philanthropy because he’s looking for counterfactual impact, which will involve betting against the efficiency of the talent-search pipeline, e.g. western universities. He has to ask, why hasn’t harvard or someone else already funded this person?
Same talent scout spotted count Basie, Billie holiday, and many many others.
Indicators of potential (intelligence) should be used to hire young people, not old people. Also Swedish CEO study shows CEO’s of even large companies are only around the 80-85th percentile. Also intelligence is priced in.
What does the multiplicative model predict? How do you actually test if a multiplicative or additive model makes more sense?
Predictors for high performance and average performance can be quite different.
Best pie in the sky improvement to hiring would be to somehow improve the incentives for references to be honest and informative, since they have the most info on the candidate. You might be able to do this by some kind of conditional incentive tied to the performance bonus/hiring outcome?