Okay, that seems reasonable. But I want to repeat my claim[1] that people are not blocked by ânot really knowing what worked and didnât work in the FTX caseâ â even if e.g. there was some type of rumor which was effective in the FTX case, I still think we shouldnât rely on that type of rumor being effective in the future, so knowing whether or not this type of rumor was effective in the FTX case is largely irrelevant.[2]
I think the blockers are more like: fraud management is a complex and niche area that very few people in EA have experience with, and getting up to speed with it is time-consuming, and also ~all of the practices are based under assumptions like âthe risk manager has some amount of formal authorityâ which arenât true in EA.
(And to be clear: I think these are very big blockers! They just arenât resolved by doing an investigation.)
Or maybe more specifically: would like people to explicitly refute my claim. If someone does think that rumor mills are a robust defense against fraud but were just implemented poorly last time, I would love to hear that!
Again, under the assumption that your goal is fraud detection.
It seems like a goal of ~âfraud detectionâ not further specified may be near the nadir of utility for an investigation.
If you go significantly narrower, then how EA managed (or didnât manage) SBF fraud seems rather important to figuring out how to deal with the risk of similar fraudulent schemes in the future.[1]
If you go significantly broader (cf. Oliâs reference to âdetecting and propagating information about future adversarial behaviorâ), the blockers you identify seem significantly less relevant, which may increase the expected value of an investigation.
My tentative guess is that it would be best to analyze potential courses of action in terms of their effects on the âEA immune systemâ at multiple points of specificity, not just close relations of a specific known pathogen (e.g., SBF-like schemes), a class of pathogens (e.g., âfraudâ), or pathogens writ large (e.g., âfuture adversarial behaviorâ).
Given past EA involvement with crypto, and the base rate of not-too-subtle fraud in crypto, the risk of similar fraudulent schemes seems more than theoretical to me.
Okay, that seems reasonable. But I want to repeat my claim[1] that people are not blocked by ânot really knowing what worked and didnât work in the FTX caseâ â even if e.g. there was some type of rumor which was effective in the FTX case, I still think we shouldnât rely on that type of rumor being effective in the future, so knowing whether or not this type of rumor was effective in the FTX case is largely irrelevant.[2]
I think the blockers are more like: fraud management is a complex and niche area that very few people in EA have experience with, and getting up to speed with it is time-consuming, and also ~all of the practices are based under assumptions like âthe risk manager has some amount of formal authorityâ which arenât true in EA.
(And to be clear: I think these are very big blockers! They just arenât resolved by doing an investigation.)
Or maybe more specifically: would like people to explicitly refute my claim. If someone does think that rumor mills are a robust defense against fraud but were just implemented poorly last time, I would love to hear that!
Again, under the assumption that your goal is fraud detection. Investigations may be more or less useful for other goals.
It seems like a goal of ~âfraud detectionâ not further specified may be near the nadir of utility for an investigation.
If you go significantly narrower, then how EA managed (or didnât manage) SBF fraud seems rather important to figuring out how to deal with the risk of similar fraudulent schemes in the future.[1]
If you go significantly broader (cf. Oliâs reference to âdetecting and propagating information about future adversarial behaviorâ), the blockers you identify seem significantly less relevant, which may increase the expected value of an investigation.
My tentative guess is that it would be best to analyze potential courses of action in terms of their effects on the âEA immune systemâ at multiple points of specificity, not just close relations of a specific known pathogen (e.g., SBF-like schemes), a class of pathogens (e.g., âfraudâ), or pathogens writ large (e.g., âfuture adversarial behaviorâ).
Given past EA involvement with crypto, and the base rate of not-too-subtle fraud in crypto, the risk of similar fraudulent schemes seems more than theoretical to me.