The essence of the problem in my view is 1) choosing (and averaging over) good reference classes, 2) understanding the heavy tails, and 3) understanding that startup founders are selected to be good at founding (a correlation vs causation issue).
First, consider the first two points:
1. Make very sure that your reference class consists mostly of startups, not less-ambitious family/lifestyle businesses.
2. The returns of startups are so heavy-tailed that you can make a fair estimate based on just the richest <1% of founders in the reference class (based on the public valuation and any dilution, or based on the likes of Forbes billionaire charts.).
For example, in YC, we see that Stripe and AirBnB are worth ~$100B each, and YC has maybe graduated ~2k founders, so each founder might make ~$100M on-expectation.
I’d estimated $6M and $10M on-expectation for VC-funded founders and Stanford-founders respectively.
A more controversial reference class is “earn-to-give founders”. Sam Bankman-Fried has made about $10B from FTX. If 50 people have pursued this path, the expected earnings are $200M.
The YC and “earn-to-give” founder classes are especially small. In aggregate, I think we can say that the expected earnings for a generic early-stage EA founder are in the range of $1-100M, depending on their reference class (including the degree of success and situation). Having said this, 60-90% of companies make nothing (or lose money). With such a failure rate, checking against one’s tolerance for personal risk is important.
Then, we must augment the analysis by considering the third point:
3. Startup founders are selected to be good at founding (correlation vs causation)
If we intervene to create more EA founders, they’ll perform less well than the EAs that already chose to found startups, because the latter are disproportionately suited to startups. How much worse is unclear—you could try to consider more and less selective classes of founders (i.e. make a forecast that conditions on / controls for features of the founders) but that analysis takes more work, and I’ll leave it to others.
I looked at some literature on this question, considering various reference classes back in 2014: YC founders, Stanford Entrepreneurs, VC-funded companies.
The essence of the problem in my view is 1) choosing (and averaging over) good reference classes, 2) understanding the heavy tails, and 3) understanding that startup founders are selected to be good at founding (a correlation vs causation issue).
First, consider the first two points:
1. Make very sure that your reference class consists mostly of startups, not less-ambitious family/lifestyle businesses.
2. The returns of startups are so heavy-tailed that you can make a fair estimate based on just the richest <1% of founders in the reference class (based on the public valuation and any dilution, or based on the likes of Forbes billionaire charts.).
For example, in YC, we see that Stripe and AirBnB are worth ~$100B each, and YC has maybe graduated ~2k founders, so each founder might make ~$100M on-expectation.
I’d estimated $6M and $10M on-expectation for VC-funded founders and Stanford-founders respectively.
A more controversial reference class is “earn-to-give founders”. Sam Bankman-Fried has made about $10B from FTX. If 50 people have pursued this path, the expected earnings are $200M.
The YC and “earn-to-give” founder classes are especially small. In aggregate, I think we can say that the expected earnings for a generic early-stage EA founder are in the range of $1-100M, depending on their reference class (including the degree of success and situation). Having said this, 60-90% of companies make nothing (or lose money). With such a failure rate, checking against one’s tolerance for personal risk is important.
Then, we must augment the analysis by considering the third point:
3. Startup founders are selected to be good at founding (correlation vs causation)
If we intervene to create more EA founders, they’ll perform less well than the EAs that already chose to found startups, because the latter are disproportionately suited to startups. How much worse is unclear—you could try to consider more and less selective classes of founders (i.e. make a forecast that conditions on / controls for features of the founders) but that analysis takes more work, and I’ll leave it to others.