5 ms·

One inefficiency could be related to this point Paul makes: We can afford to take at least 10x as much risk as Demo Day investors. And since risk is usually pr

One inefficiency could be related to this point Paul makes: We can afford to take at least 10x as much risk as Demo Day investors. And since risk is usually proportionate to reward, if you can afford to take more risk you should. What would it mean to take 10x more risk than Demo Day investors? We'd have to be willing to fund 10x more startups than they would. Which means that even if we're generous to ourselves and assume that YC can on average triple a startup's expected value, we'd be taking the right amount of risk if only 30% of the startups were able to raise significant funding after Demo Day. So- if a VC can triple a startups expected value that's great, but you'd need to do a bunch of them. I think this is essentially what Dave McClure is doing- making lots of smaller bets to "hit singles" as he says. Reminds me of my college days playing online poker. The best players would have a 20% ROI at the $55 10 person tournament tables, and each game would take an hour. If you just play one at a time, you'd make about $10/hr. That's why everyone played 10 tables at a time- we made 10 times as much.

nostrademons 14y ago
That quote was the sketchiest part of the article for me, because the same math was used to justify the subprime mortgage bubble. The implicit assumption is that the population of startups is uniform across both the set that YC funds and the set that YC does not fund, such that startups in the latter group have the same chance of being a big hit (modulo YC's mentoring, which is accounted for with the "triple a startup's expected value" clause). But the implication of that assumption is that YC is picking startups at random, and that their filtering process is totally useless! This sort of math comes up all the time whenever there's a screening process. Imagine that you're hiring for a large tech company, you currently hire 1% of applicants, and you find that among the employees hired, there is no correlation between your interview scores and the employee's eventual job performance. Can you conclude that your interviewing is useless? Should you ramp up hires so you get more workforce of equal quality? Well, maybe. Because there are a bunch of possible hypotheses that could give this result. Perhaps your interview process is designed to weed out false positives more than false negatives, so it's accurate to the 99th percentile, but then gives no discriminatory power. (Many IQ tests are like this; they're highly correlated with life outcomes up until an IQ of about 140, but beyond that they break down entirely and there's often an inverse correlation with income, happiness, etc. past that). Or perhaps your applicant pool is bimodal: 1% come from other employers and are fully qualified, while 99% are the same jobseekers that every other company rejects. Or perhaps your interview process is broken, and you would do better to find a new one. Or perhaps your interview process is okay, but a number of well-qualified applicants are not even applying to your company. Which of these is correct? You can't know without randomly sampling the population that was rejected and making an estimate of their quality. This is why all decent scientific experiments have a control, and why financial models get backtested on data that was not part of the training set. Even then, there're lots of things that can go wrong in experiment design, and lots of different ways to interpret data that don't necessarily mean "Fund 10x more startups."
- aptwebapps 14y ago
  "That quote was the sketchiest part of the article for me, because the same math was used to justify the subprime mortgage bubble." Except that YC doesn't do this. They aren't funding 10x more startups and pg says he avoids finding out how many get funded afterwards because it's the wrong thing to optimize for. To put it another way, there are a lot of ways to bring down post-Demoday funding to 30% and most of them are not going to be helpful. The observation just points out that given their high VC funding rate YC is probably not optimizing for the homeruns as well as it should from a financial perspective.
  - nostrademons 14y ago
    Perhaps I'm misunderstanding this, but: "We'd have to be willing to fund 10x more startups than they would." I guess that if currently nearly all of your startups are getting VC funding, there are two ways to end up funding 10x more startups than VCs: 1.) Make the VCs fund 10x fewer startups or 2.) Fund 10x more startups yourself. (Or various combinations, of course.) #1 seemed absurd to me, so I read the thought experiment as suggesting #2. Edit: Hmm, I guess another alternative is to fund startups that "look worse" to VCs, which is basically #1 but doesn't seem so absurd. So perhaps I was misreading the quote.
    - aptwebapps 14y ago
      Either you are or I am. ;) "We'd have to be willing to fund 10x more startups than they would." The contracted 'we would' is key. They WOULD have to be willing to do that IF they were going to pursue that goal. They aren't. He's positing the idea that if they were properly optimized for the black swans then they would see a much lower funding rate. Not that they should aim for a low funding rate for it's own sake. But, as he discusses, actually performing such an optimization is hard, not to mention the fact that making money isn't their only reason for running YC.
- resbear 14y ago
  I think the difference here is the assumption for failure rather than success for. The subprime bubble existed due to leverage - fancy models said that defaults were unlikely, so rather than extend loans from their own assets, banks decided to double (or triple, or quadruple...) down and lend out multiples of their own total assets. The models were wrong and they blew up. This is quite different from assuming a high failure rate. The marginal cost of having say, 10% more failed startups isn't a big deal when 95% of them are already failing, and the model accounts for that high failure rate. The black swan is not the unexpected failure that wipes you out, but the unexpected success that outweighs all the failures. Obviously, you'd still be screwed if none of the stuff you invest in pays off. But that would be the case even if your model was more pessimistic; if even 5% success was too much to ask for, perhaps the market for VC is busted, and there was no hope for profit anyway. I don't think pg is implying that screening processes are useless. I'm sure many obviously hopeless ideas get filtered out. It's more that, given the current state of knowledge, everything beyond that basic filter is by nature difficult to discern and we may as well assume a random fat tail distribution.

nostrademons 14y ago

That quote was the sketchiest part of the article for me, because the same math was used to justify the subprime mortgage bubble. The implicit assumption is that the population of startups is uniform across both the set that YC funds and the set that YC does not fund, such that startups in the latter group have the same chance of being a big hit (modulo YC's mentoring, which is accounted for with the "triple a startup's expected value" clause). But the implication of that assumption is that YC is picking startups at random, and that their filtering process is totally useless! This sort of math comes up all the time whenever there's a screening process. Imagine that you're hiring for a large tech company, you currently hire 1% of applicants, and you find that among the employees hired, there is no correlation between your interview scores and the employee's eventual job performance. Can you conclude that your interviewing is useless? Should you ramp up hires so you get more workforce of equal quality? Well, maybe. Because there are a bunch of possible hypotheses that could give this result. Perhaps your interview process is designed to weed out false positives more than false negatives, so it's accurate to the 99th percentile, but then gives no discriminatory power. (Many IQ tests are like this; they're highly correlated with life outcomes up until an IQ of about 140, but beyond that they break down entirely and there's often an inverse correlation with income, happiness, etc. past that). Or perhaps your applicant pool is bimodal: 1% come from other employers and are fully qualified, while 99% are the same jobseekers that every other company rejects. Or perhaps your interview process is broken, and you would do better to find a new one. Or perhaps your interview process is okay, but a number of well-qualified applicants are not even applying to your company. Which of these is correct? You can't know without randomly sampling the population that was rejected and making an estimate of their quality. This is why all decent scientific experiments have a control, and why financial models get backtested on data that was not part of the training set. Even then, there're lots of things that can go wrong in experiment design, and lots of different ways to interpret data that don't necessarily mean "Fund 10x more startups."

aptwebapps 14y ago
"That quote was the sketchiest part of the article for me, because the same math was used to justify the subprime mortgage bubble." Except that YC doesn't do this. They aren't funding 10x more startups and pg says he avoids finding out how many get funded afterwards because it's the wrong thing to optimize for. To put it another way, there are a lot of ways to bring down post-Demoday funding to 30% and most of them are not going to be helpful. The observation just points out that given their high VC funding rate YC is probably not optimizing for the homeruns as well as it should from a financial perspective.
- nostrademons 14y ago
  Perhaps I'm misunderstanding this, but: "We'd have to be willing to fund 10x more startups than they would." I guess that if currently nearly all of your startups are getting VC funding, there are two ways to end up funding 10x more startups than VCs: 1.) Make the VCs fund 10x fewer startups or 2.) Fund 10x more startups yourself. (Or various combinations, of course.) #1 seemed absurd to me, so I read the thought experiment as suggesting #2. Edit: Hmm, I guess another alternative is to fund startups that "look worse" to VCs, which is basically #1 but doesn't seem so absurd. So perhaps I was misreading the quote.
  - aptwebapps 14y ago
    Either you are or I am. ;) "We'd have to be willing to fund 10x more startups than they would." The contracted 'we would' is key. They WOULD have to be willing to do that IF they were going to pursue that goal. They aren't. He's positing the idea that if they were properly optimized for the black swans then they would see a much lower funding rate. Not that they should aim for a low funding rate for it's own sake. But, as he discusses, actually performing such an optimization is hard, not to mention the fact that making money isn't their only reason for running YC.
resbear 14y ago
I think the difference here is the assumption for failure rather than success for. The subprime bubble existed due to leverage - fancy models said that defaults were unlikely, so rather than extend loans from their own assets, banks decided to double (or triple, or quadruple...) down and lend out multiples of their own total assets. The models were wrong and they blew up. This is quite different from assuming a high failure rate. The marginal cost of having say, 10% more failed startups isn't a big deal when 95% of them are already failing, and the model accounts for that high failure rate. The black swan is not the unexpected failure that wipes you out, but the unexpected success that outweighs all the failures. Obviously, you'd still be screwed if none of the stuff you invest in pays off. But that would be the case even if your model was more pessimistic; if even 5% success was too much to ask for, perhaps the market for VC is busted, and there was no hope for profit anyway. I don't think pg is implying that screening processes are useless. I'm sure many obviously hopeless ideas get filtered out. It's more that, given the current state of knowledge, everything beyond that basic filter is by nature difficult to discern and we may as well assume a random fat tail distribution.