The basic claim of this post is that Paul Graham has written clearly and well about unlearning the desire to do perfectly on tests, but that his actions are incongruous, because he has built the organization that most encourages people to do perfectly on tests.
Not that he has done no better – he has done better than most – but that he is advertising himself as doing this, when he has instead probably just made much better tests to win at.
Sam Altman’s desire to be a monopology
On tis the post offers quotes giving evidence saying:
YC is a gatekeeper to funding and a broader network of valuable supplies
Sam Altman ambitiously wants YC to be the primary funder globally of good companies (and this could imply the sole gatekeeper)
YC creating tests
The post says that the natural way to find such people would be proactive talent scouting rather than creating a formal test, and judges YC for not doing this, and claims that the test filters for people who are obsessed with passing tests.
Here are the points made by the post, in order:
One factor they care about is extreme responsiveness. The post points out that if you are to goodhart on this metric strongly enough, it will become ‘performing’ responsiveness.
The post also quotes the YC CEO (Sam Altman) saying that the primary type of person they select for is smart, upper-middle-class people, which is the set the post thinks is most likely to have the desire to do perfectly on tests.
The post also Altman talking about the desire to maximise numbers regarding health, and then also quotes Hotel Concierge talking about a time when maximising the numbers was to the clear detriment of the reality and their personal health, suggesting that Altman is selecting for people who maximise at the expense of reality rather than in accordance with it.
The next section is about how YC forces the founders in their program to do this, to be poor and to make their companies profitable enough to earn food and living.
The post points out the obsession with growth can be goodharted on in many ways, and points out that one company advertised “fifty-per-cent word-of-mouth growth” which sounds like a straightforward nonsense metric unrelated to building a great product, created by someone who wanted to show growth.
If I were to abstract this a bit, I’d say that if you goodhart on YC’s metrics and tests, you will be able to pass them yet keep the desire to do perfectly on tests, and there is suggestive evidence that this has occured.
I think they should be praised for having built better tests. Much of society is about building better metrics to optimize for, and then when we have goodharted on them, learning from our mistakes and making better ones.
Related: I am reminded of Zvi’s Less Competition, More Meritocracy. That post talks about how, if the pool of selection gets sufficiently big, the people being selected on are encouraged to take riskier strategies to pass the filters, and the selection process breaksdown. It seems plausible to me that YC substantially changed as an organism at a certain level of success, where initially nobody cared and so the people who passed the tests were very naturally aligned people (the founders of AirBnb, DoorDash, Stripe) but as the competition increased, the meritocracy decreased.
Psychologizing Paul Graham
The post gives some arguments psychologizing Paul Graham:
The post says that Paul Graham has deceived himself on whether mean people fail, because many mean people succeed.
It also points out that Paul Graham does not follow his own advice when it comes to funding companies, because (as he says) he would find it psychologically intolerable.
It argues that “this is a case of fooling oneself to avoid confronting malevolent power”.
I think that this has some validity. My sense is that Paul Graham has made a lot of succes out of production-based strategies, and has somewhat blinded himself to the existence of adversarial strategies. He seems to me much less politically minded than other very productive people like Bezos and Musk, who I think have engaged much more directly with politics and still succeeded.
I also think that Sam Altman’s expressed desire to be something of a monopoly is not something that Paul Graham has engaged with in his writing, I think, and that this would bring with it many political and coordination issues he has not addressed and that could be harmful for the world.
Conclusion
I don’t think it’s bad for YC to have tests, even tests that are goodharted on to an extent, but I don’t think the post is actually that interested in YC. It’s more interested in the phenomenon of action being incongruous with speech.
The post is a ton of primary sources followed by a bit of psychologizing of Paul Graham. It’s often impolite to publicly psychologize someone. I generally try to emphasize lots of true and positive things about the person when I do so, to avoid the person feeling attacked, but this post didn’t choose to do that, which is a fine choice. Either way, I think it is onto something when it talks about Paul Graham being somewhat blind to adversarial and malevolent forces that can pass his tests. If I wanted to grade Paul Graham overall, I feel like this post is failing to properly praise him for his virtues. But the post isn’t trying to grade him overall, instead its focus is on the gap between his speech and his actions, and analyzes what’s going on there.
I do feel like there’s much more to be said there, many more essays, but the post does say quite valuable things about this topic. I feel like I’d get a lot of returns from this post being more fleshed out (in ways I discuss below) and elaborating more on its ideas. In its current form, I’ll probably give it a +1 or +2 in the review. I generally found it hard to read, but worthwhile.
Further work
Here are some more questions I’d like to see discussed and answered:
How did YC change over time with respect to optimizing for production-strategies and adversarial ones?
(I can imagine the answers here being (1) whenever YC became successful enough for the news media to notice it and make it prestigious, and (2) when Sam Altman became president.)
How much do the more recent YC companies care about production vs growth?
Seems like the best YC companies are the early ones, but they naturally have an age-advantage. I’ve heard rumours but would be interested in more evidence here.
I’d also be interested in a more fleshed out version of much of the discussion in this post. “This is a case of fooling oneself to avoid confronting malevolent power” → what are other cases in the modern world, and what are some of the forces at play here? “If, to participate in higher growth rates, you have to turn into something else, then in what sense is it you that’s getting to grow faster?” → this is a great point, and i’m interested in where exactly things like YC made the decision to turn into something else, and what that looked like. “He then sets up an institution optimizing for “success” directly, rather than specifically for production-based strategies.” → I’d be interested in a more detailed sketch of what that organisation would look like.
I also think that there’d be some good work in tying this into the ontology of goodhart’s law.
The basic claim of this post is that Paul Graham has written clearly and well about unlearning the desire to do perfectly on tests, but that his actions are incongruous, because he has built the organization that most encourages people to do perfectly on tests.
Not that he has done no better – he has done better than most – but that he is advertising himself as doing this, when he has instead probably just made much better tests to win at.
Sam Altman’s desire to be a monopology
On tis the post offers quotes giving evidence saying:
YC is a gatekeeper to funding and a broader network of valuable supplies
Sam Altman ambitiously wants YC to be the primary funder globally of good companies (and this could imply the sole gatekeeper)
YC creating tests
The post says that the natural way to find such people would be proactive talent scouting rather than creating a formal test, and judges YC for not doing this, and claims that the test filters for people who are obsessed with passing tests.
Here are the points made by the post, in order:
One factor they care about is extreme responsiveness. The post points out that if you are to goodhart on this metric strongly enough, it will become ‘performing’ responsiveness.
The post also quotes the YC CEO (Sam Altman) saying that the primary type of person they select for is smart, upper-middle-class people, which is the set the post thinks is most likely to have the desire to do perfectly on tests.
The post also Altman talking about the desire to maximise numbers regarding health, and then also quotes Hotel Concierge talking about a time when maximising the numbers was to the clear detriment of the reality and their personal health, suggesting that Altman is selecting for people who maximise at the expense of reality rather than in accordance with it.
The next section is about how YC forces the founders in their program to do this, to be poor and to make their companies profitable enough to earn food and living.
The post points out the obsession with growth can be goodharted on in many ways, and points out that one company advertised “fifty-per-cent word-of-mouth growth” which sounds like a straightforward nonsense metric unrelated to building a great product, created by someone who wanted to show growth.
If I were to abstract this a bit, I’d say that if you goodhart on YC’s metrics and tests, you will be able to pass them yet keep the desire to do perfectly on tests, and there is suggestive evidence that this has occured.
I think they should be praised for having built better tests. Much of society is about building better metrics to optimize for, and then when we have goodharted on them, learning from our mistakes and making better ones.
Related: I am reminded of Zvi’s Less Competition, More Meritocracy. That post talks about how, if the pool of selection gets sufficiently big, the people being selected on are encouraged to take riskier strategies to pass the filters, and the selection process breaksdown. It seems plausible to me that YC substantially changed as an organism at a certain level of success, where initially nobody cared and so the people who passed the tests were very naturally aligned people (the founders of AirBnb, DoorDash, Stripe) but as the competition increased, the meritocracy decreased.
Psychologizing Paul Graham
The post gives some arguments psychologizing Paul Graham:
The post says that Paul Graham has deceived himself on whether mean people fail, because many mean people succeed.
It also points out that Paul Graham does not follow his own advice when it comes to funding companies, because (as he says) he would find it psychologically intolerable.
It argues that “this is a case of fooling oneself to avoid confronting malevolent power”.
I think that this has some validity. My sense is that Paul Graham has made a lot of succes out of production-based strategies, and has somewhat blinded himself to the existence of adversarial strategies. He seems to me much less politically minded than other very productive people like Bezos and Musk, who I think have engaged much more directly with politics and still succeeded.
I also think that Sam Altman’s expressed desire to be something of a monopoly is not something that Paul Graham has engaged with in his writing, I think, and that this would bring with it many political and coordination issues he has not addressed and that could be harmful for the world.
Conclusion
I don’t think it’s bad for YC to have tests, even tests that are goodharted on to an extent, but I don’t think the post is actually that interested in YC. It’s more interested in the phenomenon of action being incongruous with speech.
The post is a ton of primary sources followed by a bit of psychologizing of Paul Graham. It’s often impolite to publicly psychologize someone. I generally try to emphasize lots of true and positive things about the person when I do so, to avoid the person feeling attacked, but this post didn’t choose to do that, which is a fine choice. Either way, I think it is onto something when it talks about Paul Graham being somewhat blind to adversarial and malevolent forces that can pass his tests. If I wanted to grade Paul Graham overall, I feel like this post is failing to properly praise him for his virtues. But the post isn’t trying to grade him overall, instead its focus is on the gap between his speech and his actions, and analyzes what’s going on there.
I do feel like there’s much more to be said there, many more essays, but the post does say quite valuable things about this topic. I feel like I’d get a lot of returns from this post being more fleshed out (in ways I discuss below) and elaborating more on its ideas. In its current form, I’ll probably give it a +1 or +2 in the review. I generally found it hard to read, but worthwhile.
Further work
Here are some more questions I’d like to see discussed and answered:
How did YC change over time with respect to optimizing for production-strategies and adversarial ones?
(I can imagine the answers here being (1) whenever YC became successful enough for the news media to notice it and make it prestigious, and (2) when Sam Altman became president.)
How much do the more recent YC companies care about production vs growth?
Seems like the best YC companies are the early ones, but they naturally have an age-advantage. I’ve heard rumours but would be interested in more evidence here.
I’d also be interested in a more fleshed out version of much of the discussion in this post. “This is a case of fooling oneself to avoid confronting malevolent power” → what are other cases in the modern world, and what are some of the forces at play here? “If, to participate in higher growth rates, you have to turn into something else, then in what sense is it you that’s getting to grow faster?” → this is a great point, and i’m interested in where exactly things like YC made the decision to turn into something else, and what that looked like. “He then sets up an institution optimizing for “success” directly, rather than specifically for production-based strategies.” → I’d be interested in a more detailed sketch of what that organisation would look like.
I also think that there’d be some good work in tying this into the ontology of goodhart’s law.