Great post, thanks for writing it! This framing appears a lot in my thinking and it’s great to see it written up! I think it’s probably healthy to be afraid of missing a big multiplier.
I’d like to slightly push back on this assumption:
If output scales linearly with work hours, then you can hit 60% of your maximum possible impact with 60% of your work hours
First, I agree with other commenters and yourself that it’s important not to overwork / look after your own happiness and wellbeing etc.
Having said that, I do think working harder can often have superlinear returns, especially if done right (otherwise it can have sublinear or negative returns). One way to think about this is that the last year of one’s career is often the most impactful in expectation, since one will have built up seniority and experience. Working harder is effectively a way of “pulling that last year forward a bit” and adding another even higher impact year after it. I.e. a year that is much higher-impact than your average year, hence the superlinearity.
Another way to think about this is intuitively. If Sam Bankman-Fried had only worked 20% as hard, would he have made $4 billion instead of $20 billion? No. He would probably have made much much less. Speed is rewarded in the economy and working hard is one way to be fast.
This makes the multiplier from working harder bigger than you would intuitively expect and possibly more important relative to judgment than you suggest.
(I’m not saying everyone reading this should work harder. Some should, some shouldn’t.)
Edited shortly after posting to add: There’s also a more straightforward reason that the claim “judgment is more important than dedication” is technically true but potentially misleading: one way to get better judgment is investing time into researching thorny issues. That seems to be what Holden Karnofsky has been doing for a decent fraction of his career.
A key question for whether there are strongly superlinear returns seems to be the speed at which reality moves. For quant trading and crypto exchanges in particular, this effect seems really strong, and FTX’s speed is arguably part of why it was so successful. This likely also applies to the early stages of a novel pandemic, or AI crunch time. In other areas (perhaps, research that’s mainly useful for long AI timelines), it may apply less strongly.
I agree that superlinearity is way more pronounced in some cases than in others.
However, I still think there can be some superlinear terms for things that aren’t inherently about speed. E.g. climbing seniority levels or getting a good reputation with ever larger groups of people.
The examples you give fit my notion of speed—you’re trying to make things happen faster than the people with whom you’re competing for seniority/reputation.
Similarly, speed matters in quant trading not primarily because of real-world influence on the markets, but because you’re competing for speed with other traders.
Fair, that makes sense! I agree that if it’s purely about solving a research problem with long timelines, then linear or decreasing returns seem very reasonable.
I would just note that speed-sensitive considerations, in the broad sense you use it, will be relevant to many (most?) people’s careers, including researchers to some extent (reputation helps doing research: more funding, better opportunities for collaboration etc). But I definitely agree there are exceptions and well-established AI safety researchers with long timelines may be in that class.
FWIW I think superlinear returns are plausible even for research problems with long timelines, I’d just guess that the returns are less superlinear, and that it’s harder to increase the number of work hours for deep intellectual work. So I quite strongly agree with your original point.
Great post, thanks for writing it! This framing appears a lot in my thinking and it’s great to see it written up! I think it’s probably healthy to be afraid of missing a big multiplier.
I’d like to slightly push back on this assumption:
First, I agree with other commenters and yourself that it’s important not to overwork / look after your own happiness and wellbeing etc.
Having said that, I do think working harder can often have superlinear returns, especially if done right (otherwise it can have sublinear or negative returns). One way to think about this is that the last year of one’s career is often the most impactful in expectation, since one will have built up seniority and experience. Working harder is effectively a way of “pulling that last year forward a bit” and adding another even higher impact year after it. I.e. a year that is much higher-impact than your average year, hence the superlinearity.
Another way to think about this is intuitively. If Sam Bankman-Fried had only worked 20% as hard, would he have made $4 billion instead of $20 billion? No. He would probably have made much much less. Speed is rewarded in the economy and working hard is one way to be fast.
This makes the multiplier from working harder bigger than you would intuitively expect and possibly more important relative to judgment than you suggest.
(I’m not saying everyone reading this should work harder. Some should, some shouldn’t.)
Edited shortly after posting to add: There’s also a more straightforward reason that the claim “judgment is more important than dedication” is technically true but potentially misleading: one way to get better judgment is investing time into researching thorny issues. That seems to be what Holden Karnofsky has been doing for a decent fraction of his career.
A key question for whether there are strongly superlinear returns seems to be the speed at which reality moves. For quant trading and crypto exchanges in particular, this effect seems really strong, and FTX’s speed is arguably part of why it was so successful. This likely also applies to the early stages of a novel pandemic, or AI crunch time. In other areas (perhaps, research that’s mainly useful for long AI timelines), it may apply less strongly.
I agree that superlinearity is way more pronounced in some cases than in others.
However, I still think there can be some superlinear terms for things that aren’t inherently about speed. E.g. climbing seniority levels or getting a good reputation with ever larger groups of people.
The examples you give fit my notion of speed—you’re trying to make things happen faster than the people with whom you’re competing for seniority/reputation.
Similarly, speed matters in quant trading not primarily because of real-world influence on the markets, but because you’re competing for speed with other traders.
Fair, that makes sense! I agree that if it’s purely about solving a research problem with long timelines, then linear or decreasing returns seem very reasonable.
I would just note that speed-sensitive considerations, in the broad sense you use it, will be relevant to many (most?) people’s careers, including researchers to some extent (reputation helps doing research: more funding, better opportunities for collaboration etc). But I definitely agree there are exceptions and well-established AI safety researchers with long timelines may be in that class.
FWIW I think superlinear returns are plausible even for research problems with long timelines, I’d just guess that the returns are less superlinear, and that it’s harder to increase the number of work hours for deep intellectual work. So I quite strongly agree with your original point.