One general point: My rough guess is that acceptance rates have stayed largely constant across AI safety programs over the last ~2 years because capacity has scaled with interest. For example, Pivotal grew from 15 spots in 2024 to 38 in 2025. While the ‘tail’ likely became more exceptional, my sense is that the bar for the marginal admitted fellow has stayed roughly the same.
They might (as I am) be making as many applications as they have energy for, such that the relevant counterfactual is another application, rather than free time.
The model does assume that most applicants aren’t spending 100% of their time/energy on applications. However, even if they were, I feel like a lot of this is captured by how much they value their time. I think that the counterfactual of how they spend their time during the fellowship period (which is >100x more hours than the application process) is the much more important variable to get right.
you also need to consider the intangible value of the counterfactual
This is correct. I assumed most people would take this into account (e.g. subtract their current job’s networking value from the fellowship’s value), but I might add a note to make this explicit.
you also ought to consider the information value of applying for whatever else you might have spent the time on
I’m less worried about this one. Since we set the fixed Value of Information quite conservatively already, and most people aren’t constantly working on applications, I suspect this is usually small enough to be noise in the final calculation.
there is a psychological cost to firing out many low-chance applications
I agree this is real, but I think it’s covered in the Value of Your Time. If you earn £50/hr but find applying on the weekend fun/interesting, you might set the Value of Your Time at £5/hr. If you are unemployed but find applying extremely aversive, you might price your time at e.g., £200/hr.
Thanks a lot for engaging!
One general point: My rough guess is that acceptance rates have stayed largely constant across AI safety programs over the last ~2 years because capacity has scaled with interest. For example, Pivotal grew from 15 spots in 2024 to 38 in 2025. While the ‘tail’ likely became more exceptional, my sense is that the bar for the marginal admitted fellow has stayed roughly the same.
The model does assume that most applicants aren’t spending 100% of their time/energy on applications. However, even if they were, I feel like a lot of this is captured by how much they value their time. I think that the counterfactual of how they spend their time during the fellowship period (which is >100x more hours than the application process) is the much more important variable to get right.
This is correct. I assumed most people would take this into account (e.g. subtract their current job’s networking value from the fellowship’s value), but I might add a note to make this explicit.
I’m less worried about this one. Since we set the fixed Value of Information quite conservatively already, and most people aren’t constantly working on applications, I suspect this is usually small enough to be noise in the final calculation.
I agree this is real, but I think it’s covered in the Value of Your Time. If you earn £50/hr but find applying on the weekend fun/interesting, you might set the Value of Your Time at £5/hr. If you are unemployed but find applying extremely aversive, you might price your time at e.g., £200/hr.