Arepo's Quick takes

Arepo

Arepo's Quick takes

Arepo

1 min readSep 21, 2024

Community

Frontpage

Comments 10

Sorted by

New & upvoted

Arepo

Suggested hiring practice tweak

There are typically two ways for organisations of running hiring rounds: deadlined, in which job applications are no longer processed after a publicised date, and rolling in which the organisation will keep allowing submissions until they've found someone they want.

The upside of a deadline is both to an applicant that they know they're not wasting their time on a job that's 99% assigned, and to the organisation, which doesn't have to delay giving an answer to an adequate candidate on the grounds that a potentially better one submits when you're most of the way through the hiring process, and incentivises people to apply slightly earlier than they would have.

The downsides are basically the complement. The individual doesn't get to go for a job that they've just missed and would be really suited to, and the org doesn't get to see as large a pool of applicants.

It occurred to me that an org might be able to get some of the best of both by explicitly giving a mostly-deadline, after which they will explicitly downweight new applications. So if you see the mostly-deadline in time, you're still incentivised to get your application in by the date given, and if it's passed you should rationally apply if and only if you think there's a good chance you're an exceptional fit..

Arepo

1y*

One of the problems with AI benchmarks is that they can't effectively be backcast more than a couple of years. This prompted me to wonder if a more empirical benchmark might be something like 'Ability of a human in conjunction with the best technology available at time t'.

For now at least, humans are still necessary to have in the loop, so this should in principle be at least as good as coding benchmarks for gauging where we are now. When/if humans become irrelevant, it should still work - 'AI capability + basically nothing' = 'AI capability'. And looking back, it gives a much bigger reference class for forecasting future trends, allowing us to compare e.g.

Human
Human + paper & pen
Human + log tables + paper & pen
Human + calculator + log tables + paper & pen
Human + computer with C + ...
Human + computer with Python + ...
Human + ML libraries + ...
Human + GPT 1 + ...

etc.

Thoughts?

Thomas Kwa🔹

1y*

One problem is putting everything on a common scale when historical improvements are so sensitive to the distribution of tasks. A human with a computer with C, compared to a human with just log tables, is a billion times faster at multiplying numbers but less than twice as fast at writing a novel. So your distribution of tasks has to be broad enough that it captures the capabilities you care about, but it also must be possible to measure a baseline score at low tech level and have a wide range of possible scores. This would make the benchmark extremely difficult to construct in practice.

Arepo

1y*

I think that's right, but modern AI benchmarks seem to have much the same issue. A human with a modern Claude instance might be able to write code 100x faster than without, but probably less than 2x as fast at choosing a birthday present for a friend.

Ideally you want to integrate over... something to do with the set of all tasks. But it's hard to say what that something would be, let alone how you're going to meaningfully integrate it.

Arepo

9mo*

Somehow despite 'Goodharting' being a now standard phrase, 'Badharting' is completely unrecognised by Google.

I suggest the following intuitive meaning: failing to reward a desired achievement, because the proxy measure you used to represent it wasn't satisfied:

'No bonus for the staff this year: we didn't reach our 10% sales units growth target.'

'But we raised profits by 30% by selling more expensive products, you Badharting assholes!'

Arepo

12mo*

I wonder if it would be interesting to use PPP-adjusted sale price of high-end luxury or Veblen goods as a metric for moral progress of humanity (I suspect on this metric, we'd look like we were getting progressively worse, but I'm unclear if that's a bug or a feature).

JuliaHP

12mo

To clarify, do you mean that high consumption of Veblen goods would imply worse morality?

Arepo

12mo

Yeah, it's a relatively objective expression of the ratio of how much the most successful people in society care about tiny benefits to themselves vs any alternative.

Arepo

1y*

Suppose we compare two nonprofit orgs doing related work. Let’s use some real examples: Rethink Priorities and Founders Pledge, both of who do global health and climate change research; CEA (who run EAGs) and any community groups who run EAGxes or otherwise movement build; perhaps CFAR and Khan Academy.

Ideally, in an effectiveness-minded movement, every donation to one of them rather than the other should express some view on the relative capability of that org to execute its priorities - it is essentially a bet that that org will make better use of its money.

We can use a simple combinatorial argument to show that the epistemic value of this view rapidly approaches 0, the more things either or both of those organisations are doing. If AlphaOrg does only project A₁, and BetaOrg does only project B₁ (and for the sake of simplicity, that both projects are have the same focus) then donating to Alphaorg clearly shows that you think Alphaorg will execute it better - that A₁ > B₁.

But if AlphaOrg adds a single (related or unrelated) project, A₂, to their roster, the strength of the signal drops to 1/6th: now in donating to AlphaOrg, I might be expressing the view that A₁> B₁> A₂, that A₂> B₁> A₁, that A₁> A₂ > B₁, or that A₂> A₁ > B₁, or (if I think the lesser projects sum to more than the greater one), that B₁ >A₂ > A₁or B₁ >A₁ > A₂.

In general, the number of possible preference orderings we can have between the projects of just two orgs respectively running m and n projects between them is (m + n)![^end] (meaning 3*2=6 for three, 4*6=24 for four, 5*24=120 for five, and so on). So if I give to AlphaOrg, my donation is consistent with (m + n)! - (m! * n!) possible preferences, where (m! * n!) is the number of preferences for all BetaOrg projects over all AlphaOrg projects, which my donation rules out.

In general, for k orgs in a comparable space, a donation to AlphaOrg with x1 projects is (based on some hasty interpretation) consistent with is (x1 + x2 + ... xk)! - x1! * (x2 + x3 + ... + xk)! preference orderings.

Assuming a typical EA org receives money from a couple of hundred donors a year, each of which we might consider a ‘vote’ or endorsement, that means on naive accounting (where we divide votes by preference orderings), as few as 6 projects between two relevant orgs give us less than a single endorsement’s worth of info on which of their projects effectiveness-minded donors actually support.

Obviously there are other considerations. Reduced administrative burden from combining is perhaps the foremost; also major donations can be restricted, somewhat mimicking the effect of donating to a more focused org (though if the org also receives unrestricted donations, it can undo this effect by just redirecting the unrestricted money) ; also one might want a very strong team to expand their focus on priors - though doing so would strongly reduce our confidence that they remain a strong team for the expanded purpose.

Nonetheless, with the central EA orgs typically having at least 3 or 4 focus areas each (giving ~40320 possible preference orderings between two of them), and more if you count in-house support work - software development, marketing etc - as separate projects, I think the magnitude of this epistemic cost is something a purportedly effectiveness-minded and data-driven movement should consider very seriously.

Jason

Some of the combinatoric effect here is picking up on the number of projects, I think? If you have three projects in three separate orgs, your vote to fund one conveys information on which you rank first but not the rank order between the other two. If you have ten orgs with one project each, there are I think 36 pairwise comparisons and a first-place donation vote addresses only nine of them.

More generally: it may be worthwhile to distinguish more between donation/vote as an information mechanism and as an influence mechanism. It's plausible to me that other features of the ecosystem could significantly impair both the potential informational and influential power of "votes" before we even got to considering the issue you describe here.

Donation/vote as a donor influence mechanism has some significant limitations in an ecosystem where the bulk of the funding comes from a few megadonors. To the extent that smaller donors think their donations funge with those of the megadonors, and that megadonors are more capable of adjusting to enact their global preferred funding allocations, the small donors may not believe that their votes have any meaningful influence on overall funding allocations. To the extent that smaller donors believe that, I expect the belief would have an significant effect on small-donor willingness to invest in casting informed "votes." So it may seriously affect the epistemic / informational value of the votes too.

You could get some of the informational effect by merely asking donors to identify the specific program they'd like to donate to in a non-binding fashion. Of course, the advisory nature of the project-specific vote would likely make donors less willing to spend time on casting informed votes.

Comments