Deferring Well

This is a linkpost for a pair of blogposts I wrote on the skill involved in "deferring to others" You can find the originals here and here, but I've copied the text below. The second post was prompted by the fact that Coefficient Giving, where I work, is hiring, but I think the content will be useful even after the round closes, so I'm bundling them together and will keep the post up, with minor edits, after the round.

I heavily used Claude as part of the drafting process, but nearly everything in the final drafts was either written by me or significantly edited. As well as Claude, I benefitted from useful feedback from humans including (in alphabetical order) Arden, Catherine, Ed, and Neel.

Deferral is a skill

(Issue?)

Not everyone needs to be a researcher. But “not being a researcher” doesn’t mean “not thinking for yourself”.

I think the skill involved in “secondary research” / “deferring well” / [whatever the right label is, part of the problem might be a lack of a good label] is frequently underrated.

It’s easy to carve people up into “thinkers who can and should develop inside views” or “people who defer.” But the second camp is massive and heterogeneous; many tasks/jobs/roles involve “just” deferring in the sense that nearly all of the relevant facts/theories/ideas don’t need to be generated de novo by the person doing the job, but where the skill+effort that gets put into trying to form true beliefs is still pretty much the whole ball game.

• Often, especially in small/new fields, there isn’t a consensus you can just identify and defer to
• Even when there is a consensus on a specific question, there’s a big gap between “defer to the consensus” and “understand what the consensus actually implies for the decision you’re actually facing”
• If the plan is to defer to a specific expert on something, how good you are at extracting relevant info from them matters a lot: are you asking good questions, do you need things explained twice, can you work out what to ask them vs. find out yourself?
  • This matters because experts in a subject area aren’t always experts at explaining stuff
  • And because how easy it is for the experts to convey what they need to to you affects how much future facetime you’ll get.
• Even the literal best thinker in the world on topic X isn’t necessarily going to be that hard to beat at making a specific decision where topic X is relevant but not the entire story, especially if they only have a moment to think about it.

This is especially tricky in fields like AI safety and governance, where the experts genuinely disagree, the field is moving fast, and the “consensus view” on many important questions just doesn’t exist. You can’t defer to a consensus that isn’t there.

I had to do a lot of [this thing, where I’m trying really hard to work out what’s true and what I actually think, but not “doing novel research”] both as an advisor at 80k and now as a grantmaker. In both jobs it wasn’t my role to produce new knowledge, but there was still a huge amount of original thinking required. The decisions I face (”what should this person/org do?”) involve deep understanding of areas where I’m not an expert, meaning that relying on others’ expertise is extremely important. It’s almost never the case that I can just ask the relevant expert the answer to the actual question I’m facing.

I need to understand something close to cutting-edge knowledge, but don’t need to produce it. Instead I need to quickly work out which things are relevant to my decision, get to the truth of them (often by asking), work out how much variation in belief exists among genuine experts, and how sensitive the decision is to that variation etc.

Whether you can think of a better label for the thing I’m trying to describe, or just want to call it “deferring, but, like, the high skill kind of deferring”, don’t underrate it, in yourself or others.

More on Deferral

And we're hiring

I’m looking for people with technical backgrounds in ML, hardware, or information security to join the AI Governance and Policy team at Coefficient Giving. If you¹ have such a background, and you’ve never seriously considered grantmaking, or if your instinct is that your technical skills wouldn’t be especially helpful, this post is an attempt to change your mind. We’re also hiring for a wide variety of other grantmaker and generalist roles across our Global Catastrophic Risks (GCR) teams.

We expect to move around $1 billion in GCR grants this year, a significant step up from last year, and we expect that figure to keep growing. Trying to make sure this happens has prompted big internal changes aimed at helping us get faster, be less bureaucratic, and push for more ambition on individual grants. I felt the effects of this viscerally² on returning from parental leave a few weeks ago.

But however good we get on speed and ambition, we can only say “yes” to what comes in. And a lot of what I think matters most – work where currently the world is completely dropping the ball – isn’t coming in. We have to go out and create those opportunities. That takes a lot more time and deeper understanding to prioritise.

Which means our team is much more capacity-constrained than funding-constrained. And we’re particularly constrained on technical capacity.

A few months ago I wrote Deferral is a skill, because I thought people were underrating how much serious thinking is involved in working out what’s true even if your sources are other people rather than original research. A lot of that thinking is involved in grantmaking, and helped by the kinds of background I’m looking for, which I’ll try to articulate below.

I’m never going to be the world’s top expert on the thing I’m funding, but I’m usually going to have access to people who are.

Unfortunately, there often isn’t an obvious person to ask, and sometimes there isn’t an obvious question. Even when an expert can give you a clear answer on the technical point, it’s on you to work out what that means for the actual decision you’re facing.

I’ve had questions where tens of millions of dollars were on the line, I asked two of the world’s top ~five experts in the relevant domain, and I then got contradictory answers. You can’t just shrug, pick the expert you like most, and move on. You have to actually try to develop a view. Listen to the mechanistic arguments people give, evaluate whether those arguments hold up, push back where they don’t, and then evaluate the responses to your pushback. All of that is far easier to do when you’re fluent in the material.

Even if there’s a clear consensus, or a single best-placed expert who you can ask questions of³, you have to be able to notice when an expert has answered a nearby question rather than the one you needed⁴.

Holding the map

To the extent that grantmakers do end up “world class” in some technical domain, it’s in something like “holding a detailed, opinionated map of a field”. People working on one part of the terrain will know that part much better than I ever will. But it’s not their job to track every adjacent effort, notice persistent gaps, compare five proposals trying to solve nearby problems, or work out which two groups should probably be talking to each other.

The best grantmakers will have a live, gears-level model of the area they work in, against which any specific project can be evaluated, and within which gaps can be noticed and filled. This is a hard thing to do, and without enough depth, you risk just tracking the legible social surface instead. Sometimes that’s useful. Often it isn’t enough.

With a detailed map, deferral itself gets easier. If I understand an area well, I can get more out of a half-hour call with someone working in it, and form a better view of how good a project’s outputs actually are. If I’m really engaging with the views of the experts I call on, I can quickly build a picture of when and how I should just fully take their answer as gospel.

If you like learning across a range of different areas, including by literally talking to some of the world’s foremost experts about the questions you most want answered – this role gives you an unusual amount of that. You get to speedrun learning something up to near expert level, and then move on to something else.

I find this very much my thing. Most of the people I know who are good at it find it very much theirs. But not everyone is going to have a good time with it.

If you love going as deep as possible into one specific technical area, and you’d hate spending much of your time one step removed from the object-level work, you probably wouldn’t like this role. If jumping between topics, and spending nearly all of your time talking to people about stuff that they’ve spent way longer thinking about than you sounds hard, or frustrating, or stressful, you aren’t going to have a good time.

Similarly, if your reaction to noticing that there are a bunch of things which need to happen is “I’m going to pick one of those things, and fix it myself”, the kind of gap-filling we’re likely to do is probably not going to appeal. We want to make things happen, and that involves a lot more agency than one might expect, but it’s still going to look more like scoping a project, finding someone to run it, and then giving them a ton of resources, than taking on the project yourself.

If you’re on the fence, apply. When I was advising at 80,000 Hours I’d often tell people that application processes are a great way to get information about whether a role is right for you, and I think ours is a very good simulation of (parts of) the job. If you get through the early stages, I’d be excited to talk through what the work tests do and don’t capture, and help you figure out fit. I’m worried that we won’t fill these roles, and I want to find great people for them, but we’re all on the same team, and if you’d do more good elsewhere, I’d rather help you figure that out than hire you.

Round closes 11:59pm Pacific Time, 24 May 2026. Apply here.