Why AI Safety Camp struggles with fundraising (FBB #2)

Crossposted on The Field Building Blog.

Epistemic status: I don't have experience in grantmaking so take all this with a grain of salt. I don’t claim that I’m saying anything that AI Safety Camp’s organisers haven’t considered before, or that there aren’t other reasons why it’s hard to fundraise for AISC. I tried to put myself in the shoes of (potential) donors and think about what their concerns might be. I also asked a few knowledgeable people to weigh in.

AI Safety Camp is the project that everyone knows, almost everyone loves, yet it remains underfunded. Why is that?

I have written up my thoughts about this situation, hopefully, it’s going to start a useful conversation in the community. I also share some easy and not-so-easy fixes that I think would make their case easier.

To get this out of the way: I think AISC has an excellent track record and has been able to do a lot with very little. In my opinion, their funding issue is not caused by their lack of impact. It’s sort of everything else. You could make the case that impact is the only thing that should matter, but as you will see below, it’s not that simple. If I had to summarise the rest of this post in one sentence, I would say they don’t do a good job of framing and structuring their organisation so that it appeals to various funders.

Reasons it’s hard to donate to AISC for (some) people:

1. Organisational structure

AISC has a very broad focus. This is an issue in itself, but their current organisational structure does not allow for funders to pay for different types of work, let alone specific projects. I can imagine many people wanting to fund technical research, but not advocacy, and vice versa.

I raised this issue to Remmelt, and in response, he wrote:

So I think we wanted to prevent a dynamic where people split into factions each raising funds for their own faction. For lack of a better word, that breaks down the ‘collaborative spirit’ at the camp. [...] I would not feel comfortable though with eg. raising funds in a way that could lead to Robert receiving a lower pay than me, or the other way around. Each of us organisers spend at least half of our time doing general communications and ops for the program, so earmarking pay to us by the area we oversee does not make much sense anyway.

I empathise with where Remmelt is coming from, and I agree that restricted funds have downsides^[1] and might be just annoying to deal with. Still, this is a point they should concede (or at least try once) if the alternative is for AISC not to be funded.

If AMF were doing both bednets and cataract surgeries, the case for restricted funding is understandable. If they were doing vaccines and bednets, it would be understandably annoying.^[2] Despite having a similar format, I think AISC’s technical research vs. advocacy projects are more similar to the former comparison. In an ideal world, AISC would be able to secure funding to run the operations of both types of projects, albeit from different funders.^[3] To be able to achieve that, they need to make more specific funding requests for specific types of work.

It’s also worth mentioning that for some people, such as those working in governments, it might be risky to be a research lead at an org that also hosts projects working to pause AI or grassroots advocacy.

A weaker version of separation would be earmarking stipends for different (types of) projects. This is easier to do than splitting the organisation, but later I will point out why I think stipends are unnecessary in the first place.

2. Leadership’s image

I talked to Remmelt a couple of times and from what I could tell he is a competent person who has thought deeply about his views. However, some of the public stances he has taken make it difficult for grantmakers to associate themselves with him. Even if OP were otherwise very excited to fund AISC, it would be political suicide for them to do so. (Update: see Eli's comment below.) They can’t even get away with funding university clubs.^[4] Unless donors don’t care about optics at all, paying Remmelt’s salary is a difficult ask. I think this further makes the case for (1), where Remmelt could take on leading the advocacy parts of the project or fully spin out those into a separate org.

3. Nonessential stipends

AISC has a huge funding range of 15k USD to 300k. On the upper end, the major driver of the costs are stipends, as opposed to operational costs.

To get this out of the way, I’m someone who benefited from the ~900 USD stipend they offered last camp,^[5] but I think they would have been better off if they used the money to increase their runway. At this point, SPAR has sufficiently demonstrated that part-time research programs don’t need stipends to succeed. AISC also had iterations without stipends.

I’m not saying paying participants or research leads wouldn’t help, but it’s likely much less cost-effective than spending on the organisers’ salaries. I’m also not saying that projects should have a minimum budget. For example, SPAR offers several thousand worth of compute credits to teams that need it, which I think is awesome.

Fieldbuilding is hard, but one thing that makes it easier is that if you are doing what I call “phase two communitybuilding”^[6], ie. organising programs for people who already decided that they want to work in AI Safety, you have an almost endless supply of highly motivated and intelligent (mainly early career) people.

Given this, you should be able to find plenty of people who are willing to take part in the camp for free. Yes, for some people it really does make a difference to receive a ~1000 USD. In AISC’s current (2025) iteration, they only offer stipends for people from low-income countries. In other rounds such as last year, it was more expansive. I’m not sure if this “opt-in but otherwise no rejections”^[7] policy selects these people well enough. If they don’t have the capacity to be able to evaluate stipend requests more rigorously, I think it’s better to be more frugal.

And yes, it totally sucks if some people who are less privileged couldn’t take part in the camp, the same way it sucks for someone to struggle to pay for rent in the world. But in AI safety we have limited resources and offering a small stipend won’t necessarily solve this anyway.

Ultimately, to be able to produce better projects, they could instead focus on increasing the prestige of the organisation, which I will cover next.

4. The bar is likely too low for project leads

AISC has a hit-based approach, meaning that even if most projects don’t end up producing much value^[8], those couple of home runs make it worth paying for the whole package. This is well understood by grantmakers.

I want to flag here I haven’t spent time studying the different types of projects they put on, but enough knowledgeable people referred to this as an issue that I feel fine deferring to them. I recommend making up your mind on this yourself, by looking at their requirements for project leads and the projects they put on this year.

Taking on additional projects is going to cost money, even if they get efficiency gains from scaling.

The bigger issue I think, is that the worst projects make AISC less prestigious. If their bar is too low, that makes some of the most competent potential research leads reluctant to associate themselves with the organisation. I understand not wanting to be an elitist, but I think this level of inclusivity should be reserved for people who are earlier in the funnel, where I agree that you should give everyone a chance if you can. You might say that if the alternative for these people is to be rejected and not do anything, then it’s still a net positive for the project to happen for the sake of upskilling.

However, there are some great programs out there. Opportunities such as ARENA are hard to get into, but the curriculum is available online. Perhaps AISC could have people apply to lead such upskilling projects on vetted curriculums from the start, so they can offer something to both rejected research leads and participants who could otherwise not take part.^[9]

A colleague of mine said that [if you want to attract high-profile research leads], “you are only as strong as your weakest project” - which I thought was well put.

5. External communications

I shared an earlier version of this post with the organisers, and they were kind enough to provide a lot of valuable feedback. Despite being a previous participant, I got many things wrong (and likely still do). Arguably I could have done a better job of reading up on AISC, but I think it’s not an accident that they are one of the more misunderstood AIS organisations. Many people do not understand the reasoning behind some of the strategic decisions they make. Perhaps they could add an appendix to their funding proposal where they answer some common objections they would expect people to have, such as those in this post or the issues that came up here. This way people have a better picture of what they are trying and not trying to do. As a heuristics, if you think they have a similar theory of change to MATS, you should learn more about them.

A word of caution for funders

As you can read above, there are legitimate reasons to have reservations about AI Safety Camp. However, I want to caution any potential donors from automatically deferring to other funders. I have talked to dozens of people at conferences who think AISC is great and are not sure why they are not funded, and this somewhat updates them to think it has less EV.^[10]

I’m worried that we could end up in a cyclical reinforcement loop where people are not donating to AISC because others are also not donating, and no one really knows why.^[11] For a community that values rationality as ours, it would be a pity to run into such a failure mode. If you don’t trust Arb’s impact assessment, it might be particularly useful to investigate this opportunity yourself. If you are a (potential) funder, sharing your thoughts publicly about your choice would be especially valuable in this case.

Closing thoughts

To have a bit of fun at the end, wanted to share how I perceive AI Safety Camp as an outsider.

I imagine a very kind and slightly melancholic robot, who has been working really hard over the past years. The humans around don’t appreciate him enough, except for those who know him well. Some of his parts are rusty. He could use an oil change, but humans like the new and shiny robots better. I hope one day we realize how much this Iron Giant has done for the world, and show him the love he deserves.

Thanks to AISC organisers Linda and Remmelt for providing extensive feedback on the draft of this post. Also thank you to Peter and Milan for helping with editing.

1. ^{^}

   Thanks to Kola for sending me this link.

2. ^{^}

   Assuming the cost-effectiveness is similar for both.

3. ^{^}

   The organisations I founded, EA and AIS Hungary had successfully fundraised from different funders, even though there is a big overlap where there is only one ops team working for both organisations.

^{^}

OP recently stopped funding university clubs’ organiser salaries, but public perception was only one of the considerations they had. They still fund general group support costs, such as buying books and snacks.

^{^}

In case you are wondering, I opted in for the stipend because my salary for the next 10 months was still being evaluated by grantmakers at the time and I wanted the little bit of extra security. My fundraising was successful so I could have donated it back to AISC. I didn’t because I wouldn’t be able to get a tax-deduction and I don’t think an additional ~700-900 USD would have been worth a lot more in their pocket than mine.

^{^}

I’m borrowing and slightly expanding the scope of this term from this post. I’m slightly worried that this could cause some confusion, so in case anyone has a cooler term, I’m happy to hear it. :)

^{^}

Remmelt’s comment on this: “This round, we offered stipends of $1000 to 42 participants from low-income countries who opted in. A handful of those had actually emailed to say they were in a tight financial situation. My guess is that overall that is pretty well targeted actually to people who could use the money to focus. I don't know how many participants SPAR gets from low-income countries. You're right that AISC virtual has run fine without offering stipends as well. Side-note: research leads can also opt in to stipends of $1500 total each.”
My thoughts: To me, it's not clear that people from lower-income countries need the money more. While it's true that money goes further in non-western countries, I can imagine a Cambridge student struggling to buy groceries to get more out of 1k USD than a student in a poor country who otherwise gets money from their parents. I understand that it would take a lot of time to individually evaluate all of these small grants though, which is why they use an imperfect proxy. For context, this round they rejected 81 stipend requests, I assume of those who live in high-income countries.

^{^}

However, it’s worth mentioning that taking part in these projects is a great way to upskill.

^{^}

Let’s call them Campstone projects :)

^{^}

You can also read three comments where people are deferring to others’ judgement to various degrees.

^{^}

Linda makes an interesting point here: “There has been this persistent idea (that is now slowly breaking down) that if you can't get funded in AI Safety there has to be something wrong with you. When the reality has always been that the funders have certain tastes, and if you do things that don't follow their taste you will struggle. And also their taste are not at all representative for what most people think is good work.”

Show all footnotes