And how do we fix it?

I’m currently working on a BlueDot AI Safety Grantmaking Fundamentals course, and in the process of recruiting grantmakers for curriculum development, I’ve been having chats with several people thinking about field strategy. The following is inspired by a conversation with Evan Miyazono from Atlas Computing (more on Atlas below).

In this post, I’ll be attempting to address four questions:

1. What’s wrong with passive grantmaking?
2. What’s wrong with active grantmaking?
3. How do we solve these problems?
4. How could we scale this approach?

1. What’s wrong with passive grantmaking?

Passive grantmaking = evaluating proposals.

A grantmaker has to evaluate incoming proposals along two axes:

1. Project value: Do I think this project is important? Will it contribute to the goals my organization has set (e.g. reducing existential risk from AI)?
   1. What are the downside risks?
   2. If this project fails, what is the likeliest reason for the failure?
   3. In what worlds is this project actively harmful?
2. Applicant fit: Is the person who submitted the proposal the right person to do this project?
   1. How competent are they? Do I think they can execute this proposal well?
   2. What do their references look like? How strongly do people they’ve worked with endorse them?

Some notes on the above:

• Finding people who can think critically about what’s needed in the field (henceforth referred to as “field strategists”) is difficult.
  • Usually, the people who are best positioned to do this are high-context in AI safety, and have been thinking about what’s missing in the field for a while.
  • People with this kind of strategic taste are also likely to already be working in high-level, impactful roles, which means they’re rarely willing to transition away from those roles.
• Being a good field strategist does not necessarily require execution ability.

Passive grantmaking forces every applicant to have both skills. It doesn’t need to.

2. What’s wrong with active grantmaking?

Active grantmaking is when a grantmaker:

1. Comes up with a good idea for a project
2. Searches for people in their network to seed the project with

Problems with the above:

AI safety grantmakers are very very busy

• Grantmaker time is already extremely limited with all of the proposals they receive; grant evaluations at Coefficient Giving, for example, regularly take months to be completed. It’s difficult for an already capacity-constrained grantmaker to set aside time for substantial high-level strategy work.
• Successfully searching for people in your network requires:
  • Thinking about who would be a good fit;
  • Attempting to pitch people to potentially leave their jobs to do this other thing;
  • Having sufficiently high confidence in the people you pitch.
    • An unfortunate failure mode of active grantmaking: pitching someone on a project, getting them to put in the effort to apply for funding, realizing they aren’t a good fit while evaluating their application, and having to say no.

The above assumes that you happen to know people who would be a good fit for your project at all.

3. How do we solve these problems?

The key to optimizing grantmaking is specialization: separating project proposals from people– separating evaluations of project value from applicant fit.

Two case studies:

1. The Institute For Progress (IFP)

IFP’s Launch Sequence is an open call for project proposals “that prepare the world for advanced AI.” They write:

“You should submit a proposal if you have a good idea, even if you don’t expect to be the person actually leading or working on the resulting project… If your proposal is selected, we’ll work closely with you to develop the idea into a concrete project plan, publish your plan, connect you with philanthropic funders, and help you headhunt for a project lead or co-lead.”

If you’re interested in engaging in field strategy work part-time, consider authoring a proposal or offering ideas to IFP. They pay a $10,000 honorarium for published proposals and $1,000 bounties for ideas they publish.

In effect, IFP treats field strategy as freelance work: anyone can contribute an idea, get paid, and walk away. Atlas Computing, Evan Miyazono’s organization, is betting on the opposite model: field strategy as a full-time job.

2. Atlas Computing

Evan in this post discusses his plans to make Atlas a hub that “rais[es] funding to gather and support [field strategists]” to address “[t]he problems that don’t fit neatly into any existing fiefdom.” He writes:

“[These orphaned problems] need someone whose job it is to find them, scope them, and make it easy for the right person to step in and solve them.”

If you’re interested in being a Field Strategist in a more official capacity than what IFP’s Launch Sequence offers, Atlas Computing’s job listing is here.

If you’re interested in supporting Atlas’s work, you can also email [email protected].

4. Scaling this approach

How might we scale this approach to major funders? Here are a couple of ideas:

1. Hire full-time field strategists. Based on conversations I’ve had with grantmakers working at major funding organizations, there doesn’t seem to be an explicit, written high-level strategy to win that proposals are evaluated against. Organizations like Coefficient Giving and Longview should hire strategists to consider:
   1. What is our theory of victory for AI safety?
   2. Given this theory of victory, what is on the critical path?
   3. What projects and organizations on the critical path do we need to initialize?
2. Separate people from proposals. Publish detailed, pre-scoped project plans that applicants apply to execute.
   1. Instead of evaluating both the project proposal’s merits and the person’s fit, once this is done, the only question to answer becomes: can this person execute this project? That’s a much easier judgment than assessing a novel idea and an unknown person simultaneously.

Summary

In sum, the grantmaking pipeline currently requires three distinct jobs:

1. Field strategy: diagnosing what the field needs.
2. Project design: scoping those needs into concrete, fundable plans.
3. Execution matching: pairing each plan with someone who can run it.

Today, all three are combined into a single application. Applicants must diagnose the field, design a project, and nominate themselves to execute it. Grantmakers must then evaluate all three at once.

It doesn’t need to be this way. Field strategists and scoping processes like IFP’s can own jobs one and two; application evaluation can then shrink to job three alone: can this person run this plan?

Conclusion

Photo by Brecht Corbeel on Unsplash

AI safety is facing an unprecedented wave of funding with the upcoming Anthropic and OpenAI IPOs. As I’ve written previously:

When [Anthropic IPOs], a few thousand people, including some of the wealthiest people on the planet, will become liquid, and a meaningful fraction of those people will want to give to AI safety.

No philanthropic field this young, with this little absorptive capacity, has ever successfully channeled a funding influx this large. The closest precedents ended badly.

We need a radically different approach to deploying capital if we want to be the first to pull it off.