The AI safety ecosystem is still in need of generalists: people who will go out and solve the tasks that people who feel limited by their job description won’t address. This post will explain how to break into generalist work, then provide an important set of projects that, if completed, could strengthen many crucial parts of the AI safety ecosystem, as well as your own career.
If you’re looking for something concrete to do, there are a bunch of good entry points:
The rest of the post has more detail on choosing projects, getting mentorship, and other ways to get involved.
University Student: get heavily involved with your university AI safety group. University groups are real, impactful organizations, and they can be run as such.
If your school doesn’t have a group, you should start one. Convincing an equally talented person to join AI safety can straightforwardly double your impact, and many of the most impactful people working in the field today credit their university groups for convincing them to make this career choice.
A university group is also a great place to cut your teeth doing generalist work. Most group organizers find themselves working on operations, strategy, events, headhunting, and research in a single week, testing their fit for each type of work and operating under enviably tight feedback loops. It’s rare that an early-college generalist finds a better opportunity for both impact and career capital than “making their school’s group a lot better,” which is why many of the field’s best generalists were once university group organizers.
If you would like to get access to funding or mentorship, the Pathfinder Fellowship and the AI Safety Seeding Initiative are both fantastic resources to get mentorship and funding. University groups can run existing educational programs like ARENA, or they can support students’ efforts to get into SPAR, MATS, Generator, or one of the many other advanced fellowships that teach people how to do impactful work.
Early or Mid-Career Professional: Make sure you have relevant context on the AI safety space. Take some BlueDot courses, meet people at AI safety events, and read voraciously from a variety of sources.
Once you have a general understanding of the field, look into projects. Completing an AI safety project is a good way to support the field, meet new people, and gain valuable skills and context. You can just do many of these things; you don’t need permission or a full-time role to contribute!
This section is basically a TL;DR version of 80k Hours’ posts on how to get into AI safety in 3 months. There’s a wealth of information on their Substack, their website, and they just released a book!
The AI safety space is littered with unresolved problems and inefficiencies, because many of the people who were skilled enough to notice these problems don’t have enough spare time to address them. While preparing for the Generator Residency, we conducted interviews with dozens of AI safety professionals from organizations like Coefficient Giving, AI Futures Project, Forethought, and more.
The residents couldn’t do all of them, but that doesn't mean they shouldn’t be done.
Some of these projects might make the talent pipeline more efficient, making it easier for passionate people to learn about AI safety and gain the skills that make them impactful. Other projects might improve our data collection systems, make concerns about existential risks visible to a broader audience, make it easier to convert political will into impactful legislation, or fill some broader need that can’t be easily described.
This list is far from exhaustive, and many of the programs listed will be continuously maintained with new projects as the needs of the field evolve. We hope that this section serves as a jumping-off point for you to explore the programs and shared resources outlined below.
SPAR’s generalist projects are a new category under SPAR with mentors from Generator, Constellation, Kairos, universities, and other AI safety organizations on projects in fieldbuilding, communications, research, and organizational infrastructure.
Current projects include:
Plus a dozen more, as well as a Generalist Megastream for strong applicants who do not see a perfect fit among the existing listings. Applications close August 18.
We’ve added a number of projects we find promising to Mangrove, a platform for coordinating AI safety projects and communities.
We think a lot about AI safety talent pipelines, and by far the biggest drop-off for potential generalists is the stage between “having context in AI safety” and “building signal to get into fellowships”. Much time gets lost due to lack of coordination, and much motivation gets lost due to lack of accountability and community.
We’re excited about Mangrove as a place for people to hit the ground running on interesting projects, while developing credibility and connections in the space. Mangrove is in development, and plans to host hackathons, community projects for university groups, a contracting platform for orgs (!), and more.
Mangrove’s official launch announcement is coming in a few days, but you can check out the platform right now.
Some projects we seeded on Mangrove:
A lot of generalist shaped things don’t have discrete projects, and rather require building specific skills like effective comms, hiring/recruiting, org strategy, management, and much more, to pick up dropped balls. As a sample of these sorts of fuzzier projects:
Consider being an influencer: As “warning shots” become scarier and more frequent, there’s potential and demand for high-quality media that connect these events to broader AI safety concerns and convert public will into action. Some great media has come out of the AI safety ecosystem as the Overton window is shifting, and programs like The Frame and Tarbell have emerged to target this.
Political organizing: There’s a lot of useful work in recruiting people, running events, building local groups, organizing constituent outreach, and helping turn public will into useful legislation and strong leadership. This work is often bottlenecked on reliable organizers, and many consider it incredibly high-leverage. Political organizing can become quite sensitive and prone to backfire risk, so if you’re interested in this, we recommend filling out this short form before launching something yourself and we'll connect you with someone who can advise you.
Writing: Writing publicly is one of the fastest ways to build a track record. The spreadsheet below includes a running list of posts people want written.
As a general rule, you should be applying for more things. New programs are springing up to accelerate generalists:
Many people are too scared to apply for funding, when probably they shouldn't be! Even if you don’t get the money, oftentimes submitting a (non-slop) grant proposal can be a good forcing function for execution.
It can be extremely helpful to know people in the community who can give you mentorship. A good mentor can make a big difference in acquiring taste, skills, and connections. Many of the best projects require specific context to complete well.
As a general rule, your bar for reaching out to people is too high.
For example, if you’re especially interested in a series of blog posts someone has written, consider reaching out to discuss the content of their post further (but try not to ask questions that you can easily answer by reading more of their work, or a 2 minute search).
Alternatively, if there are sets of projects you are interested in pursuing, consider reaching out to people who have done similar work, or find a SPAR mentor who can help you do so. You can reach the authors of this post at [email protected], [email protected] (Harry), or [email protected], and you can meet many more people at events like EA Global or a Global Challenges Project workshop.
AI safety can be a confusing field to enter, and a lot of useful context lives inside organizations, private conversations, and people’s heads. Still, we think one of the fastest ways to build that context is to work on something concrete. Projects expose you to fast feedback from the outside world, give you reasons to talk to knowledgeable people, and get you oriented faster than reading alone.
