We’re announcing Project Tailwind: a call for founders to launch ambitious new AI safety initiatives, in the spirit of Y Combinator's Requests for Startups.
The initiatives we want to see
The world remains unprepared for transformative AI. There are critical, basic problems that no one owns. Any industry whose failures can cause serious harm needs a safety ecosystem. Aviation has certification authorities and incident investigators; medicine has trial regulators and disease surveillance; finance has independent auditors and deposit insurance. Until recently, AI had nothing like that. People have started building these institutions (and we're very excited about the number of strong organizations relative to 2–3 years ago), but many of these gaps persist because no organization is dedicated to closing them.
We've published a list of initiatives we want to exist, spanning evidence generation, independent assessment, AI security and verification, research institutions, talent acceleration, and more. Each is endorsed by a grantmaker on our team who has thought about the problem and is ready to fund a solution.
This set of initiatives isn't comprehensive. We’d be excited to fund many initiatives that aren’t listed, and an initiative’s inclusion doesn’t mean no one is working on that problem — only that we want to fund more work on it, including from existing teams.
We also think the best founders will bring their own conviction about what needs to be done. If you have an idea that could do more to reduce catastrophic risks from AI than anything on our list, we want you to pitch it to us.
Some of these initiatives may also be better housed at existing organizations — we'd like people to seriously explore that option and pitch us on it instead if so.
Founders are the bottleneck, not funding
For most of the AI safety field's history, money has been the binding constraint. We don't think that's true anymore. Last year, our technical AI safety, security, and fieldbuilding programs directed roughly $350 million. This year, we're on track to direct more than $1 billion, and we expect substantially more funding could soon materialize.
What we're short on is organizations that can effectively turn that funding into impactful work at scale, and founders to build them. Teams that received $1–2 million from us four years ago can now productively deploy 20x that amount every year. Some of the most important institutions in the field today, including Epoch, Redwood Research, and MATS, started within the last five years as just a few people with an idea. Today, they advise governments, guide frontier developers, and lead their own subfields. We'd like many more organizations on that trajectory, moving even faster. Institutions take time to build — so if the world will need something in two years, someone should be founding it today.
For the right person, we think founding an AI safety organization is plausibly the highest-impact thing they could do.
What funding looks like
We're funding at three levels:
- Pre-seed ($200k–$2m): Develop an idea, assemble a founding team, and validate your approach over 9–18 months. You don't need a legal entity during this period.
- Seed ($2m–$20m): Launch and scale over 18–36 months, with competitive salaries and serious compute budgets.
- Scale ($20m–$200m+): Target the field's most important problems with a world-class team, backed at the scale the problem demands. (Our recent grant to Resolution is an example.)
If you’re funded at one level, you don't have to wait until your funding window ends to apply for the next. For example, if your pre-seed work is going well after three months, you should feel empowered to ask for seed funding then.
To move at the pace this requires, we're operating less like a traditional foundation and more like an impact-seeking venture firm. This ties into changes we’ve made to improve our grantmaking process and get money out the door a lot faster than we did historically.
Who we're looking for
Founding an AI safety nonprofit presents challenges beyond those of a typical startup: there's no revenue signal to tell you what's working, and experts disagree on fundamental assumptions. The win condition is avoiding an outcome (catastrophe), and progress toward that is harder to measure than progress toward creating a specific outcome. We're looking for people who can articulate a strong theory of change, ruthlessly prioritize impact, and build excellent teams as they scale.
Check it out
Project Tailwind is live. Start here, then check out the list of initiatives. If you're ready to build something on our list (or you have a better idea), express your interest here.
