Manifund is raising money for our 2026 AI safety regranting program!
TL;DR
Early grants can have huge returns.
Dispersed part-time grantmakers can find opportunities before centralized funders.
Donating to Manifund Regranting puts your money in these opportunities.
See also our 2025 announcement for more thoughts on “why regranting?” (answer: it’s fast, low-friction, proactive), and “what makes a good regrantor?” (answer: taste, capacity, dealflow).
Six top hits
We’ve run AI safety regranting every year since 2023, so we’ve now had enough time to see how some of our early grants have played out. And because giving is hits-based, looking at a few successes can prove the value of our program.
$143k in August 2023 from Evan Hubinger, Ryan Kidd, Rachel Weinberg, and Marcus Abramovitch
This grant provided the initial funding for Timaeus to get started. Since then, they grew their organization and eventually merged with Resolution, earning a $160m grant — Coefficient Giving’s largest-ever AI safety grant. See this post for our argument that this grant alone makes all of our regranting so far worth it:
Multiplied out, our $143k stake via our 2023 regrants would have grown 53x to now be worth $7.6m. This first regrant to Timaeus, alone, would more than pay back the entire Manifund regranting budget to date ($5.2m). Since the Resolution grant is so much larger than most other AIS grants, Timaeus could be the single best 2023 AI safety grant in these terms.
This situation struck me as being preposterous. Here is one of the world’s top experts on AI consciousness, claiming a nerve-wracking chance of AI consciousness in the not-too-distant future, with fairly strong professional links at top AI labs and ~shovel-ready ideas for empirical projects, preparing a not-terribly-costly ask (give me a work desk, ~0 engineering time, and model access to work on these research questions)... and he is unsure whether he should make the ask?!
Joel solved this by paying for 5 weeks of Rob’s time, allowing him to research AI consciousness and moral patienthood, rather than applying for jobs.
Rob went on to found Eleos AI a year later, which has raised more than $2m from funders like Macroscopic and Longview, and performed the first officially-commissioned welfare evaluations of frontier models on Claude Opus 4. Rob’s final report: “This grant was absolutely clutch.”
$2500 for Neuronpedia and $25k for Joseph Bloom in July 2023 from Austin Chen and Marcus Abramovitch
For Neuronpedia, this was the first funding they’d received; for Joseph Bloom, he’d been working as an independent researcher for 6 months at the time, and the grant enabled 6 months of his research.
These were separate grants, but Joseph and Neuronpedia ended up collaborating to produce tools that are widely used in mechanistic interpretability. Joseph used the funding to release the SAE Lens library and partner with DeepMind on Gemma Scope, among other projects. He’s now the model transparency lead at UK AISI.
Neuronpedia pivoted from its initial game concept to produce open-source interpretability infrastructure used by DeepMind and Anthropic. It went on to get donations from LTFF, Superalignment Fast Grants, AISTOF, and SFF.
This grant was to a research project by a PhD student and an undergrad and was used to pay consultants to develop a 3600-question dataset of questions surrounding hazardous knowledge in CBRN and cybersecurity. This dataset became the WMDP benchmark, the standard benchmark for hazardous knowledge in the unlearning subfield.
$38k in September 2023-August 2024 from Joel Becker and Evan Hubinger
At the time of the regrant, ChinaTalk was a part-time operation running on a budget of $35k/year, and Manifund was the first outside funding it received. The grant allowed the founder, Jordan Schneider, to spend less time on paid consulting to support himself.
ChinaTalk now has five full-time staff and has gotten funding from Emergent Ventures, Macroscopic, and Coefficient Giving. They’re listened to by many influential sources and did notable early reporting on DeepSeek before it made major news in the West.
$335k in August through December 2023 from Tristan Hume, Evan Hubinger, and Marcus Abramovitch
Apollo Research launched in May 2023 as an eight-person organization. Since then they’ve built the first empirical evaluations for AI scheming, worked with OpenAI and Anthropic to test models before shipping, and have grown to around forty people.
This one is less exciting than some others because it came after $2m of funding from Open Phil and SFF. But Manifund regrantors were still quite early in the lifetime of the organization, and provided a significant fraction of their early funding. (We’d estimate that they raised a ~$20m seed round for their PBC in 2026, which would imply a ~10x impact ROI from this regrant.)
See for yourself
If you’d like to review all the regrants we’ve ever made, check out this dashboard. I also asked Claude Cowork to look through our regrants and rank the best ones in retrospect in this spreadsheet. I think it did a pretty good job—we decided to include its 1-5 and 8 in our “top hits” section.
The case for part-time grantmakers
One argument we’ve heard is, “There are organizations full of specialized full-time grantmakers. Why give your money to part-time grantmakers instead?”
We think it’s good for a few reasons:
Many people’s first $100k of grants will beat a professional grantmaker’s last $100k.
Any given grantmaker’s field of view can only be so big: people are limited by their networks, geographic locations, and areas of expertise. Lots of people have specialized knowledge or personal contacts that allow them to notice a few good opportunities.
It’s way more scalable.
The number of researchers and projects and the amount of funding in AI safety is growing massively in 2026, but it’s going to be difficult for our capacity to allocate the funding to the right researchers and projects to keep up. Scaling up a foundation full of professional grantmakers is hard for the same reasons scaling orgs generally is hard. It’s easy to get bottlenecked on hiring, management, training, or trust.
On the other hand, finding talented people to give regranting budgets to and increasing their budget if they make good grants is a much lower-friction process.
It helps find, empower, and upskill regrantors.
In the past, we’ve given regranting pots to a mix of those well-known in the community and people who were less connected but who we thought were promising. Some of those people have ended up being (in our opinion) great regrantors and gone on to jobs in AI safety—like Joel Becker, who started as a regrantor in 2023, and in 2025 became a member of technical staff at METR and coauthor on some of their most-discussed work.
Should you donate?
We think regranting is a good fit for donors who are earning-to-give $50k+/year, who care about AI safety but may not be as steeped in the AI safety ecosystem. Think of regrantors as part-time program officers who can provide taste and dealflow without the logistics of hiring or creating a foundation.
Everyone expects a lot of Anthropic and OpenAI money to flow into AI safety soon, but there’s a lack of good opportunities. Regranting is particularly well suited to seeding new opportunities; it puts money in the hands of individuals with taste and/or dealflow.
And if you are steeped in the ecosystem and want to have more fine-grained control over where your money goes, regranting could also be a good fit! Historically, many of our best regrantors have been lab employees, who’ve had taste, but not liquidity — until now. This year, we’d like to ask them to fund part of their own regranting budget, while Manifund handles fiscal sponsorship and payment logistics. (We piloted this with Leo Gao’s microgrants program.)
Logistics
In 2025, our pool of 13 regrantors gave away $2m. At a minimum, with $2m again, we could run a similar program, with 10-15 regrantors each receiving a starting pot of $100k, and top-ups for those who are making exciting grants. We’d like to target $5m for a more ambitious program: say, 15 proven regrantors getting pots of $200k to start and 20 more getting a starting pot of $50k. And we expect to be able to easily scale regranting beyond that.
We’re flexible on the structure, whether you’d rather make a donation to our regranting program as a whole, donate to specific regrantors, or suggest your own regrantors. We do ask that donors contribute 5% to Manifund’s operating budget.
Mostly written by Carol, edits by Austin.
Manifund is raising money for our 2026 AI safety regranting program!
TL;DR
See also our 2025 announcement for more thoughts on “why regranting?” (answer: it’s fast, low-friction, proactive), and “what makes a good regrantor?” (answer: taste, capacity, dealflow).
Six top hits
We’ve run AI safety regranting every year since 2023, so we’ve now had enough time to see how some of our early grants have played out. And because giving is hits-based, looking at a few successes can prove the value of our program.
1. Timaeus
$143k in August 2023 from Evan Hubinger, Ryan Kidd, Rachel Weinberg, and Marcus Abramovitch
This grant provided the initial funding for Timaeus to get started. Since then, they grew their organization and eventually merged with Resolution, earning a $160m grant — Coefficient Giving’s largest-ever AI safety grant. See this post for our argument that this grant alone makes all of our regranting so far worth it:
2. Rob Long’s research on AI consciousness
$7200 in October 2023 from Joel Becker
Joel explains:
Joel solved this by paying for 5 weeks of Rob’s time, allowing him to research AI consciousness and moral patienthood, rather than applying for jobs.
Rob went on to found Eleos AI a year later, which has raised more than $2m from funders like Macroscopic and Longview, and performed the first officially-commissioned welfare evaluations of frontier models on Claude Opus 4. Rob’s final report: “This grant was absolutely clutch.”
3. Neuronpedia and Joseph Bloom
$2500 for Neuronpedia and $25k for Joseph Bloom in July 2023 from Austin Chen and Marcus Abramovitch
For Neuronpedia, this was the first funding they’d received; for Joseph Bloom, he’d been working as an independent researcher for 6 months at the time, and the grant enabled 6 months of his research.
These were separate grants, but Joseph and Neuronpedia ended up collaborating to produce tools that are widely used in mechanistic interpretability. Joseph used the funding to release the SAE Lens library and partner with DeepMind on Gemma Scope, among other projects. He’s now the model transparency lead at UK AISI.
Neuronpedia pivoted from its initial game concept to produce open-source interpretability infrastructure used by DeepMind and Anthropic. It went on to get donations from LTFF, Superalignment Fast Grants, AISTOF, and SFF.
4. Weapons of Mass Destruction Proxy (WMDP) benchmark
$190k in January 2024 from Dan Hendrycks
This grant was to a research project by a PhD student and an undergrad and was used to pay consultants to develop a 3600-question dataset of questions surrounding hazardous knowledge in CBRN and cybersecurity. This dataset became the WMDP benchmark, the standard benchmark for hazardous knowledge in the unlearning subfield.
5. ChinaTalk
$38k in September 2023-August 2024 from Joel Becker and Evan Hubinger
At the time of the regrant, ChinaTalk was a part-time operation running on a budget of $35k/year, and Manifund was the first outside funding it received. The grant allowed the founder, Jordan Schneider, to spend less time on paid consulting to support himself.
ChinaTalk now has five full-time staff and has gotten funding from Emergent Ventures, Macroscopic, and Coefficient Giving. They’re listened to by many influential sources and did notable early reporting on DeepSeek before it made major news in the West.
6. Apollo Research
$335k in August through December 2023 from Tristan Hume, Evan Hubinger, and Marcus Abramovitch
Apollo Research launched in May 2023 as an eight-person organization. Since then they’ve built the first empirical evaluations for AI scheming, worked with OpenAI and Anthropic to test models before shipping, and have grown to around forty people.
This one is less exciting than some others because it came after $2m of funding from Open Phil and SFF. But Manifund regrantors were still quite early in the lifetime of the organization, and provided a significant fraction of their early funding. (We’d estimate that they raised a ~$20m seed round for their PBC in 2026, which would imply a ~10x impact ROI from this regrant.)
See for yourself
If you’d like to review all the regrants we’ve ever made, check out this dashboard. I also asked Claude Cowork to look through our regrants and rank the best ones in retrospect in this spreadsheet. I think it did a pretty good job—we decided to include its 1-5 and 8 in our “top hits” section.
The case for part-time grantmakers
One argument we’ve heard is, “There are organizations full of specialized full-time grantmakers. Why give your money to part-time grantmakers instead?”
We think it’s good for a few reasons:
Many people’s first $100k of grants will beat a professional grantmaker’s last $100k.
Any given grantmaker’s field of view can only be so big: people are limited by their networks, geographic locations, and areas of expertise. Lots of people have specialized knowledge or personal contacts that allow them to notice a few good opportunities.
It’s way more scalable.
The number of researchers and projects and the amount of funding in AI safety is growing massively in 2026, but it’s going to be difficult for our capacity to allocate the funding to the right researchers and projects to keep up. Scaling up a foundation full of professional grantmakers is hard for the same reasons scaling orgs generally is hard. It’s easy to get bottlenecked on hiring, management, training, or trust.
On the other hand, finding talented people to give regranting budgets to and increasing their budget if they make good grants is a much lower-friction process.
It helps find, empower, and upskill regrantors.
In the past, we’ve given regranting pots to a mix of those well-known in the community and people who were less connected but who we thought were promising. Some of those people have ended up being (in our opinion) great regrantors and gone on to jobs in AI safety—like Joel Becker, who started as a regrantor in 2023, and in 2025 became a member of technical staff at METR and coauthor on some of their most-discussed work.
Should you donate?
We think regranting is a good fit for donors who are earning-to-give $50k+/year, who care about AI safety but may not be as steeped in the AI safety ecosystem. Think of regrantors as part-time program officers who can provide taste and dealflow without the logistics of hiring or creating a foundation.
Everyone expects a lot of Anthropic and OpenAI money to flow into AI safety soon, but there’s a lack of good opportunities. Regranting is particularly well suited to seeding new opportunities; it puts money in the hands of individuals with taste and/or dealflow.
And if you are steeped in the ecosystem and want to have more fine-grained control over where your money goes, regranting could also be a good fit! Historically, many of our best regrantors have been lab employees, who’ve had taste, but not liquidity — until now. This year, we’d like to ask them to fund part of their own regranting budget, while Manifund handles fiscal sponsorship and payment logistics. (We piloted this with Leo Gao’s microgrants program.)
Logistics
In 2025, our pool of 13 regrantors gave away $2m. At a minimum, with $2m again, we could run a similar program, with 10-15 regrantors each receiving a starting pot of $100k, and top-ups for those who are making exciting grants. We’d like to target $5m for a more ambitious program: say, 15 proven regrantors getting pots of $200k to start and 20 more getting a starting pot of $50k. And we expect to be able to easily scale regranting beyond that.
We’re flexible on the structure, whether you’d rather make a donation to our regranting program as a whole, donate to specific regrantors, or suggest your own regrantors. We do ask that donors contribute 5% to Manifund’s operating budget.
If you’re interested in donating, reach out to austin@manifund.org!