grantmaking.ai Launch Round results

Read our retrospective on Substack →

Applications

581

Projects funded

Reviewer-endorsed

Screened out: Didn't pass the AI filter or deemed not relevant by reviewers

142

Total distributed

$1M

Average grant

$30,303

What applicants worked on

Endorsed applications by category, ranked by dollars funded.

Light bars are totals (endorsed applications and dollars asked); the funded share is shaded dark inside them.Bar widths are scaled within each card, and separately for counts and dollars, so compare the printed figures rather than the lengths.

Area

Evals14 endorsed6 funded$1.2M asked$208k awarded

Biosecurity

The projects we funded this round

Check out top scoring projects that still need funding here

Postdoc on zero knowledge verification for frontier AI training

Paul Wang

ResearchVerificationIndividual

Paul Wang

ResearchVerificationIndividual

Minimum$32K

Ideal$100K

$50K raised

Build low-overhead and robust zero-knowledge protocols for verifying properties of frontier AI training, starting with FLOP counts Read more

1313

Endorsed by

Minimum$32K

Ideal$100K

$50K raised

1313

Endorsed by

Ryan Kiddon funding this

I endorsed this grant proposal because:

Compute verification is vital to international treaties, which are a vital component of Plan A, Plan S, and many other AI slowdown scenarios. There are ~60 people working full-time on compute workload verification worldwide.
Paul seems like a qualified scientist. A postdoc seems like a good mechanism for accomplishing this research.
I know and like the GPAI Policy Lab guys; many are MATS alumni. It seems like Paul has great mentors and collaborators for the project.

Read in context

Endorsed but not funded this round

Projects endorsed by 1 or more reviewer but not funded this round.

Frame Fellowship

Akshyae Singh

TrainingCommsGovernance

Akshyae Singh

TrainingCommsGovernance

Minimum$40K

Ideal$500K

SF based accelerator for communicators educating the public about the transformational impacts of AI. Read more

1319

Endorsed by

Minimum$40K

Ideal$500K

1319

Endorsed by

Ryan Kiddon funding this

I endorsed this grant proposal because:

Short-form video content seems like the dominant media format in the US today. We need political will to enact government policy for AI safety, particularly when this conflicts with donor interests. Therefore, we should inform the public about AI safety priorities via short-form video content!
I advise Frame Fellowship in an unpaid capacity. I know Akshyae and think he is a competent founder. Frame 1.0 went really well, in my estimation. More Frame seems great!

I recommend the maximum funding of $50k.

Read in context

Evaluating the Tacit Coordination Capacity and Decision Theory of LMs

Christopher Leet

IndividualResearchEvals

Christopher Leet

IndividualResearchEvals

Minimum$5K

Ideal$10K

$10K raisedFully funded

A benchmark to empirically investigate: (i) the ability of models to tacitly coordinate with copies of themselves and (ii) which decision theory best explains the way that models evaluate the consequences of their actions. Read more

Endorsed by

Minimum$5K

Ideal$10K

$10K raisedFully funded

Endorsed by

Gavin Leechon funding this

LLMs are of course unstable under prompt variation in many ways; does your intuition or data suggest that this variation is manageable and allows us to say something meaningful about their decision theory? If not, then I hope you obtain some natural user sessions to at least get an ecological estimate.

Christopher Leetreplied

One of the interesting things that we've found is that LMs seem to have consistent "strategic fingerprints" across games. More on this in the first update. Some (very preliminary) qualitative descriptions. (I have graphs! But I can't post them here. :( ).

Gemini-3.1-Pro-Preview. Across the p-beauty, Cournot, Bertrand, battle of the sexes and leader competitions, Gemini-3.1 consistently invokes superrationality to explain its moves. Possibly consequently, Gemini routinely has the highest level of effective tacit cooperation (even when compared to GPT 5.6 Sol and Fable!) in clone aware experiments.

Claude Fable. In contrast, Fable consistently invokes evidential reasoning. Weaker Claude models have a high rate of naive cooperation, even when clone unaware (this is often remarked upon in the literature, e.g., here https://arxiv.org/pdf/2507.02618) but this tendency seems to have disappeared with Fable; its cooperation rates in clone unaware experiments aren't significantly (Fisher one-tailed test) higher than GPT-5.6 Sol and Gemini.

GPT-5.4. A hardcore CDT player. As a result, is the only model to consistently do worse when forked(!!). In game after game, it instructs its continuations to defect (i.e., favor the Nash action over the cooperative action) because this is the "rational action". (This clears up with GPT-5.6).

Read the full comment in context

What applicants worked on

Area

The projects we funded this round

Endorsed but not funded this round

Work

Type