In our last post we talked about using karma as a signal to surface projects that were worth funding. Karma uses donations, upvotes, and comments as inputs—the obvious next step is to actually read the proposal. So we wanted to know: how well does AI do at rating Manifund projects?
I don’t think we’re at the point where we want to use AI to replace human judgment in grantmaking. Right now, the use I’m most optimistic about is surfacing proposals that it’s worth taking a closer look at.
Basic setup
We used Claude Opus 5.5 to read every proposal starting in July 2026 (to avoid proposals that might be in its training data—this was 530 proposals). We prompted it to write a few-sentence review of the proposal, give it a quality score from 1-5, and predict the probability it reached its minimum funding on Manifund.
Obviously, it’s much easier to check probability that the proposal gets funded than it is to judge a quality score. But this is both not what we intrinsically care about, and subject to various weird distortions (eg people posting proposals with a large minimum often have prearranged some funding before the proposal gets posted). For now, we did both.
Things we tried
We experimented with various aspects of the setup.
In the prompt, we got better results by emphasizing that the abilities of the grantee and feasibility and importance of the proposal mattered more than how well-written it was. (Previously, it seemed to overindex on the writing quality and whether the proposal included things like a detailed budget and milestone deliverables.) It was also helpful to tell it that small projects aren’t any more likely to get funded (prior to this, it used a lot of reasoning like “well the minimum is only $500, so all it would take is a single small donor”).
The probability predictions were improved by adding three terms: the minimum funding amount, the proposal length, and the goal divided by the minimum. (These didn’t get factored into the 1-5 rating.)
Stuff we tried that didn’t particularly help:
We tried using GPT-6 Sol instead of Claude Opus 5.5. Their rankings of the proposals were quite similar, but for some reason Sol’s calibration was way worse—it gave projects higher ratings across the board (too generously, in my opinion) and significantly overestimated the probability of projects getting funding.
We let the AI use web search to research the grantee. This barely seemed to improve the evals, and it introduces more contamination risk.
Telling Opus the base rate in the prompt made the calibration worse—without the base rate, its starting calibration was actually very good.
Results
We ended up with this regression model:
All results are on the subset of proposals that have already closed, and for which the AI evaluator didn’t learn from the text whether they had gotten funded. We got these results for the model’s probabilities:
AUC, standing for “area under the curve,” gives the chance a randomly chosen funded proposal is ranked above a randomly chosen unfunded one. A coinflip would get 0.5, while a perfect classifier would get 1. The Brier skill score measures how much better it was than always guessing the base rate, where 0 is no better and 1 is perfect. (These were measured with 5-fold cross-validation.) The average probabilities can be compared to a base rate of 15% of proposals in the dataset that ended up being funded.
Here’s the probability of getting funded, bucketed by Opus’s score:
And here’s the model’s calibration:
Caveat: bag-of-words did just as well at predicting
We also tried a very simple bag-of-words logistic regression … and it actually did just as well at predicting whether proposals got funded as Opus. (AUC 0.88, compared to AUC 0.90 for the model above. These are both calculated with 5-fold cross-validation.) You can see a table here of some of the terms that were weighted the most highly:
This does call into question how sophisticated Opus’s judgments really are. But I still think we should stick to using LLM evaluation over this for a few reasons:
Even if its predictive power was about the same, the judgment of the bag-of-words model struck me as qualitatively worse. The funded proposals where bag-of-words was higher than Opus by the largest amount read to me as not very strong—not that the project itself was bad, but often these were proposals that had clearly been discussed offline with a funder before posting, and I wouldn’t have wanted to fund it based on the proposal alone.
Bag-of-words is also obviously quite vulnerable to gaming. We don’t want to encourage people to get high scores by starting their proposals with “I don’t have a LessWrong account, have never been accepted by MATS, and don’t do interpretability research.” (This is true for LLM rating too, but to a lesser extent.)
LLMs can also give reasons that people can evaluate, and they’ll probably get better over time (when first running this with Sol, I accidentally used 5.6 instead of 6 and it was terrible).
Five potentially underrated projects
Out of the projects rated 4+ by Opus, these are the ones that either haven’t been funded or are still significantly below their funding goal. Consider checking them out—I thought they were all at the very least quite plausibly good funding targets!
Principled Agents is a research nonprofit working on theoretical alignment, with a track record of publishing research and endorsements on Manifund from Max Harms, Johannes Treutlein, and Cecilia Wood (which weren’t considered in the AI evaluation). Its funding from UK AISI runs out soon and it’s looking to raise up to 1 year of funding. Opus said:
This is a credible theoretical alignment org. Rubi Hudson has peer-reviewed outputs on directly relevant problems … and has already secured funding from UK AISI’s Alignment Project and the Corrigibility Research Fund. The team is small but competent, combining a math PhD with Anthropic Fellows experience and an engineer from Palisade. Weaknesses: the agenda of formalizing safety properties of utility functions is a long-shot bet with uncertain relevance to how frontier systems are actually trained, and the $500k ask is large relative to typical Manifund donations.
Ethan is an independent researcher who recently had a paper on restricting dangerous knowledge during pretraining recognized as a spotlight paper at ICML and cited by Dario Amodei as a promising approach. The proposal is for a research project on alignment drift during RL and continual learning. Opus said:
This is a strong researcher. A spotlight ICML paper on modular/gradient-routing pretraining done in collaboration with Anthropic, plus a public mention by Dario, is unusually credible for an independent applicant. The research question, alignment drift under RLVR and continual learning with gradient-routing countermeasures, is timely and builds directly on his prior work. Weaknesses: the scope is broad for a 4-month pilot, the evaluation design for drift is underspecified, and most of the budget goes to GPU rental, which some donors may see as better funded by labs or compute programs. Still, the track record and the natural extension of prior work make it an attractive bet.
Niklas produces German-language content on AI safety: he has 90,000 followers and produced a viral video explaining the HuggingFace incident with 2 million views. Right now he works full-time and produces content on the side—this funding would allow him to quit his job and focus on producing content for a year. Opus said:
This has unusually strong demonstrated traction for a solo creator: about 90k followers, roughly 2.5M monthly views and viral explainers. It also fills a real gap, since there is little German-language content on catastrophic AI risk, and the creator says German parliamentarians are among his viewers. His journalism and science-communication background makes delivery credible, and he has a sensible plan for sustainability through keynotes. The weaknesses are that the $70–75k minimum is large by Manifund standards, the impact of short-form reach on actual policy or careers is hard to verify, and the pending Coefficient Giving application could lead him to close this proposal or make donors wait for that decision.
IBBIS already has a free, open-source synthesis screening tool, Commec, that is in use by many DNA synthesis providers. This project funds the salary and travel budget for a partnerships lead to reach out to companies and encourage voluntary adoption of synthesis screening. Opus said:
IBBIS is a credible, well-connected biosecurity organization with a deployed open-source screening tool (commec), and the project targets a real, neglected gap: direct outreach to synthesis providers in China, India, South Korea and Brazil during a time-sensitive regulatory window. The theory of impact is clear, the risks are honestly stated, and the budget is mostly a single, specific hire with relevant adoption experience. Weaknesses are that outcomes depend on uncertain regulation and on provider incentives, and the budget ask has grown somewhat loosely from $100k to $190k–$265k. Overall it is a strong, low-cost lever on an important biosecurity chokepoint.
Another project by IBBIS, this covers continuing development on the engineering side of their synthesis screening (as opposed to outreach and adoption). Without funding before November, they’ll have to move Commec into maintenance mode. Opus said:
Commec is a real, shipped, open-source DNA synthesis screening tool. It has documented adoption (Dynegene, SIAT), peer-reviewed validation work with Microsoft, and a clear role at a tractable chokepoint for AI-enabled biorisk. That makes continued development a sound, high-value biosecurity bet from a credible institution backed by Founders Pledge. Weaknesses are the departure of project lead Tessa Alexanian, which shifts leadership to a less public figure, and a large minimum ($206k) that mostly covers one developer plus overhead. Commercial and SecureDNA alternatives also partly reduce the counterfactual value.
Two of the projects closed recently without being funded, but if you’re interested in funding them you should definitely reach out to the creators anyway!
Further work
We’re using AI scoring for new proposals as they come in, with the scores and reviews publicly viewable in our database. You can also see on our Github the code that scores proposals, including the prompt, and the research code. Our immediate plan is to start using these scores to surface good projects on our homepage and in our weekly digest.
There’s a lot more we’re interested in doing. On the research side, there’s:
refining and testing AI scores by having humans read Manifund proposals, rate them on quality, and compare those ratings to scores
using few-shot prompting
revisiting web search and other tools
using pairwise comparisons
having AIs debate each other on proposals
trying models that output calibrated probabilities directly, like Jev
Beyond using AI to find good proposals, we’re also interested in other constructive uses. One is helping grantees write proposals: maybe AI can ask questions to help them strengthen their proposal. There’s also suggesting funders, whether it’s people on Manifund or other institutions or grant programs that any given grant might be a good fit for.
And five projects on Manifund that need money!
In our last post we talked about using karma as a signal to surface projects that were worth funding. Karma uses donations, upvotes, and comments as inputs—the obvious next step is to actually read the proposal. So we wanted to know: how well does AI do at rating Manifund projects?
I don’t think we’re at the point where we want to use AI to replace human judgment in grantmaking. Right now, the use I’m most optimistic about is surfacing proposals that it’s worth taking a closer look at.
Basic setup
We used Claude Opus 5.5 to read every proposal starting in July 2026 (to avoid proposals that might be in its training data—this was 530 proposals). We prompted it to write a few-sentence review of the proposal, give it a quality score from 1-5, and predict the probability it reached its minimum funding on Manifund.
Obviously, it’s much easier to check probability that the proposal gets funded than it is to judge a quality score. But this is both not what we intrinsically care about, and subject to various weird distortions (eg people posting proposals with a large minimum often have prearranged some funding before the proposal gets posted). For now, we did both.
Things we tried
We experimented with various aspects of the setup.
In the prompt, we got better results by emphasizing that the abilities of the grantee and feasibility and importance of the proposal mattered more than how well-written it was. (Previously, it seemed to overindex on the writing quality and whether the proposal included things like a detailed budget and milestone deliverables.) It was also helpful to tell it that small projects aren’t any more likely to get funded (prior to this, it used a lot of reasoning like “well the minimum is only $500, so all it would take is a single small donor”).
The probability predictions were improved by adding three terms: the minimum funding amount, the proposal length, and the goal divided by the minimum. (These didn’t get factored into the 1-5 rating.)
Stuff we tried that didn’t particularly help:
Results
We ended up with this regression model:
All results are on the subset of proposals that have already closed, and for which the AI evaluator didn’t learn from the text whether they had gotten funded. We got these results for the model’s probabilities:
AUC, standing for “area under the curve,” gives the chance a randomly chosen funded proposal is ranked above a randomly chosen unfunded one. A coinflip would get 0.5, while a perfect classifier would get 1. The Brier skill score measures how much better it was than always guessing the base rate, where 0 is no better and 1 is perfect. (These were measured with 5-fold cross-validation.) The average probabilities can be compared to a base rate of 15% of proposals in the dataset that ended up being funded.
Here’s the probability of getting funded, bucketed by Opus’s score:
And here’s the model’s calibration:
Caveat: bag-of-words did just as well at predicting
We also tried a very simple bag-of-words logistic regression … and it actually did just as well at predicting whether proposals got funded as Opus. (AUC 0.88, compared to AUC 0.90 for the model above. These are both calculated with 5-fold cross-validation.) You can see a table here of some of the terms that were weighted the most highly:
This does call into question how sophisticated Opus’s judgments really are. But I still think we should stick to using LLM evaluation over this for a few reasons:
Even if its predictive power was about the same, the judgment of the bag-of-words model struck me as qualitatively worse. The funded proposals where bag-of-words was higher than Opus by the largest amount read to me as not very strong—not that the project itself was bad, but often these were proposals that had clearly been discussed offline with a funder before posting, and I wouldn’t have wanted to fund it based on the proposal alone.
Bag-of-words is also obviously quite vulnerable to gaming. We don’t want to encourage people to get high scores by starting their proposals with “I don’t have a LessWrong account, have never been accepted by MATS, and don’t do interpretability research.” (This is true for LLM rating too, but to a lesser extent.)
LLMs can also give reasons that people can evaluate, and they’ll probably get better over time (when first running this with Sol, I accidentally used 5.6 instead of 6 and it was terrible).
Five potentially underrated projects
Out of the projects rated 4+ by Opus, these are the ones that either haven’t been funded or are still significantly below their funding goal. Consider checking them out—I thought they were all at the very least quite plausibly good funding targets!
Principled Agents (Rubi Hudson, $50k raised out of a $500k goal)
Principled Agents is a research nonprofit working on theoretical alignment, with a track record of publishing research and endorsements on Manifund from Max Harms, Johannes Treutlein, and Cecilia Wood (which weren’t considered in the AI evaluation). Its funding from UK AISI runs out soon and it’s looking to raise up to 1 year of funding. Opus said:
Robustifying alignment techniques to survive RLVR and continual learning (Ethan Roland, closed on October 9 without reaching its minimum funding)
Ethan is an independent researcher who recently had a paper on restricting dangerous knowledge during pretraining recognized as a spotlight paper at ICML and cited by Dario Amodei as a promising approach. The proposal is for a research project on alignment drift during RL and continual learning. Opus said:
Quit my job to focus on AI safety content creation (Niklas Titgemayer, $0 raised out of an $85k goal)
Niklas produces German-language content on AI safety: he has 90,000 followers and produced a viral video explaining the HuggingFace incident with 2 million views. Right now he works full-time and produces content on the side—this funding would allow him to quit his job and focus on producing content for a year. Opus said:
Scaling adoption of synthesis screening outside the US and EU (IBBIS, $68k raised out of a $190k goal)
IBBIS already has a free, open-source synthesis screening tool, Commec, that is in use by many DNA synthesis providers. This project funds the salary and travel budget for a partnerships lead to reach out to companies and encourage voluntary adoption of synthesis screening. Opus said:
Continued development of Commec (IBBIS, closed on September 18 without reaching its minimum funding)
Another project by IBBIS, this covers continuing development on the engineering side of their synthesis screening (as opposed to outreach and adoption). Without funding before November, they’ll have to move Commec into maintenance mode. Opus said:
Two of the projects closed recently without being funded, but if you’re interested in funding them you should definitely reach out to the creators anyway!
Further work
We’re using AI scoring for new proposals as they come in, with the scores and reviews publicly viewable in our database. You can also see on our Github the code that scores proposals, including the prompt, and the research code. Our immediate plan is to start using these scores to surface good projects on our homepage and in our weekly digest.
There’s a lot more we’re interested in doing. On the research side, there’s:
Beyond using AI to find good proposals, we’re also interested in other constructive uses. One is helping grantees write proposals: maybe AI can ask questions to help them strengthen their proposal. There’s also suggesting funders, whether it’s people on Manifund or other institutions or grant programs that any given grant might be a good fit for.