Hmm well isn’t this basically the core project of EA aha? anyway thanks for sharing I chatted my two cents into chaptgpt 5.5 and had it format/write it out for me.
The Foundation should target market failures: public goods, externalities, information failures, coordination failures, etc. That means being suspicious of “important but already incentivized” domains. I start with this because of an example they have already expressed interest in—Alzheimer’s is important, but it is not obviously where a new mega-foundation has the highest marginal leverage. There are already enormous commercial incentives to develop effective treatments: rich patients, aging societies, and pharmaceutical companies all badly want a breakthrough. The philanthropic opportunity is not “Alzheimer’s” in general, but specific neglected bottlenecks inside Alzheimer’s: open datasets, repurposed generics, trial infrastructure, non-patentable interventions, prizeable biomarkers, fixing and standardizing medical ontologies, or other areas where private returns diverge from social returns.
The bigger bottleneck is that we do not have a trustworthy system for converting hundreds of billions into impact. Conventional grantmaking is too opaque, too relationship-driven, too vulnerable to value lock-in, and too dependent on a small number of people’s cached worldviews. If you give that system $180 billion and scale quickly, you mostly get larger versions of the same failure modes.
So my first rule would be: do not spend the endowment quickly (yes I also have short timelines so need to balance that but won’t get into that rn). For the first two or three years, spend a small fraction building the allocation machine: the mechanisms, audits, forecasting systems, and public models that make later spending less dumb. The Foundation should have three arms.
First, a pull-funding arm. This should be the biggest. Wherever outcomes can be specified reasonably well, the Foundation should stop trying to guess the best grantees ex ante and instead pay for results. This is the logic behind results-based financing, advance market commitments, prizes, and market-shaping work. If you want pandemic preparedness, reward verified improvements in surveillance, cheap diagnostics, vaccine platform readiness, PPE resilience, or rapid clinical trial capacity. If you want AI governance capacity, reward usable evals, security benchmarks, model-control tools, compute-accounting systems, or policy infrastructure that actually gets adopted. If you want global health impact, pay for credible QALYs or DALYs averted, while being explicit about the moral weights and assumptions underneath. This is not a magic bullet. Pull funding Goodharts whatever it measures and favors legible outcomes. But it has one huge virtue: it forces the Foundation to say what it actually wants. If the Foundation pays for QALYs, animal welfare improvements, verified safety evals, or reductions in catastrophic risk, then people can argue about those metrics directly instead of reverse-engineering the worldview of a grants committee. For every major pull-funding program, I would reserve 5–10% of the budget for adversarial audits: rewards for showing how the metric can be gamed, why the measured outcome is not the real outcome, or why the program is selecting for fake impact.
Second, a push-funding arm. Some things cannot be bought through clean outcome contracts. You sometimes need to fund inputs: weird researchers, early science, institution-building, field creation, adversarial work, and long-horizon bets where the output is not immediately measurable. But push funding should be treated as the dangerous, high-discretion part of the portfolio, not the default. Every major push grant should come with a public theory of change, a forecast distribution over key outcomes, conflict-of-interest disclosures, and a plan for retrospective evaluation. Here I would borrow from Squiggle, Guesstimate, Metaculus, QURI, the longtermist wiki/crux project and the broader EA modeling tradition. The goal is not to pretend these models are precise. The goal is to make uncertainty explicit enough that people can find the weak points. If a grant depends on “this reduces p(doom) by 0.01%,” say that. If it depends on shrimp having nontrivial moral weight, say that. If it depends on institutional lock-in being more important than technical alignment, say that. Then pay smart critics to attack the model.
Third, an infrastructure-and-audit arm. This is the least glamorous and probably the highest-leverage part. The Foundation should build a grantmaking stack that includes financial audits, evidence synthesis, reference-class forecasting, red-team review, prediction markets, grant outcome tracking, and public postmortems. It should maintain a live map of cause areas, interventions, assumptions, evidence quality, and open cruxes.
The OpenAI conflict requires special rules. The Foundation should not fund evals, governance work, safety audits, or policy organizations that may affect OpenAI through ordinary discretionary grantmaking. Those grants should go through an independently governed firewall: external reviewers, public recusals, guaranteed publication rights, and a presumption that negative findings can be published. Otherwise, even good grants will look like reputation laundering or soft capture.
In the first six months, I would make only continuation grants and small exploratory grants. The main work would be hiring mechanism designers, economists, forecasters, auditors, AI safety people, domain experts, and institutional skeptics.
In year one, I would launch pilot programs: maybe $100–300 million across pull-funding experiments, 50 mil model-based push grants, and 300-500 mil audit systems. The goal would not be to maximize immediate impact. The goal would be calibration: which mechanisms produce real information, which get gamed, which attract talent, and which reveal hidden bottlenecks?
In years two and three, I would scale only mechanisms that survive adversarial review. The Foundation could then begin spending billions annually, but only through channels that have been stress-tested. I would heavily cap opaque discretionary grantmaking (with some sort of push through mechanism requiring super majority of disagreeable people) and require retrospective public evaluation for large grants.
The Foundation’s comparative advantage is not just money. It has the capability to be the most tech savy/automated/inference dense granter ever. If it just becomes a giant grantmaker with more zeros, it will lock in the worldview and social network of whoever happens to be close to the money. If it builds transparent pull funding, disciplined push funding, and serious audit infrastructure, it can make many other actors smarter too.
Hmm well isn’t this basically the core project of EA aha? anyway thanks for sharing I chatted my two cents into chaptgpt 5.5 and had it format/write it out for me.
The Foundation should target market failures: public goods, externalities, information failures, coordination failures, etc. That means being suspicious of “important but already incentivized” domains. I start with this because of an example they have already expressed interest in—Alzheimer’s is important, but it is not obviously where a new mega-foundation has the highest marginal leverage. There are already enormous commercial incentives to develop effective treatments: rich patients, aging societies, and pharmaceutical companies all badly want a breakthrough. The philanthropic opportunity is not “Alzheimer’s” in general, but specific neglected bottlenecks inside Alzheimer’s: open datasets, repurposed generics, trial infrastructure, non-patentable interventions, prizeable biomarkers, fixing and standardizing medical ontologies, or other areas where private returns diverge from social returns.
The bigger bottleneck is that we do not have a trustworthy system for converting hundreds of billions into impact. Conventional grantmaking is too opaque, too relationship-driven, too vulnerable to value lock-in, and too dependent on a small number of people’s cached worldviews. If you give that system $180 billion and scale quickly, you mostly get larger versions of the same failure modes.
So my first rule would be: do not spend the endowment quickly (yes I also have short timelines so need to balance that but won’t get into that rn). For the first two or three years, spend a small fraction building the allocation machine: the mechanisms, audits, forecasting systems, and public models that make later spending less dumb. The Foundation should have three arms.
First, a pull-funding arm. This should be the biggest. Wherever outcomes can be specified reasonably well, the Foundation should stop trying to guess the best grantees ex ante and instead pay for results. This is the logic behind results-based financing, advance market commitments, prizes, and market-shaping work. If you want pandemic preparedness, reward verified improvements in surveillance, cheap diagnostics, vaccine platform readiness, PPE resilience, or rapid clinical trial capacity. If you want AI governance capacity, reward usable evals, security benchmarks, model-control tools, compute-accounting systems, or policy infrastructure that actually gets adopted. If you want global health impact, pay for credible QALYs or DALYs averted, while being explicit about the moral weights and assumptions underneath. This is not a magic bullet. Pull funding Goodharts whatever it measures and favors legible outcomes. But it has one huge virtue: it forces the Foundation to say what it actually wants. If the Foundation pays for QALYs, animal welfare improvements, verified safety evals, or reductions in catastrophic risk, then people can argue about those metrics directly instead of reverse-engineering the worldview of a grants committee. For every major pull-funding program, I would reserve 5–10% of the budget for adversarial audits: rewards for showing how the metric can be gamed, why the measured outcome is not the real outcome, or why the program is selecting for fake impact.
Second, a push-funding arm. Some things cannot be bought through clean outcome contracts. You sometimes need to fund inputs: weird researchers, early science, institution-building, field creation, adversarial work, and long-horizon bets where the output is not immediately measurable. But push funding should be treated as the dangerous, high-discretion part of the portfolio, not the default. Every major push grant should come with a public theory of change, a forecast distribution over key outcomes, conflict-of-interest disclosures, and a plan for retrospective evaluation. Here I would borrow from Squiggle, Guesstimate, Metaculus, QURI, the longtermist wiki/crux project and the broader EA modeling tradition. The goal is not to pretend these models are precise. The goal is to make uncertainty explicit enough that people can find the weak points. If a grant depends on “this reduces p(doom) by 0.01%,” say that. If it depends on shrimp having nontrivial moral weight, say that. If it depends on institutional lock-in being more important than technical alignment, say that. Then pay smart critics to attack the model.
Third, an infrastructure-and-audit arm. This is the least glamorous and probably the highest-leverage part. The Foundation should build a grantmaking stack that includes financial audits, evidence synthesis, reference-class forecasting, red-team review, prediction markets, grant outcome tracking, and public postmortems. It should maintain a live map of cause areas, interventions, assumptions, evidence quality, and open cruxes.
The OpenAI conflict requires special rules. The Foundation should not fund evals, governance work, safety audits, or policy organizations that may affect OpenAI through ordinary discretionary grantmaking. Those grants should go through an independently governed firewall: external reviewers, public recusals, guaranteed publication rights, and a presumption that negative findings can be published. Otherwise, even good grants will look like reputation laundering or soft capture.
In the first six months, I would make only continuation grants and small exploratory grants. The main work would be hiring mechanism designers, economists, forecasters, auditors, AI safety people, domain experts, and institutional skeptics.
In year one, I would launch pilot programs: maybe $100–300 million across pull-funding experiments, 50 mil model-based push grants, and 300-500 mil audit systems. The goal would not be to maximize immediate impact. The goal would be calibration: which mechanisms produce real information, which get gamed, which attract talent, and which reveal hidden bottlenecks?
In years two and three, I would scale only mechanisms that survive adversarial review. The Foundation could then begin spending billions annually, but only through channels that have been stress-tested. I would heavily cap opaque discretionary grantmaking (with some sort of push through mechanism requiring super majority of disagreeable people) and require retrospective public evaluation for large grants.
The Foundation’s comparative advantage is not just money. It has the capability to be the most tech savy/automated/inference dense granter ever. If it just becomes a giant grantmaker with more zeros, it will lock in the worldview and social network of whoever happens to be close to the money. If it builds transparent pull funding, disciplined push funding, and serious audit infrastructure, it can make many other actors smarter too.