Summary
AI may produce a windfall surge in government revenues in 5 to 10 years. I want governments to spending a small fraction of that windfall on retroactively rewarding entities in proportion to how they have contributed to medical advances, measured by lives saved and suffering avoided.
Motivations
This post was inspired in part by Critch:
Healthcare, agriculture, food science, education, entertainment, and environmental restoration are all important industries that serve humans but not machines. These are industries I want to sustain and advance, in order to keep the economy caring for humans, and to avoid successionism and industrial dehumanization.
I want to mitigate panic over the coming mass retirement. For now, most people imagine that AI-induced unemployment will be a catastrophe. Some of that is because they can’t see how to pay for medical care in a world where most people depend on welfare. Medicaid and Medicare look unsustainable in a world where more people depend on those systems, and the cost of medical care keeps rising. That leads to voters being somewhat panicky about AI in a way that fails to focus their attention on the biggest risks of AI (i.e. world domination by a few AIs or by a few AI-wielding people).
A final motivation is concern that medical progress is too slow to handle my age-related health risks. I’ll highlight two contributing problems which I expect my proposal will mitigate:
- There’s as much reward for marginally beneficial treatments as for full cures. Since the former seem easier to find, companies have largely ignored the most valuable goals.
- Medical research is too dependent on patents for funding. There are likely a number of cheap, effective treatments that are being overlooked because either they’re unpatentable, or their patents have expired.
2025 is a unique time for this proposal: a small fraction of investors are seeking big AI-related profits. If they’re right, we can expect to afford much more medical research now than is currently happening. These investors forecast a dramatic increase in productivity, which seems like it would generate unprecedented economic growth, presumably enabling a dramatic increase in government spending. There are enough such investors to make a real difference in medical progress. I think it’s possible that my proposal could redirect several billion dollars to research that is valuable and neglected.
Designing the Prizes
I propose allocating 1% of windfall US government revenues between now and 2040 toward these prizes. By windfall, I mean revenues that are clearly in excess of what would be expected if future growth is similar to historical growth. E.g. if revenue growth accelerates to 15%/year starting in 2028, revenue in 2032 would be around $16.5 trillion, versus $6.2 trillion expected under normal growth. I’d classify $6 trillion of that as a windfall, for a prize fund of $60 billion that year. The appendix describes the calculation in more detail.
Who gets these prizes? I want a panel of 3 to 5 of the most respected AIs to decide. In any given windfall year, I want them to estimate who is responsible for medical advances, and divide the prize fund equally per QALY saved.
I want to leave most of the details to future AIs, who I expect will be wiser and more trustworthy than current institutions.
Benefits
The prizes would cause some investor money to be redirected toward ensuring that many people are healthy.
My proposal would reduce the risk of gradual disempowerment of the average person.
We’re heading toward a world where most institutions may end up focused almost entirely on serving machines. Many of those institutions will measure their success by criteria such as profits or donations. These prizes will ensure that some institutions measure at least some of their success by a criterion that is more closely connected to human welfare.
My guess is that gradual disempowerment will be not prohibitively hard to avoid once we have AIs that are modestly smarter than the smartest humans to advise us on how to coordinate. But I’m uncertain enough about that that I want to add in whatever precautions we can implement now.
My proposal would mitigate retirement fears by significantly increasing research into treatments that cheaply prevent cancer, cardiovascular disease, etc. That’s a step toward ensuring that programs such as Medicaid won’t collapse due to prohibitive costs. The offer of prizes wouldn’t be immediately reassuring to the large number of people who doubt that economic growth will accelerate. However, I expect that by the time AI causes noticeable unemployment, they’d see some shift in corporate strategies toward cheaper and more effective treatments.
My proposal would divert a modest amount of investment away from a general AI capabilities race, providing modestly more time for safety research. This effect will be less than I hope for, but something is better than nothing.
Examples
Most of my hopes for better medicine come from the discovery of treatments that I haven’t yet imagined. But I’ll try to whet your appetite by suggesting some examples of currently known treatments that might be used more widely as a result of the improved incentives from this prize fund. Remember that I’m just suggesting possibilities. I’m not willing to bet heavily on the value of any specific example.
Healthy Meals: A wide variety of age-related diseases are influenced by diet, most notably diabetes.
I imagine a company that sells convenient healthy prepared meals at below cost. Current institutions aren’t good enough to compensate the company for the health benefits, but it seems likely that transformative AIs will be wise enough to handle this fairly well.
Affordable Weight Loss: GLP-1 drugs appear to have widespread health benefits, yet lots of people who might benefit from one are unable to obtain one. The obstacles typically boil down to cost.
Conceivably the prizes could persuade pharma companies to reduce GLP-1 drug prices, in order to get prize money by getting the drugs to more patients. But I don’t expect pharma companies to take AI seriously enough for that effect to be significant.
AI-pilled investors could subsidize prescriptions. I don’t have a clear intuition as to how well that would work.
A somewhat more promising strategy involves Modere’s Curb (a supplement that influences GLP-1). A startup funded by AI-pilled VCs could buy Modere, pump lots of money into advertising Curb, and cut Curb’s price a bit. They’d lose moderate amounts of money for the next few years, for what they’d see as more than a 10% chance of getting $1 billion in windfall prize money.
Malaria Vaccine: A malaria vaccine recently became available. As far as I can tell, it is not being rolled out as fast as it would be if the world treated malaria deaths as a crisis. It seems likely that applying more money to the process could cheaply save lives here.
The malaria example assumes that the prizes would be awarded based on global lives saved, but I don’t intend to express a strong opinion as to whether we should aim for that, versus a more politically feasible national program.
I expect that exercise and mental health therapy are two other areas where there’s big room for improvement if the incentives become sufficiently good. I don’t have clear enough ideas there to produce realistic proposals.
As you can see from these examples, I’m not sticking to a traditional notion of medicine. I want the prize committee to focus on QALY’s saved, and use a broad definition of healthcare when deciding what treatments to reward.
Problems
A promise from Trump (as with Biden) is not, by itself, worth the tweet it’s written on.
A law passed by congress is worth a bit more, but still not as much as I would like.
I see some risk that the proposal will cause AI companies to insert biases into their AIs about how the prizes should be allocated. This seems not much of a problem at the early stages of implementation, when companies that are competing to get prizes will have trouble predicting the characteristics of the AIs that will judge the prizes. It will become a bit more of a concern later on. Maybe there needs to be some randomness in which AIs get selected for the panel?
Conclusion
I’ve designed this proposal so that there will be minimal opposition to congress enacting it into law: it looks costless under the normal scenarios that most people use to evaluate such proposals, and very affordable under more optimistic scenarios. The downside is that benefits also look fairly unlikely under scenarios that assume normalcy, so approximately zero politicians are likely to devote much thought to promoting it.
Weaker versions of this proposal are possible. E.g. Bill Gates could promise 10% of the windfall profits from his foundation’s investments.
Appendix
Here’s an outline of how I calculated the prize fund:
ExpectedRevenue(Y) = BaselineRevenue × (1 + B)^Y * (1 + ?)
Windfall(Year Y) = ActualRevenue(Y) – ExpectedRevenue(Y)
RewardableWindfall(Y) = max(0, Windfall(Y) – ThresholdPercentage × ExpectedRevenue(Y))
Where:
- BaselineRevenue = Tax revenue from the year before policy implementation
- B = Historical baseline growth rate
- ? = Additional buffer (perhaps 2 standard deviations of historical growth variation)
- Y = Years since policy implementation
I neglected inflation. An actual law should adjust for inflation.
I’ll put some code that I used in a comment.
What time period should the prize fund from any given year apply to?
Let’s imagine an example case: The law was enacted on January 1, 2025. In 2031, the prize fund is $10 billion. In 2032, the prize fund is $50 billion.
I suggest applying the $10 billion solely to reward actions made in 2025 that save QALYs compared to the baseline. Then in 2032, apply the $50 billion to reward actions from both 2025 and 2026. And so on, up to a total of 5 historical years being rewarded each year.
There should probably be some provision for altering the prize fund after, say, 2040, since it will become less appropriate to use forecasts from 2025 to classify revenues as a windfall. Maybe set it to some stable value that’s modestly lower than the 2040 prize fund?
I’m tempted to offer prize money to legislators and FDA employees who help with these advances, but I’m concerned that conflict of interest rules would make that inappropriate.
import pandas as pd import numpy as np from typing import Optional class WindfallCalculator: def __init__(self, historical_data: pd.DataFrame, threshold_percentage: float = 0.20, sigma_multiplier: float = 2.0): """ Initialize the windfall calculator with historical data. Args: historical_data: DataFrame with columns ['year', 'revenue'] threshold_percentage: Additional threshold above expected revenue (default 20%) sigma_multiplier: Number of standard deviations to add to baseline (default 2) """ self.data = historical_data self.threshold_percentage = threshold_percentage self.sigma_multiplier = sigma_multiplier # Calculate historical growth rates self.data['growth_rate'] = self.data['revenue'].pct_change() # Calculate baseline parameters self.baseline_growth = self.data['growth_rate'].mean() self.growth_std = self.data['growth_rate'].std() self.baseline_revenue = self.data['revenue'].iloc[-1] self.baseline_year = int(self.data['year'].iloc[-1][:4]) def expected_revenue(self, year: int) -> float: """Calculate expected revenue for a given year based on baseline growth.""" years_from_baseline = year - self.baseline_year simple = self.baseline_revenue * (1 + self.baseline_growth) ** years_from_baseline buffered = simple * (1 + self.growth_std * self.sigma_multiplier) return (simple, buffered) def calculate_windfall(self, year: int, actual_revenue: float) -> dict: """ Calculate windfall and reward payment for a given year's actual revenue. Args: year: The year to calculate for actual_revenue: The actual tax revenue for that year Returns: Dictionary containing calculations and intermediate values """ (simple_expected, expected) = self.expected_revenue(year) raw_windfall = max(0, actual_revenue - expected) threshold = self.threshold_percentage * expected rewardable_windfall = max(0, raw_windfall - threshold) reward_payment = 0.01 * rewardable_windfall return { 'year': year, 'actual_revenue': actual_revenue, 'expected_revenue': expected, 'raw_windfall': raw_windfall, 'threshold': threshold, 'rewardable_windfall': rewardable_windfall, 'reward_payment': reward_payment } def print_parameters(self): """Print the calculated baseline parameters.""" print(f"Historical baseline growth rate: {self.baseline_growth:.1%}") print(f"Growth standard deviation: {self.growth_std:.1%}") print(f"Growth buffer (baseline + {self.sigma_multiplier}?): {(self.baseline_growth + self.sigma_multiplier * self.growth_std):.1%}") print(f"Threshold percentage above expected: {self.threshold_percentage:.1%}") print(f"Baseline revenue (from {self.baseline_year}): ${self.baseline_revenue:.1f}B") # Example usage with the FRED data: def load_fred_data(csv_path: str) -> pd.DataFrame: """Load and preprocess the FRED tax revenue data.""" df = pd.read_csv(csv_path) df.columns = ['year', 'revenue'] return df # Example scenarios def run_scenarios(calculator: WindfallCalculator): """Run some example scenarios to demonstrate the calculator.""" scenarios = [ (2026, 2025, 1.12), (2026, 2025, 1.25), (2032, 2028, 1.15), (2032, 2028, 1.30), ] print(f"\nScenario Analysis") print("-" * 64) for target_year, inflection_year, growth_rate in scenarios: multiplier = (1 + calculator.baseline_growth) ** (inflection_year - calculator.baseline_year) multiplier *= growth_rate ** (target_year - inflection_year) (simple_expected, expected) = calculator.expected_revenue(target_year) actual = simple_expected * multiplier result = calculator.calculate_windfall(target_year, actual) print(f"\nScenario: {target_year} with growth rate {(growth_rate-1):.0%} starting in {inflection_year}; multiplier {multiplier:.1f}") print(f"Actual Revenue: ${result['actual_revenue']:.1f}B") print(f"Expected Revenue: ${simple_expected:.1f}B") print(f"Two Sigma Revenue: ${result['expected_revenue']:.1f}B") print(f"Raw Windfall: ${result['raw_windfall']:.1f}B") print(f"Rewardable Windfall: ${result['rewardable_windfall']:.1f}B") print(f"Reward Payment: ${result['reward_payment']:.1f}B") if __name__ == "__main__": # Load historical data, from https://fred.stlouisfed.org/series/W006RC1A027NBEA data = load_fred_data('/home/pcm/invest/data/econ/stls/w/W006RC1A027NBEA.csv') # Create calculator calculator = WindfallCalculator(data) # Print baseline parameters calculator.print_parameters() # Run example scenarios run_scenarios(calculator)