Dwarkesh Patel received 600 essay submissions after posing a set of open questions about AI’s biggest unresolved decisions, then named three winners: Jassi Pannu, an assistant professor at Johns Hopkins University; Ege Erdil, co-founder of Mechanize; and Michael Li, a Master of Public Policy candidate at Harvard Kennedy School. The contest asked entrants to tackle questions with no consensus answer: what a lab-funded foundation should prioritize, how bystander countries should respond to transformative growth, and how AI labs will ultimately turn a profit. Each winning essay supplies a concrete mechanism rather than a general prediction.
Pannu, who also sits on the board of Blueprint Biosecurity, answered the question of what the OpenAI Foundation should fund. Her proposal: roughly $40 billion to $60 billion over ten years to deploy far-UVC lighting, a wavelength safe for humans but lethal to airborne pathogens, across schools, hospitals, and transit hubs in OECD countries. She frames the case around a “dual-payoff principle”: interventions that reduce catastrophic pandemic risk while also delivering measurable everyday gains, in this case an estimated 60 percent cut in seasonal flu mortality by year ten. Her essay leans on the precedent of smallpox eradication, noting the gap between Edward Jenner’s 1796 vaccination breakthrough and the successful 1967 eradication campaign was closer to a coordination failure than a technology gap.
Erdil, previously a researcher at Epoch AI, addressed a different question: what should countries outside the AI supply chain do to avoid being left behind by automation-driven growth. His answer avoids the more speculative proposals the contest apparently attracted, including one submission that suggested threatening China and the United States over their chip fabs and data centers. Erdil instead argues for conventional growth policy: strong property rights, low capital taxes, and light-touch regulation, on the logic that these levers become more powerful, not less, once capital can substitute for labor entirely. He is candid that most dysfunctional governments will not adopt these reforms, and that his realistic path for lagging countries is less about active reform and more about incumbents like the United States and China making policy mistakes that create an opening.
Li, who writes the blog Ceteris Paribus, took on the question of how AI labs will make money given that inference pricing keeps falling toward marginal cost. He compares the labs to Hong Kong’s Mass Transit Railway, a transit operator that never recovers construction costs through fares alone but funds expansion by owning and developing the real estate around its stations. Applied to AI, Li’s argument is that labs should stop expecting API revenue to fund training runs and instead pursue four categories of “adjacent property”: government-granted deployment rights over sensitive data systems, accumulated reinforcement-learning data that does not depreciate the way model weights do, forward-deployed service delivery that captures the value AI creates rather than charging per token, and formal data-trusteeship arrangements over datasets like patient records.
The three essays converge on a shared premise: the biggest near-term AI decisions will be made through infrastructure and policy design rather than through model capability alone. Pannu’s proposal treats biosecurity as physical infrastructure, Erdil’s treats national competitiveness as a policy-stability problem, and Li’s treats lab profitability as a real-estate problem in disguise. None of the three essays argues that better AI models alone determine these outcomes.
For operators and policymakers, the practical takeaway is where to look for AI’s economic center of gravity over the next decade. Li’s framing suggests the AI labs most likely to become durable businesses are the ones securing exclusive deployment rights and data-trusteeship arrangements now, not the ones with the lowest API price.
Reported by Dwarkesh Patel on his blog, “The Winning Essays for the Big Questions About AI.”
