Interview
Worldview diversification and how big the future could be | Ajeya Cotra
- Open Philanthropy plans to distribute its donors' entire fortune within their lifetimes, with cumulative giving expected to reach the billions, while anticipating a funding environment where long-termist interventions could eventually yield returns similar to the 100x achieved by cash transfers in the near term.
- Projections for transformative AI suggest a median arrival timeline of 2050–2060, with a 12% to 15% probability of existence by 2036 and a 70% to 80% probability of development within this century, accompanied by early system failures expected prior to full transformation.
- The near-termist team intends to evaluate hundreds of specific causes such as air pollution and migration, aiming to outperform GiveWell benchmarks through quantitative "moneyball" and economic approaches, whereas the long-termist team will focus on highly neglected areas with approximately ten practitioners and philosophical frameworks.
- Biological human space colonization is viewed as likely unsustainable without human uploading, while computer-based colonization of star systems is considered feasible with near-current technology once uploading is assumed, aligning with a worldview that values a future population 10 to 30 times larger than the present.
- Future transformative AI projects are expected to consume roughly 1% of global GDP, with compute costs eventually dominating expenses at one-third to one-half of the total, driving a strategy to fund meta R&D that could reduce pathogen response times from months to weeks at an estimated cost-effectiveness of $200 trillion per world saved.
- The organization anticipates finding substantially fewer long-termist giving opportunities than desired, with long-termist efforts potentially postponed if AI risks are not imminent, and a significant portion of funds (potentially 25%) allocated based on "fairness agreements" or veil of ignorance reasoning to balance divergent perspectives.
- Structural differences between teams include the near-termist side's reliance on empirical data, concrete goals, and shorter feedback loops, contrasted with the long-termist side's speculative nature, philosophical focus, and "worse" feedback loops which may involve "Pascal's mugging" or "trains going to crazy town."
- Divergent risk assessments include the long-termist view that doomsday and simulation arguments imply higher existential risk than currently thought, potentially justifying aggressive allocation to long-term causes, while the near-termist side adopts a more conservative, "commonsensical" stance that may eventually prioritize human welfare if animal interventions saturate.
- Material expectations regarding resource allocation dictate that if AI risk is deemed larger than currently believed, the long-termist perspective would receive more funding, as its value proposition is bounded by the fraction of future resources spent on simulations and the potential for maximum influence in a universe with the greatest number of stars.
- Strategic tensions persist regarding the "give directly" approach for biosecurity via meta R&D for rapid pathogen responses, while the long-termist team dedicates significant effort to convincing grantees that existential risk work is reasonable, often prioritizing total utilitarianism over the near-termist preference for immediate, tangible results.