In short: I took the EA Brazil course and have been looking for a cause to invest in using pairwise comparisons. I created a brief personal checklist to decide whether I should work directly on AI safety.
Effective altruism poses a question I find really difficult: should I try to directly combat AI risk, or invest in something smaller and safer?
I set four conditions for myself. If I met any of them, I would consider prioritizing direct work on AI:
1- I can design a version of human moral reasoning that is computationally efficient enough to be used in practice. (80-90% confidence)
2- I can understand how current AI systems process information well enough to contribute technically within at least 10 years. (80-90% confidence)
3- I have enough power to have a chance of stopping two superpowers in a prisoner's dilemma scenario? (60-70% confidence)
4- I can create a competitive lab capable of matching a cutting-edge AI company, in order to protect AIs using other AIs. (60-70% confidence)
I meet none of these criteria.
This reduces "Buck's personal critical points for working on AI safety" to a single question: am I the right person for the job? It is no longer about whether the work itself matters.
So, since it seems I cannot update my beliefs enough to invest in AI, I have invested in something else:
In Samambaia, Brazil, I am adapting the Friendship Bench model (originally created to treat depression in Zimbabwe) to train adult community mentors to lead peer-to-peer emotional support sessions.
For the Patronato de Liberados (Parole and Probation Board) in Argentina, I am designing a training course for peer mentors who will support young people involved with the justice system.
Open question: when does local "training of trainers" truly substitute for direct work on AI safety, rather than serving as a convenient excuse? My assumption: only when training trainers, not just beneficiaries.
