paper / natashajaques / Jun 10
Current AI development paradigms focus on capability via a solipsistic approach, treating the world as a static feedback source. This leads to superintelligence that is unlikely to be cooperative due to the self-undermining property of unilateral optimization, where AI deployment introduces non-stationarity and a train-test-deploy gap. A non-solipsistic paradigm is needed that designs interdependence and cooperation as core principles, requiring dynamic evaluations, institutional design, and preservation of human agency.
ai-safetysuperintelligencecooperationai-ethicsai-alignment
“Superintelligence developed through a solipsistic approach to AI design, which treats the world as an external and static feedback source, is unlikely to be cooperative.”
paper / natashajaques / Jun 10
This work introduces Learn2Match, a novel benchmark and framework for two-sided matching markets with temporally extended feedback. Unlike traditional models, it captures gradual information revelation and evolving preferences using a partially observable Markov game formulation. Experiments show that while multi-agent reinforcement learning methods like PPO improve social welfare and reduce regret compared to baselines, they still struggle with information friction, highlighting the need for algorithms that combine RL adaptiveness with the statistical rigor of bandit methods and the structure of stable-matching mechanisms.
matching-marketsreinforcement-learningmultiagent-systemsdynamic-systemspreference-learning
“Existing matching models fail to capture the dynamic nature of two-sided matching markets where information unfolds over time influencing decisions.”
paper / natashajaques / May 12 / failed
paper / natashajaques / May 12 / failed