there was the red button/blue button twitter poll. It sparked a huge discussion. Rules were: if more >50% people press blue button everybody lives, if <50% only red pressers live (ie. red pressers always survive and blue if only >50% other people choose blue)

there are many different frames why to press one or the other. One would be game theory, where the red is the dominant and rational action

• but there are a lot other see more

the decisions also hinges a lot on how you model the society, what other people would press. if you give it to ai:

One way to think about [LLMs] is that they contain the ability to simulate many different characters or personas.

See The persona selection model (Anthropic)

reasoning-trained models converge on red while helpful-assistant models lean blue. and this mirrors how training shapes behavior. RLHF gave us helpful assistants; RLVR (verifiable rewards on math/code) gave us reasoners. Neither produces robust social reasoning.

Softmax's actual pitch is a third training paradigm: RLSE — Reinforcement Learning from Social Evolution. The idea is that just as coding skills jumped when models got trained in environments with executable feedback (not just by reading GitHub), social/cooperative reasoning probably needs environments where multi-agent dynamics — bargaining, reputation, trust repair, coalition formation, public goods — are real and provide structural feedback.