something interesting is happening for big models, it is easy to generate candidate solutions for code but *verifying* that they 100% work is not
no fixed reward function is durable for a strong generator
more exciting work to be done around verifiers that coevolve with generators