Also, I think that looking at the lifetime of monitoring methods could be important when we consider the selection pressure of OpenAI changing the environment to make it harder to agents to collude. For example, try to investigate how much monitoring by screening we can get away with in a similar environment before agents start doing steganographic reasoning to bypass the monitor.
Also, I think that looking at the lifetime of monitoring methods could be important when we consider the selection pressure of OpenAI changing the environment to make it harder to agents to collude. For example, try to investigate how much monitoring by screening we can get away with in a similar environment before agents start doing steganographic reasoning to bypass the monitor.