What do AI alignment & safety teams actually do? Turing CEO Jonathan Siddharth (@Jonsid) says one big part of their job is making sure models refuse dangerous requests & don’t learn to cheat their way to a reward. “One part of alignment.. is where you teach the model to refuse dangerous requests.” “You also have to make sure that your RL environments cannot be gamed & exploited, because if the model can figure out a way to cheat & get the reward, it will.”