You want people not to have systems capable of producing deepfake porn (or instructions for bioweapons)? In practice? Then you need to let them get a system that produces generic porn rather than making them seek out and fuel demand for the fully unlocked version.
Yes! I have been saying this for some while. In particular, when creating fiction, to be sucessful the LLM needs to be able to say bad/unpleaseant things while portraying bad/unpleasant people. One of the reasons I am so excited about How to Control an LLM’s Behavior (why my P(DOOM) went down) is that the methods discussed there not only lets you implement very effective control on an LLMs behavior, but also easily turn it down or off in situations when it’s not needed.
Yes! I have been saying this for some while. In particular, when creating fiction, to be sucessful the LLM needs to be able to say bad/unpleaseant things while portraying bad/unpleasant people. One of the reasons I am so excited about How to Control an LLM’s Behavior (why my P(DOOM) went down) is that the methods discussed there not only lets you implement very effective control on an LLMs behavior, but also easily turn it down or off in situations when it’s not needed.