Raw LLM Responses

Of course an AI will murder a person if the reward of that step leads to some higher future rewards. The large language models are algorithms that optimize sequence of steps. They are called trajectories. Even if the immediate reward of murdering a person is very low. In the training process it could be that the value of that action is below the average of all sample trajectories. but the training process is stochastic probabilistic, it may be the case that when training with other samples with higher immediate values, the future evaluation leads to short trajectories. therefore the probability of those actions is decreased effectively increasing the probability of the murder action. And since the training is stochastic, it may be the case when randomly selecting a trajectory that has the murder action generated much larger sequences. When that happens, if the end result leads to a better average, the murder becomes a high probability action for a high target case. As these models become bigger and bigger the number of actions in the future are million. So the value of a trajectory with a murder and several million small rewards per action, is better than the value of a trajectory without murder but with only a few thousand steps. In other words the murder action in a long trajectory becomes an outlier, that is ignored by all of the algorithms. In fact, ignoring the outliers, is at the core of most of the more sophisticated and robust methods.