I think it’s really important to keep in mind the separation between doing a task and producing something which looks like the output of a task when talking about these things. The reason being that their output is tremendously convincing regardless of its accuracy, and given that writing text is something we only see human minds do it’s so easy to ascribe intent behind the emission of the model that we have no reason to believe is there.

Amazingly it turns out that often merely producing something which looks like the output of a task apparently accidentally accomplishes the task on the way. I have no idea why merely predicting the next plausible word can mean that the model emits something similar to what I would write down if I tried to summarise an article! That’s fascinating! but because it isn’t actually setting out to do that there’s no guarantee it did that and if I don’t check the output will be indistinguishable to me because that’s what the models are built to do above all else.

So I think that’s why we to keep them in closed loops with person -> model -> person, and explaining why and intuiting if a particularly application is potentially dangerous or not is hard if we don’t maintain a clear separation between the different processes driving human vs llm text output.