2 ms·
It's somewhat like what we had before DNN based LMs, where humans were trying to craft grammar trees that could account for all the ways of saying something. Th
by verdverm 4y ago
It's somewhat like what we had before DNN based LMs, where humans were trying to craft grammar trees that could account for all the ways of saying something. There are a lot of ways to give instructions, and it is hard for the developers of these LLMs to anticipate them all. That's how the indirection in the article's example works. If you know a line in the hidden pretext provided to the model, you can get at the text by asking it to recite the surrounding text, rather than asking it to explain the text or the rules. Something quite literal