Long tasks take models longer, causing failures due to distribution shift or self conditioning (which models may suffer from).
I am not sure if I understand this correctly, and I am not an expert, but my interpretation is that if the AI makes a mistakes, the mistakes start accumulating, because it sees a mistake in the previous text, so the correct prediction of the existing text is making more mistakes. I mean, if someone makes a mistake in the first step, they are more likely to make a mistake in every step, rather than make no more mistakes.
I wonder if it may also be related to training. Is it ever rewarded when an AI makes a mistake and then notices and corrects it? If not, then perhaps it does not know how to correct its own mistakes. And the first mistake (which is more likely to happen in a longer task) then becomes fatal.
I am not sure if I understand this correctly, and I am not an expert, but my interpretation is that if the AI makes a mistakes, the mistakes start accumulating, because it sees a mistake in the previous text, so the correct prediction of the existing text is making more mistakes. I mean, if someone makes a mistake in the first step, they are more likely to make a mistake in every step, rather than make no more mistakes.
I wonder if it may also be related to training. Is it ever rewarded when an AI makes a mistake and then notices and corrects it? If not, then perhaps it does not know how to correct its own mistakes. And the first mistake (which is more likely to happen in a longer task) then becomes fatal.