>> The arithmetic questions only have 1000 tests where they add two digits. This
is certainly in the training data.
Yeah, it's confirmation bias. People do that sort of thing all the time in machine learning research, specially in the recent surge of LLM-poking papers. If they don't do that, they don't have a paper, so we're going to see much more of it before the trend exhausts itself.
Yeah, it's confirmation bias. People do that sort of thing all the time in machine learning research, specially in the recent surge of LLM-poking papers. If they don't do that, they don't have a paper, so we're going to see much more of it before the trend exhausts itself.