3 ms·
Surely this is more related to the fact that "May" has multiple meanings, but "December" is quite clearly just the name of a month. Does it happen with April a
by SeanAnderson 3y ago
Surely this is more related to the fact that "May" has multiple meanings, but "December" is quite clearly just the name of a month.
Does it happen with April and June? Does the response length trend downward as the months approach December? Does it change its behavior if it believes its in a hemisphere where May is snowy?
If it were performing better on standardized testing then I would find the results compelling, but generating longer responses isn't even necessarily desirable.
- recursive 3y ago> Surely this is more related to the fact that "May" has multiple meanings. Surely? Whence the certainty? If humans act this way, and the training data is mostly from humans, I would be surprised if this wasn't a real effect.
- SeanAnderson 3y agoOkay, fair. I'm not especially confident. I just would've expected a little more to back it up.
- vineyardmike 3y agoWell I can surely say that you’re wrong. Mostly because the source appears to be using numerical dates not the English spelling of the month name. But yea, the data could be better (but the hypothesis is very plausible, considering the training data).