I believe so, see my result with Haiku extended thinking on. I think the weights are just too biased towards blurping out the majority of the training data of 'next year is xxx'. Interesting problem to solve indeed.