Comment by golol

Because it's a straight forward stochastic sequence modelling task and I've seen GPT-3.5-turbo-instruct play at high amateur level myself. But it seems like all the RLHF and distillation that is done on newer models destroys that ability.