The counting rs in strawberry problem was a example of people not understanding how the models work but I guess good to show the limitations of the current arch

The counting rs in strawberry problem was a example of people not understanding how the models work but I guess good to show the limitations of the current architectures. But thing is, those architectures haven't improved a whole lot. Now when it answers that correctly it's either in training data or by virtue of "count letters" or code sandbox tools.