I’ll have to find the post, but you did it in two steps, changed the units of mass, and object.

The post, which is extremely hard to find with the latest slop release from Nvidia, asked the chatbot to consider the exact wording, without babying it into the correct answer. All because the close variations of the phase “X pounds of bricks and X pounds of feathers weigh the exact same” have been used in various textbooks and such for at least the last hundred years or so.

That means that the chatbot has seen that exact combo of words, in roughly that order, quite a bit more than your use of “100 kilograms of rice”. At least in English.

You can baby it through when the training data is sparse, but not when there are hundreds of uses of the same phrase over and over again in the training.