[ux266478]

The distinction does make sense, as castles (formally speaking) are medieval structures, but clearly that didn't matter because the map has other ancient acropoleis. Even putting that aside, the relationship between 'aκρόπολη' and 'castle' is a pretty trivial latent feature that a transformer is literally, mathematically designed to capture. Briefly interrogated an LLM on the matter and they struggled with the classification too. Definitely an interesting failure point. I wonder if it's because of semantic dilution. Like, because the training data had so much on the Athens acropolis which is non-castle related, that the fact ended up living in a long tail for that specific acropolis. I bet a higher precision model wouldn't have done this. Quantization sucks!