A fascinating discovery in an Icelandic-language Facebook group: the phrase “Það eru margar undur í höfuðkúpu”. A bunch of online articles will teach you this delightful saying and explain that it means “There are many wonders in a cow’s head”. Only, oddly enough, none of the results for the full phrase are Icelandic websites. It’s all blogs, fiction and social media posts by foreigners…
That would be because “Það eru margar undur í höfuðkúpu” is not a real Icelandic saying. Not only is it not a real saying, it’s grammatical nonsense: “margar” is the feminine plural of the word many, but “undur” meaning “wonder” is a neuter word. Moreover, “Það eru margar undur í höfuðkúpu” features absolutely no mention of a cow: it just means “There are many wonder in a skull”.
However, “There are many wonders in a cow’s head” is recognizably a translation of a genuine Icelandic saying - it’s just decidedly not that one. The real saying is “Það er margt skrýtið í kýrhausnum” (literally “There’s a lot of strange stuff in the cow’s head”). Somehow, somewhere along the way, someone had heard there was an Icelandic saying that meant “There are many wonders in a cow’s head” and just ran that back through a machine translation and called it a day - and then a bunch of other websites aped it after them without even a cursory fact-check by anyone who actually knows the language.
Fun fact: I bet the reason the cow disappeared from the machine translation is that the word “höfuðkúpu” happens to contain the letters -kú-, which is coincidentally the accusative and dative of the word kýr. Neural network sees a token it associates with cows in there and just figures yeah, checks out. Likewise, the grammatical error makes some warped sense for a neural network to output: while “undur” the word meaning wonder is a neuter word, many feminine words genuinely end in -ur in the plural.
(Sponsored by the Icelandic government, Icelandic linguistic tech company Miðeind collaborated with OpenAI to make GPT-4 understand and write Icelandic. They contributed a deluge of training data of Icelandic text and reinforcement learning. The results were initially pretty disappointing - because GPT is pre-trained on text scraped from the internet, and the vast majority of purportedly Icelandic text on the internet is machine-translated slop, because there just aren’t enough actual Icelandic speakers and Icelandic websites to drown out all the spam. Very recently, though, they managed to genuinely substantially improve it.)
snowghoul
annaxmalina
vaporwave
andulkaphoto
blakekathryn
yousharknotpass
