I recall that they trained a part of an insect's head like the nasal passage and it was able to be used for language better than the model at that time chatGPT2. So there's something innate in nature that can learn human languages.
I think you may have slightly over-sold the study, though -- or at least what you remembered from it.
My reading of the first study was that they took a simulated version of a relatively large (2000 node) neutral network that makes up part of a fruitfly brain, and were able to do standard neutral network training on it to do some language prediction.
I'm not sure that this says anything about fruitfly noses being wired for language though. I expect that they could have taken that same simulated architecture and trained it to do anything that regular neutral networks could do -- detect faces, make stock market predictions, play a (poor) game of Go, or learn a homeowner's thermostat patterns.
I think it's just more a statement about the power of neutral networks in general.
I did misremember, it seems they used a network of algorithms based on how flies use 2000 neurons and the original article I read it from a while ago may have oversold it. The fact it was able to predict well is them looking for inspiration in insects, and that they'll be using more insect inspired behavior.
