Comment by Aachen

And presumably the energy efficiency of a dedicated translator compared to a generic language system, assuming they didn't build this on top of a GPT. The blog post doesn't say but I'm assuming (perhaps that's no longer accurate) that it's prohibitively expensive for a small team without huge funding to build such a model as a side project