There are hundreds of languages inside any language, but no apps to translate between them


GPT-3 is amazing at spotting a pattern given only a few examples, in some cases zero examples. Few-shots and zero-shots. It can even translate between languages, presumably because it has read a few language courses while making its way across all public texts ever published. But no one really knows for sure, not even its creators.

There's a lot of patterns within a language. Sub-languages. Mistakes. Slang. In the end, we are 6 billion very similar human brains on this planet.

One such pattern I find fascinating is the pattern found in dyslexic brains. It's probably as simple as just writing out the word sounds rather than the word images. That would make sense if only dyslexia wasn't the inability to couple words and sounds. So some kind of pattern arises, even though it was difficult to find enough datapoints of dyslexic writing. If you know where to find more, let's chat.

Another pattern is dialects. Some dialects are famously hard to understand (when people from certain parts of Denmark are on national tv, they add subtitles). Here's an example of a heavy dialect from the Northern part of Jutland in Denmark. GPT-3 has no problem translating.