Google’s AI Translation Tool Has Invented Its Own Secret Internal Language


Yes, read it again- Google’s artificial intelligence (AI) program has invented its own language. Computers are only as smart as humans make them because humans program them; this means that humans write ‘the rules of the game’, or rather how smart computers are and how smart they can become.  This article is one of the few that has provided us a glimpse into the future where software programs begin ‘thinking for themselves’ and expand beyond how smart humans programmed them to be.

The AI program is called ‘Google Neural Machine Translation’ (GNMT).  It was launched in November, 2016, to assist in the Google Translator application that supports more than 100 languages.  The idea was simple: Program the GNMT to learn from the requests that Google Translator receives…so the system learns to create better, more natural translations.

GNMT was less than a month old when Google’s AI researchers realized that the program created its own language (which it was not programmed to do).  GNMT created short-cuts while translating; it is these short-cuts that are not a language as we conceive it, but rather an unknown code, where the new language exists.  The following is how it works:

Let’s say that GNMT is programmed to translate English to Korean, and English to Japanese, but it is not programmed to translate Japanese to Korean.  This would result in English being the ‘base’ language, so Japanese would have to be translated to English before it could be translated to Korean.  English is the gateway language.

Read the source article at Home | NAU