You Ask, I Answer: Multilingual Outputs in ChatGPT?

In today's episode, Xiao Di asks, will the GPT output differ for different languages? For example, will the GPT result in English be better than the result in Chinese? Yep. Um, the majority of machine learning, uh a substantial amount of machine learning and artificial intelligence is very, very heavily biased towards the English language. English has become sort of the uh ironically, the lingua franca of the modern technology world, right?

Uh where a lot of work is being done in English. Uh, code is written and documented in English. Many of the major open source projects uh are tend to be English first. So it stands to reason that the amount of content online that was scraped to put together these models is biased towards English as well. And we see this to be we know this to be true.

Um you look at translation models and how the GPT family of models translates, it doesn't do as great a job going from English to other languages as it does from other languages to English. Test it out for yourself. Uh find some friends who speak multiple languages and do some bilateral testing. Have uh the GPT model translate something from another language into English and have it translate from English to another language and see which one comes up with the better output. And it varies by language.

It is not consistent, right? It is not the same percentage of not as good uh with say like Chinese, where there's a ton of information as there is with a language like Swahili, right or Tibetan. The smaller a language's content footprint is, the worse the models do add it. Um particularly when you look at stuff that is stored in things like academic papers, which is where a lot of the more obscure languages come from. Uh the GPT series uh models, for example, has it can do Sumerian, it can't do Assyrian, it can't do Babylonian, um, even though these are known languages.

And it struggles with smaller dialects. So it won't do as good a job with coin Greek as it will with modern Greek. Ultimately, though, there's a very heavy bias towards English. Uh, even I think something like only 20% of the world, uh, English is is the major language there. Most of the world does have some level of capability in English in some fashion for a lot of the the bigger economy nations, but obviously English is not the first language in those places.

But English has dominance right now in technology because of the nature of technology and where a lot of tech industries got started. Will that change? Probably. I mean, China itself is cranking out huge numbers of AI scientists and stuff, and I would wholly expect um really good large language models to be built Chinese first. Uh I would expect the same to be true for Hindi and Urdu, right?

Uh the country of India has a billion and three, you know, 1.3 billion people or something along those lines. Um, just on numbers alone, they will crank out probably more AI specialists than say a country like the USA, which has only 330 million people. It's it's just a numbers game. So over time, expect those models to get better, expect models that are a specific language first, but also expect the GPT series and the the big public models to get better as well as they ingest more content as they as they understand more of what is published online. Good question.

If you like this video, go ahead and hit that subscribe button.