Google Translate Learns Inuktut

blog.google

42 points·joshdavham··7 comments

My immediate thought was, isn't that Inuktitut? But today I learned that Inuktut is as term which encompasses that and also Inuinnaqtun.

My second thought was surprise that they didn't already handle any language with an appreciable digitized corpus.

Funny enough, Claude 3.5 Sonnet seems to understand it perfectly, even though it's a language spoken by 39k people in the world.

Claude's multilingual capabilities never cease to amaze me. ChatGPT (GPT-4o) just outputs gibberish in Inuktut instead.

It's still surprising how Anthropic seems to be ahead of OpenAI in many ways. Claude models were able to speak Georgian at a native level when ChatGPT was still outputting gibberish (it has improved, but not to Claude's level). Claude can also speak and understand Mingrelian, not even Google Translate supports this language.

How likely is it that the text you're testing with was in the training set?

It may be that one LLM had text from the language in its training set along with translations, and the other didn't.

"Translate this into Inuktitut using syllabograms: "How likely is it that the text you're testing with was in the training set?"

"ᖃᓄᓪᓗᐊᖅ ᑖᓐᓇ ᐅᖃᓕᒫᖅᑕᐃᑦ ᖃᐅᔨᓴᕈᑎᒋᔭᐃᑦ ᐃᓕᓐᓂᐊᕈᑎᒋᓯᒪᔭᑦᑎᓐᓃᓐᓂᕐᒪᖔᑦ?"

The translation seems to make sense overall, it could be perfect or slightly off, only an Inuktitut speaker can verify that. Still, it's a better translation than what Google Translate offers.

Is it not pushed yet? I went to try it and it's not in the list for me.

now I can sound like an asshole in Inukut, neat

Reply on news.ycombinator.com