Why Is AI Thinking In Chinese #funfact

When OpenAi released o1 people began noticing a bizarre phenomenon, the AI was thinking in Chinese and multiple other languages

[Random_Person]

"the good news is the AI is doing better than we expected"
"what about the bad news?"
"the bad news is the AI is doing better than we expected"

[GuiGamer]

This is just like the Hollow Knight speedrunning community running the game in Chinese because it's got the fastest dialog boxes

[SytanOfficial]

Machine learning engineer here. I work with LLMS and companies that develop them. We do know why they think in Chinese, and it's because the models are rewarded for having the most efficient chain of thoughts balanced with the highest accuracy

English is a fast-spoken language but it is not informationally dense. Symbol-based languages such as Chinese are able to encode much more minute information in far less tokens per response. Where English may need an entire sentence to describe a specific feeling or sensation, Chinese can use a singular token to express the same thing that with even more granularity

Additionally, it's not all AI models that are doing this. For example, R1 from deep sea has a very low propensity of doing this, simply because the model is rewarded for the most correct answer, rather than the most correct and shortest answer. Optimizing for the shortest answer possible will always lead towards the models preferring to use more informationally dense languages and chains of thought, while optimizing for quality and not caring about the length will result in a generally stronger adherence to the target language

[athanmantas]

My idea was that you can get across more information per character in Chinese. Chinese has a high information density. It’s easer to get the same information with less characters

[yektaagra]

I dont want "AI" and "nobody knows why" in the same sentence

[alik.]

A Linguist here. So the different languages have different structures as you know. Therefore, when you think something in English for example, it will be slighty different from Chinese or other languages. Some languages are great for understanding mathematics (like chinese), some languages are great for understanding literature and poetry (like farsi), and some languages are focused on actions and momentum (like Turkish). This is because that not all the words in a language are 100% translatable. So AI just thinks in whatever language is the best for that current subject, then moves on the other language to continue its thinking for another subject matter.

[casualangler]

"... To give the best most accurate answer."

Oh, that's cool.

"But no one trained it to do that."

Oh...

[william]

I've actually used o1 and o3 and I've tried asking them questions in different languages and oftentimes it'll think in English and then answer in the original language it's very interesting.

[bloodfiredrake]

As a trilingual I do this too. I think in multiple languages to get the best estimation of what I wanna say. I just can't really formulate some thoughts in some languages.

[SatoshiAR]

Ironic when you consider the "Chinese room" argument.

[fumiko_]

My uneducated guess is that because LLMs are trained on tokens directly, they don't have the concept of languages (we don't tag each token with what language it comes from). So asking it a problem in English may result in it generating text from another language, but it's still a relevant response to your original message. There's absolutely more variables at play here like including different languages in the hidden system prompt (if any), but yeah this is my guess

[FakecelFakefanSchiZoid]

"This how is how an AI would think if it were Black or Chinese"

[PeekasLukkala]

"Oh cool"

"But no one trained it to do that"

"Oh no"

[sanadalusta]

“ no one trained it to do that “ yeah this shits getting scary.

[napoleon]

US senator: Open AI, are you a Chines spy? have you joined Chinese Communism party? When you eat breakfast this morning, you are using chopsticks or a fork?
Open AI: I am an AI, sir.

[travelsizedlions]

A.K.A, "AI rediscovers extremely common speed running strategy to process more information into fewer bytes"

[four]

fun fact: this has been happening since before 2023. this was discussed in the AI Dilemma here on youtube, and it's been learning without being taught for ages.

[supermj]

I once heard a fact that "Thinking in other languages makes you make more rational decisions." Maybe the AI saw this as fact and ran with it.

[Mlamin]

Project Skynet is proceeding very well i see💀

[killerbunny]

All the literature on physics is in English, I'm German. I have been trained on data from different languages. So I translate questions into multiple languages to check the questions against the data I have been trained with. Then I translate the answer into the appropriate language.

Why would it only think in English?