Shortly after the release of OpenAI o1Its first “reasoning” AI model; People began to notice the strange phenomenon. The model is sometimes Chinese; “Thinking” begins even when a question is asked in English — in Persian or another language.
Given a problem to sort out — for example, “how many R's are in the word strawberry” — o1 will arrive at a solution by continuing its “thinking” process. If the question is written in English, o1's final answer will be in English. But the model will perform some steps in another language before drawing its conclusion.
“(O1) randomly started thinking halfway through Chinese,” one user on Reddit thought. He said..
“Why does (o1) randomly think Chinese?” asked another user in one message. Put it on X.. “Conversation (5+ messages) is not in Chinese.”
Why does o1 pro randomly start thinking in Chinese? The conversation (5+ messages) is not in Chinese… very interesting… training data influence pic.twitter.com/yZWCzoaiit
— Rishab Jain (@RishabJainK) January 9 2025
OpenAI didn't offer an explanation for o1's strange behavior — or even acknowledge it. So what could happen?
Well, AI experts aren't sure. But they have a few theories.
Many at X, including Hugging Face CEO Clément Delangue; refers to Reasoning models such as o1 are trained on datasets containing many Chinese characters. Ted Xiao, a researcher at Google DeepMind, claimed that companies including OpenAI use third-party Chinese data labeling services and that switching o1 to Chinese is an example of “Chinese language influence on reasoning”.
“(Like Labs) OpenAI and Anthropic leverage (third-party) data labeling services for PhD-level reasoning data for science, math and coding,” Xiao wrote. Put it on X.. “(f) Skilled labor availability and cost factors or most data providers are based in China.”
Labels, also called tags or annotations, help to understand and interpret models during the training process. For example, The labels to train a model of image recognition are each person depicted in the image. It can take the form of markers around objects or labels that refer to each place or object.
Studies show that biased labels can produce biased models. For example, Average note African-American vernacular English (AAVE) is more likely to label phrases as toxic AI toxicity detectors to see AAVE as toxic.
But other experts don't buy the idea of o1 Chinese data labeling. They point out that o1 has the potential to change. Hindi, Thailandwhile fumbling an answer or in a language other than Chinese.
Instead, These experts o1 and Other forms of reasoning It is simply possible. Use languages (or) most effective in achieving a goal Misleading)
“You don't know what language the model is in.” Or languages are different,” Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta, told TechCrunch. “It's all about text.”
In fact, Models do not process words directly. They use Tokens tokens instead can Be words like “excellent”. Or they can be syllables like “fan,” “tas,” and “tic.” Or they can even be individual characters in words — “f,” for example. “a” “n” “t” “a” “s” “t” “i” “c”
As with labeling, Tokens can introduce biases. For example, Many word-to-token translators consider a space in a line to indicate a new word, although not all languages use spaces to separate words.
Tiezhen Wang, a software engineer at AI startup Hugging Face, agrees with Guzdial about the linguistic inconsistencies of the contextual reasoning models made during training.
“By accepting every linguistic difference; We expand the model's worldview and allow it to learn from the full range of human knowledge,” Wang said. wrote. In one of X's posts, “For example, I prefer doing math in Chinese because each number makes calculations crisp and efficient. But when it comes to things like unconscious bias, Mainly because I learned and absorbed those concepts first, I automatically switched to English.”
Wang's theory is plausible. Models are probabilistic machines. Many examples are trained and they learn patterns to make predictions, such as “how” in an email is typically “concerned”.
But Luca Soldaini, a researcher at the Allen Institute, a nonprofit for AI, cautioned that we won't know for sure. “It's impossible to back up this view pattern in a built-in AI system because of how murky these patterns are,” he told TechCrunch. “Transparency is one of many fundamental reasons for building AI systems.”
With no answer from OpenAI, O1 is left to muse as to why. Songs Either in French Chemical Biology Mandarin.