OpenAI o1 AI Model Sometimes Thinks in Chinese and No One Knows Why

The reasoning AI model OpenAI o1 surprises with its linguistic abilities, switching to different languages ​​in the process of solving problems, including Chinese, even if the request was made in English. Experts are at a loss, but OpenAI is in no hurry to explain.

Image source: hdhAI

OpenAI last year introduced its first artificial intelligence (AI) model with reasoning capabilities, o1. However, users noticed a curious feature: the model sometimes begins to think in Chinese, Persian or other languages, even if the question was asked in English. For example, when solving the problem “How many letters R are in the word strawberry?” o1 may perform part of the reasoning in Chinese before producing the final answer in English.

Experts put forward several theories to explain this phenomenon. One of them is the hypothesis about the influence of the Chinese data on which the model was trained. Clement Delangue, CEO of Hugging Face, which develops tools for building applications using machine learning, and Ted Xiao, a researcher at Google DeepMind, note that many companies, including OpenAI, use the services of third-party Chinese companies to label the data needed to train models. This, in their opinion, may lead to the fact that the o1 model in some cases prefers the Chinese language in its reasoning, as it is more effective for processing information.

However, not everyone agrees with this theory. Some experts point out that the influence of Chinese data is not confirmed and the model may in fact switch to different languages, including Hindi and Thai. This may be due to the fact that o1 simply chooses the most convenient way to solve the problem. According to Matthew Guzdial, an AI researcher at the University of Alberta, the o1 model does not treat languages ​​as separate entities, but rather just plain text consisting of a set of tokens that it processes.

Ultimately, the exact reasons for the behavior of the model reasoning in different languages ​​remains a mystery. Luca Soldaini, a research fellow at the Allen Institute for AI, emphasizes that due to the “opacity of AI models, it is impossible to establish exactly what is happening inside” and how it all functions. OpenAI itself remains silent for now.