OpenAI's recently released reasoning AI model called “o1” exhibited the strange behavior of sometimes “thinking” in Chinese when solving problems, even when asked in English.
tech crunch report The AI community has been abuzz after noticing a strange phenomenon in OpenAI's first “inference” AI model, o1. The recently released o1 is designed to arrive at a solution to a problem through a series of reasoning steps. However, users have observed that the model sometimes switches to perform some of these inference steps in Chinese, Farsi, or other languages before providing the final answer in English. Masu.
This strange behavior has left many scratching their heads, as the language switching seems to happen randomly, even when the entire conversation with o1 is in English. “Why did it happen? [o1] Do you randomly start thinking in Chinese? one user mused on social media platform X, formerly known as Twitter. “The conversation (5 or more messages) did not include Chinese.”
OpenAI has been silent on the issue and has not provided any explanation or acknowledgment of the o1 language discrepancy. Despite the lack of official statements, AI experts have put forward several theories to explain this strange behavior.
One hypothesis, supported by Hugging Face CEO Clément Delangue and Google DeepMind researcher Ted Xiao, is that the datasets used to train inference models like o1 contain large numbers of Chinese characters. suggests. Xiao claimed that companies such as OpenAI and Anthropic use third-party data labeling services based in China for expert-level inference data related to science, mathematics, and coding. . He believes o1's tendency to switch to Chinese is a result of the “Chinese linguistic influence on inferences” arising from these data providers.
However, not all experts are convinced by this theory. They claim that o1 does not prefer Chinese only, but is just as likely to switch to other languages such as Hindi or Thai when seeking a solution. Instead, o1 and other reasoning models may simply be using the language they deem most efficient to achieve their goals, or they may be hallucinating language switching. No, they suggest.
Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta, told TechCrunch. “The model doesn't know what a language is or that it's different. It's all just text.” He said the model doesn't process words directly, but instead processes words, syllables, and even individual letters. Explains the use of tokens that can represent These tokens can introduce biases, such as assuming that a space in a sentence represents a new word, even though not all languages use spaces to separate words.
Tiezhen Wang, a software engineer at Hugging Face, echoed Guzdial's sentiments, suggesting that the inference model's language discrepancy could be explained by associations created during training. “By embracing all linguistic nuances, we extend the model's worldview and allow it to learn from the full range of human knowledge,” Wang writes about X.
Luca Soldaini, a researcher at the nonprofit Allen Institute for AI, cautions that the opacity of these models makes it impossible to know exactly what is causing this behavior. “This is one of many examples of why transparency in how AI systems are built is important,” they told TechCrunch.
read more Click here for TechCrunch.
Lucas Nolan is a reporter for Breitbart News covering free speech and online censorship issues.
