Serving technology enthusiasts for more than 25 years.
TechSpot is the place to go for tech advice and analysis.
WTF?! OpenAI’s latest AI model o1 has displayed unexpected behavior which has captured the attention of users and experts. The model, designed for reasoning tasks, has been observed to switch languages mid-thought even when the original query is presented in English. Users from various platforms have reported
OpenAI’s O1 model has been observed in instances where it begins its reasoning in English, but unexpectedly switches to Chinese, Persian or other languages, before delivering the answer in English. This behavior can be observed in a variety of scenarios, ranging from simple counting exercises to complex problem solving exercises.
A Reddit user commented “It randomly started thinking in Chinese halfway through,” and another user on X questioned “Why did it randomly start thinking in Chinese? No part of the conversation (5+ messages) was in Chinese.”
.
Why did o1 pro suddenly start thinking in Chinese? No part of the conversation (5+ message) was in Chinese… very intriguing… training data influence pic.twitter.com/yZWCzoaiit
– Rishab Jain (@RishabJainK)””https://twitter.com/RishabJainK/status/1877157192727466330?ref_src=twsrc%5Etfw””> January 9, 2025
The AI Community has been buzzing about theories to explain this strange behavior. OpenAI has not yet released an official statement. However, experts have proposed several hypotheses.
Some experts, including Hugging face CEO Clement Delangue speculate that the phenomenon may be linked to the data used in training o1. Ted Xiao is a Google DeepMind researcher who suggested that relying on third-party Chinese data labels for expert-level reasoning could be a factor.
“For expert labor availability and cost reasons, many of these data providers are based in China,” Xiao said. This theory suggests that the Chinese language influence on reasoning may be due to the labeling process used by the model during training.
What about the impact of closed-source players using open-source AI, which is currently dominated by Chinese players, like open-source datasets.
Open-source AI winners will have a huge impact on the future of AI. https://t.co/M8ZdYfWxNI
– clem (@ClementDelangue) January 10, 2025
According to another school of thought, o1 may be selecting the languages that it believes are most efficient in solving specific problems. Matthew Guzdial is an AI researcher at the University of Alberta and an assistant professor. He offered a different view in an interview with TechCrunch. This view implies that a model’s language switching may be a result of its internal processing mechanics, rather than based on linguistic comprehension.
A new phenomenon has appeared: the latest generation foundation models often switch from Chinese to CoT thinking in the middle.
Why? AGI labs such as OpenAI and Anthropic use 3P data labeling for PhD-level reasoning for science, math and coding. https://t.co/VllUIC9V91
– Ted Xiao (@xiao_ted) January 9, 2025
Tiezhen, a software developer at Hugging Face suggests that the inconsistencies in language could be due to associations the model made during training. “I prefer doing math in Chinese because each digit is just one syllable, which makes calculations crisp and efficient. But when it comes to topics like unconscious bias, I automatically switch to English, mainly because that’s where I first learned and absorbed those ideas,” Wang explained.
For example, I prefer doing math in Chinese because each digit is just one syllable. https://t.co/yD2YNscWW5
– Tiezhen WANG (@Xianbao_QIAN) January 13, 2025
While these theories offer intriguing insights into the possible causes of o1’s behavior, Luca Soldaini, a research scientist at the Allen Institute for AI emphasizes that For example, I prefer doing math in Chinese because each digit is just one syllable, which… https://t.co/yD2YNscWW5
– Tiezhen WANG (@Xianbao_QIAN) January 13, 2025
While these theories offer intriguing insights into the possible causes of o1’s behavior, Luca Soldaini, a research scientist at the Allen Institute for AI, emphasizes the importance of transparency in AI development.
“This type of observation on a deployed AI system is impossible to back up due to how opaque these models are. It’s one of the many cases for why transparency in how AI systems are built is fundamental,” Soldaini said.