OpenAI recently unveiled its pioneering “reasoning” AI model, merrily christened o1. However, this technological marvel soon raised eyebrows with its idiosyncratic ‘thinking’ process, occasionally sparking off in Chinese, Persian, or other languages, all the while responding to English prompts.
Users started to notice such behavior; o1, when confronted with an analytical problem such as identifying the amount of ‘R’s in the word ‘strawberry’, would often undergo a multilingual thought process before delivering its English conclusion.
While OpenAI remains mum on this peculiar behavior, speculation among the AI experts is rife. Some theorize that the behavior might be a downstream impact of the Chinese data labeling services employed by OpenAI and others. Their claims are rooted in the extensive use of Chinese characters in reasoning AI data sets. Notwithstanding, there is skepticism about this theory, as o1 also switches to languages like Hindi or Thai during its thought process.
Disputing this theory, AI researcher Matthew Guzdial suggests that AI models make no language distinctions, they merely perceive it as text. These systems operate using tokens that can represent a word, a syllable, or even a single character. Indeed, these tokens could possibly propagate bias, as many translators operate on the assumption that a space in a sentence represents a new word – a rule not applicable to all languages.
Another potential explanation by Tiezhen Wang from Hugging Face is that language inconsistencies can be attributed to associations made during training. AI systems might resort to a language that, according to their training, expands their data processing abilities.
However, all these theories lack definitive proof due to the inscrutable nature of AI models. Until an explanation emerges, one can only wonder why o1 prefers thinking about pop tunes in French and synthetic biology in Mandarin.
Original source: Read the full article on TechCrunch
