If you speak multiple languages, you’ve probably had the experience of thinking or dreaming in a language that’s not your native tongue.
I learned some French in high school, and I’ll occasionally have a dream where I’m speaking French.
Apparently, OpenAI’s o1 model does the same thing. According to recent reporting by TechCrunch, the model will sometimes randomly begin to “think” in Chinese.
To understand why, it helps to know a bit more about how the o1 model works.
Mulling it Over
OpenAI’s O1 model is a reasoning model. Whereas earlier generations of large language models (LLMs) and chatbots took in a prompt and immediately returned a response, reasoning models work differently.
When you give them a prompt, there’s an intermediate step where the model first expands on your prompt, works to understand your intentions, and then maps out its plan for solving the problem you’ve presented.
For example, if you ask the model to write a blog post, it might first consider the general format that blog posts follow, review other blog posts you may have asked it to create in the past, and take into account…
It will use all of this information before formulating its response.
Likewise, if you ask a scientific question, the model might break the question into smaller parts in order to solve it more effectively.
These intermediate reasoning steps add a huge amount of time, cost, and need for computing power to reasoning models. Whereas older LLMs would respond in a few seconds, o1 can take up to two minutes to solve a complex problem.
To most users, the intermediate reasoning step is relatively hidden. You put in your question, wait a while, and get a solid response back.
But if you pull back the curtain and look at that intermediate step of reasoning and planning, you can get a sense of how these models approach the problems they’re trying to solve.
And in the case of o1, apparently, you’ll see that the model will sometimes switch into Chinese during this step.
Crucially, this can happen even when the model is given an English input and is returning an English output. It will first translate the user’s prompt into Chinese, “think” about it in Chinese for a while, and then translate the response back into English.
Ghosts in the Machine
Why is o1 doing this? Researchers believe it has a lot to do with how the model was trained. o1 was likely trained on essentially all of the publicly available text on the internet and in published books and papers.
Much of that text is in English, but a large amount is in Chinese. Because the model has seen so much Chinese text, it may have discovered that “thinking” in Chinese gives it access to more information and thus allows it to generate better responses.
It’s also possible that something about the Chinese language lends itself to particular kinds of computational thought. The complex and varied symbols in the Chinese language might allow for more abstraction in the model’s thinking processes, or it might enable the model to be more flexible in the kinds of responses it generates.
Researchers have also observed that o1 tends to switch into Chinese when it’s solving complex, scientific problems. Again, that might represent the fact that it’s seen a lot of Chinese characters in the scientific papers in its training set.
Many Chinese research papers are published online for free — at least 60 million on a single, popular website — so the model may have a lot of examples of Chinese scientific data to pull from.
Either way, it’s a fascinating insight and another example of a situation where LLMs behave wildly differently than their creators expect.
When OpenAI first released ChatGPT, researchers were shocked to find that the model had taught itself how to code in Python.
No one explicitly asked it to do this — it simply figured out how the Python programming language works from looking at millions of pages of Python code published online.
These unique abilities of large language models are called emergent properties. Again, they’re not explicitly taught to the model — they simply emerge from the training data and the way the models are constructed.
The choice to think in Chinese is likely another example of an emergent property.
My favorite insight from the study, though, hit closer to home. Apparently, when the o1 model is thinking about literary or poetic topics, it often switches to French rather than Chinese.
There’s something beautiful about the idea that the French language is as uniquely suited to poetry and artistic thought as Chinese is to mathematical and scientific topics.
o1, in short, sometimes muses and dreams in French — just like me.
I’ve tested thousands of ChatGPT prompts over the last year. As a full-time creator, there are a handful I come back to every day. I compiled them into a free guide, 7 Enormously Useful ChatGPT Prompts For Creators. Grab a copy today!
