According to news on December 12, xAI, an artificial intelligence startup owned by Tesla and SpaceX CEO Elon Musk, recently released its chat robot Grok, but something seems to have gone wrong. On Friday, security tester Jax Winterbourne posted a screenshot on social media platform
This aroused people's curiosity because Grok was not developed by OpenAI, and the two parties are still competitors.
What’s even more interesting is that the folks at xAI didn’t deny that this behavior occurred in their AI models. Igor Babuschkin wrote in his reply: "The problem is that when we were training Grok, the network was full of ChatGPT outputs, and we accidentally used some of them. We were also very surprised when we first noticed this problem. Anyway, this problem is very rare, and now that we are aware of it, we will make sure that future versions of Grok do not have this problem. Don't worry, no OpenAI code was used to develop Grok."
In reply to Babushkin, Winterbourne wrote: "Thank you for your reply. I will say that this situation is not uncommon and happens frequently when it comes to code creation. Still, I will leave it to those who specialize in large language models and artificial intelligence to discuss this further. After all, I am just a layman."
To many experts, however, Babushkin's explanation seemed unconvincing. Large language models typically don't spit out their training data verbatim, and it would be understandable if Grok happened to mention OpenAI's policies online. Refusing to respond based on OpenAI policy may require specialized training. The most likely reason for this happening is that Grok fine-tuned the output data of the OpenAI large language model.
AI researcher Simon Willison said in an interview: "I'm a little skeptical about this. Is Grok behaving like this just because the Internet is flooded with ChatGPT-generated content? I've seen a lot of open-weight models on HuggingFace exhibiting the same behavior, like They are also the same as ChatGPT. This is because these models are fine-tuned on datasets generated using OpenAIAPI, or using datasets generated by ChatGPT itself. I think it is more likely that Grok is tuned on the dataset containing ChatGPT output, rather than an accident based on network data.”
As OpenAI's large language models become more and more powerful, it has become increasingly common for some artificial intelligence projects (especially open source projects) to use synthetic data generated by other large language models to fine-tune the output of their own artificial intelligence models. Fine-tuning the behavior of an AI model can serve specific purposes, such as improving coding capabilities after an initial training run. In March this year, a group of researchers at Stanford University caused a stir when they used the output of OpenAI's GPT-3 model (named Text-DaVinci-003) to fine-tune instruction tracking.
Online, we can easily find several open source datasets that researchers have collected from ChatGPT output. So xAI may have fine-tuned Grok using one of the datasets to achieve some specific goals. This practice is so common that there is even an article titled "How to use ChatGPT to create a data set" on the encyclopedia-style life guide website WikiHow.
This is one way in the future to use artificial intelligence tools to build more complex artificial intelligence tools, just as people started using microcomputers to design microprocessors that were more complex than paper and pencil drawings. However, xAI may be able to avoid this in the future by filtering the training data more carefully.
While borrowing the work of others may be common in the machine learning community (although this is usually against the terms of service), this incident has intensified the competition between OpenAI and xAI, which even dates back to Musk's past criticism of OpenAI. As news spread that Grok might borrow OpenAI model data, the official ChatGPT account even posted a post mocking xAI, saying, “We have a lot in common” and quoting Winterbourne’s post. Musk fired back: "You collect all the training data from this platform, you should know that!"