The researchers chose the Harry Potter books due to their universal familiarity, making it easier for others to evaluate the effectiveness of their technique. The model was posted on Huggingface, an open-source data science platform, for others to test. This development could be significant for an industry that relies heavily on data, as forgetting or unlearning could become as important as learning. However, as AI tools improve, they also become more capable of spreading well-written and persuasive misinformation, as noted by NewsGuard.
Key takeaways:
- Microsoft researchers Ronen Eldan and Mark Russinovich have developed a method to make an AI model forget things it had learned from private user data. They used the model Llama 2-7b and made it forget the Harry Potter universe in 30 minutes.
- The research aims to address the challenges of bias, untruths, and misinformation in the massive amounts of data used by large language models (LLMs). Traditional models focus on adding or reinforcing knowledge, but do not provide straightforward ways to “forget” or “unlearn” knowledge.
- The researchers chose the Harry Potter books for their universal familiarity to test their technique. The model was posted on Huggingface, an open-source data science and machine-learning platform, for others to test.
- The New York Times is suing OpenAI, the creator of ChatGPT, and Microsoft for copyright infringement, claiming that millions of its articles were used to train chatbots that compete with the NYT for a source of information.