Researchers from MIT studied how AI models like ChatGPT work. AI models have been working in complex ways for chatting, customer support, coding and translation but many still don’t understand how AI models work. Upon researching, the researchers found out that many Large Language Models (LLMs) use linear functions to decode and recover facts and information. The AI models use the same functions to decode the facts that are similar.
The researchers of the study said that they can find where the information is stored in LLMs by decoding these equations for different facts. The researchers found that even when an AI model answers incorrectly, the information stored in it is still correct and the model just doesn’t know how to convey it. In the near future, scientists will probably find a way to train AI models in such a way that they will be able to give information as correctly as it is stored in it.
The co-author of the study, Evan Hernandez, says that even if these AI models who are trained on linear functions and their data are hard to understand, the mechanisms they work under are really simple and can easily be understood. Most of the AI models work as neural networks. This means that they are developed by keeping the structure and function of the human brain in mind. These AI models are also called transformer models and the information stored in these models is often related to each other. When additional information is added in a transformer model, it relates it to an already stored subject and decodes it when it is relevant to the prompt. They do this by finding the relation with the subject due to linear functions.
The researchers also used linear functions to decode what AI models believe is true for a specific subject. The AI model can choose to focus on different information when text is being produced, but it does encode all the information that is provided to it. This study shows that complex linear functions in an LLM can easily be replaced by simple linear functions.
Image: DIW-Aigen
Read next: New Vulnerability Targets Open Source AI Giant Ray Leaving Thousands Of Companies Exposed
The researchers of the study said that they can find where the information is stored in LLMs by decoding these equations for different facts. The researchers found that even when an AI model answers incorrectly, the information stored in it is still correct and the model just doesn’t know how to convey it. In the near future, scientists will probably find a way to train AI models in such a way that they will be able to give information as correctly as it is stored in it.
The co-author of the study, Evan Hernandez, says that even if these AI models who are trained on linear functions and their data are hard to understand, the mechanisms they work under are really simple and can easily be understood. Most of the AI models work as neural networks. This means that they are developed by keeping the structure and function of the human brain in mind. These AI models are also called transformer models and the information stored in these models is often related to each other. When additional information is added in a transformer model, it relates it to an already stored subject and decodes it when it is relevant to the prompt. They do this by finding the relation with the subject due to linear functions.
The researchers also used linear functions to decode what AI models believe is true for a specific subject. The AI model can choose to focus on different information when text is being produced, but it does encode all the information that is provided to it. This study shows that complex linear functions in an LLM can easily be replaced by simple linear functions.
Read next: New Vulnerability Targets Open Source AI Giant Ray Leaving Thousands Of Companies Exposed
