Understanding the Impact of Parameter Changes in AI Models

Recently, Meta released its new AI model Llama 3.1. As an open-source AI, researchers can easily experiment with it. A group of researchers led by Colorado Reed, an Applied Research Scientist at Apple, conducted an experiment to see how easily AI models can be disrupted. According to New Scientist, even models with billions of parameters can be affected by changing just one number.

The researchers used the Llama-7B model, which has seven billion parameters. These numbers are responsible for generating text based on prompts. Each parameter represents probabilities for specific text elements that the AI can string together. The researchers altered these numbers to see how the output changed. They found that many parameters could be removed without significantly affecting the output. However, some parameters, called superparameters, are crucial. Changing one of these can cause the AI model to produce gibberish or hallucinations.

In their tests, the researchers altered about 7,000 parameters without affecting the AI’s output. Only the superparameter caused Llama-7B to malfunction. According to Yingzhen Li from Imperial College in London, these parameters are often at the beginning of the AI model. If they are incorrect, the error propagates through all subsequent parameters.

Thomas Wolf, co-founder of Hugging Face, confirmed these findings. In 2017, there was a large language model by OpenAI trained on 82 million Amazon reviews. The AI was supposed to distinguish between negative and positive reviews. According to Wolf, this decision depended on just one parameter in the AI model.

This insight could be valuable for future AI development. AI researchers need to find ways for models to function without relying on these crucial parameters and find alternative paths to correct answers.

The ability to alter AI models locally is also becoming more accessible. With the right tools, running an AI locally is not a problem. This advancement in AI technology opens up new possibilities for software and development. As AI continues to evolve, staying informed about these changes is essential.

Exit mobile version