Large AI networks like language models make mistakes or contain outdated information. MEND shows how to update LLMs without changing the whole network.
Large AI models have become standard in many AI applications, such as natural language processing, image analysis, and image generation. The models, such as OpenAI’s GPT-3, often have more diverse capabilities than small, specialized models and can be further improved via finetuning.
However, even the largest AI models regularly make mistakes and additionally contain outdated information. GPT-3’s most recent data is from 2019 – when Theresa May was still prime minister.
Comments are closed.