Stanford Removes Low-Cost Alpaca AI Model Over Safety and Hosting Costs
Stanford University takes its Alpaca AI web demo offline due to concerns about the model generating toxic content and the rising expenses of hosting it. The low-cost language model was built by fine-tuning Meta's LLaMA system.
Stanford University takes down the web demo for Alpaca, a small AI language model built on Meta's LLaMA system, due to safety concerns and increasing hosting costs. Researchers develop the seven-billion-parameter model for less than $600 to help academics study why language models generate false information and toxic text.
The Stanford team fine-tunes LLaMA using 50,000 text samples so it functions similarly to OpenAI's text-davinci-003. The project quickly gains attention from developers who successfully run the open-source model on modest hardware like Raspberry Pi computers and Pixel 6 smartphones.
Despite the goal of making research accessible, the developers remove the public demo after discovering the model still hallucinates, produces misinformation, and propagates social stereotypes. The team states it achieves its primary objective of disseminating the research, even though the interactive webpage remains permanently closed.