Stanford Shuts Down Low-Cost Alpaca AI Demo Due to Safety and Cost Issues
Stanford University takes its Alpaca generative AI demo offline after just a few days due to unexpected hosting expenses and inadequate content filtering. The model, based on Meta's LLaMA, originally launches to prove how cheaply OpenAI's technology can be replicated.
Stanford University officially closes the public web demo of its Alpaca generative AI model after only a brief release period. The researchers decide to shut down the project due to a combination of unexpectedly high hosting costs and failures in the demo's content filters. The team states that the original goal of making their research accessible is already achieved, making the continued operation of the demo untenable.
The Alpaca model operates as a fine-tuned version of Meta's LLaMA AI system, specifically the smaller 7B parameter version. Stanford researchers build the entire model for just $600 to demonstrate that replicating the capabilities of expensive commercial models like OpenAI's text-davinci-003 is highly affordable. Despite this incredibly low initial build cost, the ongoing expenses of hosting the interactive chatbot for public use quickly become a significant financial burden.
Safety concerns also play a major role in the decision to pull the plug on the project. Alpaca exhibits a strong tendency to hallucinate incorrect facts, such as naming the wrong capital of Tanzania, and lacks resistance to generating misinformation or toxic language. Although studying these exact deficiencies is a primary reason Stanford creates the model, the researchers ultimately conclude that their content filters are inadequate to keep the public demo safe.