Stanford researchers create ChatGPT-like AI, spend only $600 for its training

March 22, 2023  10:25

Stanford University researchers have developed their own artificial intelligence (AI) model, the Alpaca AI, which similar to ChatGPT. This model cost only $600 to train because it was built based on an open source platform.

Their experiment has shown that today's advanced AI systems can be easily replicated, and it doesn't even require huge costs; all that is needed is a knowledge base on which the given AI will learn. The base cost the researchers $500, and they spent about $100 on creating the AI, which required eight computers with 80GB NVIDIA A100 accelerators for three hours; they "leased" this service in the cloud infrastructure.

Stanford researchers based their aforesaid AI on the open language model LLaMA 7B, which is the most widely available of the LlaMA series developed by Meta. Its capabilities are actually quite limited, and it does worse than the ChatGPT on most tasks. And this is not surprising, because training the OpenAI GPT models took more time and resources. ChatGPT has read billions of books, while Alpaca AI has learned some—albeit quite a few—questions and answers given by humans.

On the other hand, the Alpaca AI does some tasks quite well—and sometimes even better than the ChatGPT. In the first testing, Alpaca AI successfully completed 90 tests—writing emails, posting to social media, helping with work, etc.—, whereas ChatGPT successfully completed 89 tests.

However, it should be noted that the researchers have not yet addressed the most important issue for modern AI; they still need to make additional adjustments to ensure this model works safely and ethically.

According to the researchers, they could have achieved such a result with even less material costs if they had had such a goal from the beginning. Moreover, theoretically, anyone with corresponding technical training and at least $600 can repeat their experiment.

True, it is worth considering that the OpenAI's license does not allow the use of data from its models to develop competing systems. Meta allows academic researchers to use non-commercial licenses for now.

  • Archive