Llama? Vicuña? Alpaca? You might be asking yourself, “what do these camelids have to do with licensing LLM artificial intelligence?” The answer is, “a lot.”
LLaMa, Vicuña, and Alpaca are the names of three recently developed large language models (LLMs). LLMs are a type of artificial intelligence (AI) that uses deep learning techniques and large data sets to understand, summarize, generate, and predict content (e.g., text). These and other LLMs are the brains behind the generative chatbots showing up in our daily lives, grabbing headlines, and sparking debate about generative artificial intelligence. The LLaMa model was developed by Meta (the parent company of Facebook). Vicuña is the result of a collaboration between UC Berkeley, Stanford University, UC San Diego, and Carnegie Mellon University. And Alpaca was developed by a team at Stanford. LLaMa was released in February, 2023; Alpaca was released on March 13, 2023; and Vicuña was released two weeks later on March 30, 2023.
LLMs like these are powerful tools and present attractive opportunities for businesses and researchers alike. Potential applications of LLMs are virtually limitless, but typical examples are customer service interfaces, content generation (both literary and visual), content editing, and text summarization.
While powerful, these tools present risks. Different models have diverse technical strengths and weaknesses. For example, the team that developed Vicuña recognizes “it is not good at tasks involving reasoning or mathematics, and it may have limitations in accurately identifying itself or ensuring the factual accuracy of its outputs.” Thus, Vicuña might not be the best choice for a virtual math tutor. Moreover, in a general sense, the most popular type of LLM – the recurrent neural network (RNN) – is well-suited for modeling sequential data, but suffers from something called the “vanishing gradient problem” (i.e., as more layers using certain activation functions are added to neural networks, the gradients of the loss function approach zero, making the network hard to train). Meanwhile, transformers (the “T” in GPT), are great with long-range dependencies which help with translation style tasks, but are limited in their ability to perform complex compositional reasoning.
The waters are muddied further when these large corporations start lending and sharing availability of LLMs with each other. There are further indications that Meta is opening up access to its LLaMa model beyond the world of academia as reports surface about partnerships with Amazon and Microsoft. For example, Meta’s LLaMa large language model is now available to Microsoft Azure users.
Thus, in selecting LLMs for various purposes, users must weigh the technical advantages and drawbacks of the different models (e.g., network architecture, weights and biases of algorithms, performance parameters, computing budget and the actual data on which the model was trained) with the legal liabilities that may arise from using these LLMs. Critically, before investing too much time or resources into a product or service that makes use of an LLM, business leaders must review the terms associated with the model in order to fully understand the scope of legally permissible use and take actions to ensure legal compliance with those terms so as to avoid liabilities.