What do you mean, what do you mean? Artificial intelligence word soup

Written by Ben Esplin

LLMs, and other recent generative AI technologies, are so powerful they almost seem like magic. I believe one of the challenges for entrepreneurs with projects built on generative AI is a thorough lack of understanding of the underlying technology on the part of principles within the technology ecosystem. By that I mean very few people, even within “tech,” understand the “nuts and bolts” of a system created to leverage generative AI. This results in a vernacular around generative AI that is fragmented and inconsistent. This inconsistency in language is a roadblock for generative AI projects. Of course, my particular frustration is in trying to describe technical innovation in the space for the purpose of patent protection, but it also creates difficulties in explaining a project to investors, potential clients, potential partners, and in other circumstances.

For example, many use the term “AI” when they mean machine-learning. Technically, “artificial intelligence” is a term broad enough to include a state machine, one of the earliest forms of machine-based reasoning that can be embodied in a simple analog electrical circuit.

As another example, I would characterize most early projects leveraging GPT or other proprietary, third-party models as “agents” that used prompt engineering to perform specific tasks in a certain way, sometimes with a particular style. However, I still hear efforts like this referred to as “wrappers.” But “wrappers” are actually the tools implemented by these projects to provide a higher-level interface with an LLM that is useful in getting the LLM to provide the features and functions desired of the agent. Calling these projects “wrappers” would be like calling a project that utilizes a complex database (e.g., a modern videogame) a “query language.”

Often, terms like “an AI,” “a model,” and “an agent” are used interchangeably (at least between different speakers), even though under rigorous scrutiny they actually refer to concepts which are semantically very distinct. The problem extends to other areas of innovation that can be deployed with generative AI models, like embeddings datastores that often make up the backbone of Retrieval Augmented Generation (“RAG”) systems which extend and enhance the capabilities of LLM functionality.

What is to be done? We must be rigorous in our language, and expect the same from others.