An AI system that analyzes enormous quantities of natural language (millions and billions of sentences). The large language model (LLM) uses the word patterns in this input to create an "inference engine" that generates original content based on prompts from the user. The LLM is "trained" on the language data, and the training phase, which takes the most computer time, is the first step. See
supervised learning.
Large language models (LLMs) differ from AI robots designed to handle mechanical tasks. While robots may employ language models to interact with the user in natural language, they employ machine vision and other technologies to determine how to perform their mechanical routine.
The World's Information Is Online
Since the mid-1990s, the bulk of the world's information has become available over the Internet from websites, blogs and social media. A large language model uses huge amounts of this data as input to a "neural network" to learn about a subject. For example, OpenAI's GPT-3 uses more than 150 billion data items. See
neural network and
AI emergent properties.
Smaller Models: Fine-Tuned and Edge
Designed for a very specific purpose such as writing source code, fine-tuned language models are smaller than an LLM. Edge language models are smaller yet and generally do not require processing in the cloud. See
GPT,
ChatGPT,
LLaMA,
LaMDA,
PaLM and
Gemini chatbot.