QuickTake
How AI-Driven Chatbots Know What to Say
Trained on vast amounts of information, large language models identify language patterns, then use them to convincingly mimic the way people communicate.
Large language models, or LLMs, are the building blocks of chatbots, such as ChatGPT and Google’s Bard, that are driven by generative artificial intelligence. The mechanics of LLMs explain how these platforms are able to impersonate Shakespeare, elegantly explain complex topics and even write computer code. They also demonstrate how convincing responses produced by generative AI can sometimes be biased, or simply wrong.
An LLM’s basic architecture includes a kind of multidimensional map of meanings that reflects how words are similar or dissimilar and an analytical component known as a neural network. An LLM puts a user’s question or command through layers of decision-making nodes in the neural network to come up with a response. Here’s a simplified depiction of the process.
Training
Before an LLM is able to handle user prompts, it goes through a training process in which it is fed vast amounts of information (for example, every book available freely on the internet, news articles, online forum content and Wikipedia pages). It observes language patterns in the material and tweaks the parameters, or weights, of the nodes in the neural network in response.
Responding to Prompts
Once an LLM is sufficiently trained, it can provide responses that convincingly mimic natural language.
An LLM has not absorbed facts. It’s learned what a plausible answer would sound like. That’s why some answers created by generative AI sound convincing but are untrue. Developers call these responses hallucinations.