Large Language Models

Background

Providing the foundation for tools like OpenAI's ChatGPT and Google's Gemini, large language models are sophisticated computer programs that process and generate natural language.

By finding patterns in the sequence of words and subwords (prefixes, for example) in massive amounts of text, these models can predict the most likely next word in a generated sequence and repeat this prediction until the output is complete. Predictions improve with more data, and more specific outputs can be created using specialized data.

AI chatbots are user-friendly interfaces that act as intermediaries between users and the LLM framework.

How Are LLMs Created?

Engineers follow three key steps to build and improve an LLM: data, architecture, and training.

Data can be any digital resource containing text, including online textbooks, website articles, and podcast transcripts. As of 2024, the Internet Archive estimates there are 44 million digital books and texts and 835 billion websites on the internet, and hundreds of trillions of words from these sources can be in an LLM’s data set.

The software that processes this data is called transformer architecture, which Google researchers developed in 2017 to improve language translation.

First, transformers assign each word or subword in a block of text a set of numbers called a token. All tokens are then reviewed in parallel, identifying which go together most often to “learn” patterns in meaning, grammar, and context. These patterns are coded as parameters—numeric values that can be changed to improve output quality.

After the model sets parameters from data analysis alone, humans train the LLM, ranking the quality of its outputs and identifying boundaries for what should be excluded from outputs (e.g., violence). The model then adjusts its parameters, changing the associations between specific tokens to mimic natural language accurately. See how ChatGPT is trained here.

How Are LLMs Used?

Anyone using generative AI tools to produce text-based content interacts with an LLM, usually through a chatbot or other user interface.

When prompting, the chatbot sends the input to the LLM, which sends back unique responses by combining user instructions (e.g., “Using an engaging tone …”) and a randomization parameter called temperature, which sometimes tells the model to choose less probable next tokens.

LLMs have also been fine-tuned, where an existing model’s parameters are adjusted based on new data to accomplish a specific task more effectively. For instance, a business can fine-tune an LLM with product-specific troubleshooting knowledge so that its customer service chatbot can provide instructions on fixing a hardware problem in a friendly tone.

Over time, the LLM framework will be more commonly used in agentic AI, where a system takes actions to fulfill a goal given by a user, such as booking a vacation. While built on generative AI, these systems would not require subsequent human direction or refinement.

Limitations of LLMs

Because of the temperature parameter, insufficient data, or inadequate training, LLMs may hallucinate, generating false or misleading information. Due to the complex nature of transformer architecture, explicitly identifying what led to the hallucination is often impossible.

Language models are limited to text-based generation; image, audio, and video generators use other tools, such as diffusion models.

They are also confined to the information in their data and may propagate stereotypes, discrimination, and bias in their outputs (listen to an expert discuss the issue here). LLMs need ongoing updates and training to incorporate new information and ensure accuracy.