Large Language Model (noun, “LARJ LANG-wuhj MAH-del”)
A large language model — or LLM — is an artificial intelligence computer program that can read text and write responses. LLMs power many modern artificial intelligence, or AI technologies, such as ChatGPT.
LLMs generate human-sounding text using computer programs that “learn.” These algorithms learn by analyzing huge amounts of data. The data might be the text on webpages, online books or social media posts. Scientists “train” algorithms by providing them these data.
An LLM first breaks the training data into smaller pieces, such as words or parts of words. These are called tokens. The system then divides the tokens into categories. It also maps out relationships and patterns between them. Computers don’t use words to solve problems. They use math. So, an LLM must assign numbers to text tokens. That allows the LLM to calculate responses based on probabilities.
Once trained, an LLM can read a new text prompt — such as a question — and break it into tokens. Then, the LLM puts together a response one token at a time. It calculates which tokens are most likely to come next based on patterns learned in training.
Humans are very good at detecting these patterns. For example, imagine a friend telling you to “turn right.” The word “right” has several meanings. “Right” could mean accurate. It could also mean morally good. But these alternate definitions wouldn’t confuse you. In this context, coming after the word “turn,” the word “right” clearly refers to a direction. You’ve learned that from experience.
LLMs can also use their experience from training to figure out meanings. LLMs do this with another program called a transformer. A transformer focuses on a bit of text, such as a sentence. It considers multiple meanings that tokens in sentences might have. Then, it decides which meaning is most likely intended.
LLMs are powerful AI tools. However, they also come with problems. The language data used to train LLMs sometimes bakes biases into these systems. They also tend to oversimplify some topics. And they have limited ability to verify their information. That’s because they do not understand the concept of truth or accuracy as a person does.
Despite how humanlike LLMs may sound, it’s crucial to remember their limits. LLMs do not truly comprehend the text they generate. While they mimic some patterns of human reasoning, they do not have feelings and cannot understand emotions like a person.
In a sentence
Large language models (LLMs) can power AI job-screening technology that comes with cooked-in biases, favoring applicant names that sound male.
