Large language model (LLM): the engine behind AI assistants

How it works, in plain terms

Think of your phone’s predictive text, which suggests the next word. A large language model does the same thing, but on a massive scale and with a much finer grasp of context. It splits text into small chunks called tokens, then works out, chunk after chunk, the most likely continuation given everything that came before.

To get there, the model was trained on a very large body of text (its training data): web pages, books, code, articles. In the process, it tuned its parameters to capture the grammar, common facts, styles and ways of reasoning found in those texts.

Where LLMs come from

Most current LLMs are built on the Transformer architecture, introduced in 2017 by Google researchers in the paper “Attention Is All You Need” (arXiv). In 2020, the paper describing GPT-3 showed that a large enough model could handle many tasks from plain instructions and a few examples, without task-specific retraining. ChatGPT, launched in November 2022, then put this technology in the hands of the general public.

Why it matters to you

Knowing that an LLM predicts rather than looks things up explains a lot:

• it writes fluently, even on topics it barely knows;
• it can produce a hallucination, meaning false information stated with confidence;
• its knowledge stops at the end of its training, unless it can search the web.

Models are not interchangeable: some are faster, others better at reasoning or cheaper. The compare models tool puts their prices, capabilities and public results side by side.

Common myths

• “The model learns from every conversation in real time.” No: its parameters are frozen once training ends. A conversation may be used to train a future version, depending on the service’s settings.
• “Bigger is always better.” Not necessarily: for a simple task, a smaller model may be enough, answer faster and cost less.