LLM
The term behind everything on this site — and what it actually means is more interesting than the name suggests
What is an LLM
LLM stands for Large Language Model. The name tells you three things: it works with language, the model is large, and it is a model — a mathematical system that has learned patterns from data.
The large part is not marketing. The earliest language models in the 1990s had millions of parameters. By 2020, GPT-3 had 175 billion. Today's frontier models are estimated to have hundreds of billions to trillions. The scale is what changed everything.
Where it came from
For decades, computers processed language through rules. If the sentence has this structure, do this. If this word appears, flag it. Rules written by humans, applied by machines.
In 2017, researchers at Google published a paper called "Attention Is All You Need." It introduced a new architecture — the Transformer — that learned patterns from text rather than following rules. Feed it enough text. It learns grammar, facts, reasoning patterns, writing styles. Not because anyone told it to. Because the patterns were in the data at sufficient scale.
GPT-3 in 2020 was the first model to demonstrate that scale alone — more parameters, more training data, more compute — produced capabilities that nobody had specifically programmed. It could write, reason, translate, and summarise without being explicitly trained on any of those tasks.
The number that makes it real
sourc.dev currently tracks 30 large language models. Input prices range from $0.03/1M tokens to $15.00/1M. Context windows range from 8,000 to 1,000,000 tokens. The cheapest model today costs 97% less than comparable capability cost in 2020. Verified March 2026.
Why this matters to you
Understanding what an LLM is — a statistical model that predicts the next token, trained on vast amounts of text at scale — helps you understand both its power and its limits.
It is extremely good at language because language is what it was trained on. It is unreliable on facts it was not trained on, or facts that have changed since training. It does not reason the way a human reasons — it generates plausible continuations. Most of the time that produces correct, useful output. Sometimes it produces confident nonsense.
Knowing the mechanism does not make you a researcher. It makes you a better builder.
Verified March 2026