what’s an LLM?
A large language model is a neural network trained on huge amounts of text to predict the next token in a sequence, which at sufficient scale produces fluent language.
modern LLMs (ChatGPT, Claude, Gemini, Llama) are all built on the transformer architecture introduced by Vaswani et al. in 2017. the transformer's core innovation is self-attention: when processing a word, the model directly weighs its relationship to every other word in context, without reading strictly left-to-right.
LLMs are trained autoregressively: given a sequence of text, predict the very next token. (a token is roughly a word-piece; “tokenization” is the LLM's vocabulary.) repeat this billions of times across the public internet, plus books, code, and scraped data, and the model learns grammar, facts, reasoning patterns, and stylistic conventions implicitly.
at generation time, the model produces text one token at a time, each conditioned on everything written so far. every token is chosen from a probability distribution over the vocabulary, and that probability distribution is exactly the signal text detectors exploit. “large” in LLM refers to parameter count. modern frontier models have hundreds of billions to trillions of parameters.
three things a layperson needs to know:
LLMs don't “know” facts. they predict statistically likely text. that's why they hallucinate confidently when their training data is thin or contradictory on a topic. confident-sounding output is not the same as correct output.
every word an LLM writes is sampled from a probability distribution. AI-text detection exploits this directly: the average probability of an LLM-written sequence under a reference language model is detectably lower than the average probability of a human-written sequence of comparable topic. more on this in perplexity.
the same architecture (transformers) now powers most image, audio, and video generation too. diffusion transformers like Sora are transformers under the hood. the transformer is the most general-purpose AI substrate of the decade.
LLMs generate plausible-sounding token sequences. behavior that looks like reasoning emerges from pattern-matching at scale. that distinction shapes how you evaluate LLM output and how detectors flag it.
questions
what is a large language model (LLM)?
a large language model is a neural network trained on huge amounts of text to predict the next token in a sequence, which at enough scale produces fluent language. ChatGPT, Claude, and Gemini all sit on the transformer architecture from 2017. its core idea, self-attention, lets the model weigh every word against every other word in context instead of reading strictly left to right.
do LLMs understand or know facts?
no. an LLM predicts statistically likely text; it does not store and retrieve facts the way a database does. that is why it hallucinates with full confidence when its training data runs thin or contradictory on a topic. confident output and correct output are separate things, which shapes how much you should trust anything an LLM writes.
how does AI-text detection use the way LLMs work?
an LLM samples every word from a probability distribution over its vocabulary, and detectors read that pattern. the average probability of an LLM-written passage under a reference language model sits measurably lower than a comparable human-written passage. perplexity-based detectors measure that gap.
do LLMs reason like humans?
an LLM generates plausible token sequences, and behavior that looks like reasoning comes from pattern-matching at scale. that distinction shapes how much you trust the output and how detectors flag it.
sources.
- 01Vaswani et al. — Attention Is All You Need (arXiv:1706.03762)The 2017 transformer paper that made modern LLMs possible.
- 02
- 03