What a language model is
A transformer turns text into numbers that carry meaning. Encoders read, decoders write, and which one you need depends on the job.
A language model is a next-word guesser trained until the guessing became something close enough to understanding to be useful. Modern ones are transformers. Text is split into subword pieces, every piece gets a vector, and layers of attention let each piece adjust its meaning by looking at every other piece. That is how “Riegel” next to “Milka” resolves to a chocolate bar rather than a door bolt.
The family splits in two. Encoders read a whole text at once and output vectors: coordinates in a space where distance is meaning. They power search, classification, and clustering, and they are small enough to run fast on a CPU. Decoders generate, one token at a time, each conditioned on everything before it. The chat models everyone knows are decoders, and the L in LLM mostly belongs to them: scale is what made generation good enough to hold a conversation, follow instructions, and write code.
The split matters commercially, not just academically. A categorization or search problem usually wants a fine-tuned encoder: milliseconds, no per-request fees, runs on your hardware. Generation, summarization, and agents want a decoder, which is bigger, slower, and worth it exactly where producing text is the job. A lot of money gets spent pointing a decoder at problems an encoder solves for a hundredth of the cost.
So “do you use AI” is the wrong question. The right one is which model, at what size, sits at which step of the pipeline.