Explain like I'm five · AI

What is an LLM?

Large language models, told as a story about a parrot that read a mountain of books and learned to continue any sentence.

Paris is the capital of Francecapital of France? …and write me a poemask anything, and a sensible answer comes out — how?

Ask the parrot anything, and a sensible answer comes out.It knows capitals, writes poems, fixes code. How?

not thinking —Todaytheweatheris?NEXT BEANnice ★★★★cold ★★banana ☆it picks the bean that most often came next in its books — again and againthe parrot is a bird that continues sentences well — not one that understands them

The parrot doesn't understand — it continues the sentence.It picks the word piece that most often came next in its books — again, and again, and again. That becomes a sentence.

more books than the village libraryone bean at a timetrainerWHAT IT READencyclopedias, novels,letters, code…an LLM is a parrot that read a mountain of books and got very good at picking the next bean

An LLM is a parrot that read a mountain of books and got very good at picking the next bean.It read encyclopedias, novels, letters, even code. So whatever you say, it can continue — one bean at a time.

Knows what it readnot what wasn't in the books
One bean at a timecontinuing, not thinking
Answers are draftsfast and plausible
A person checksnames, numbers, dates most
big parrotknows more, slower, priciersmall parrotfast and cheap, sometimes shortDRILLbe politeafter the reading, a short drill fixes its mannerssize is parameters; the drill is fine-tuning and RLHF

Parrots come in sizes, and after the reading they get a short drill.A big parrot knows more but is slower and pricier. The drill fixes its manners — be polite, say so when you don't know.

GOOD ATsummarizing, translating, draftsfixing code, changing toneanything it read a lot aboutfaster than people, and sometimes betterthat book came out in 1987(no such book exists)it continues just as confidently when it doesn't knowso a parrot's answer is a draft — a person checks it

It's good at a lot. But it continues just as confidently when it doesn't know.So a parrot's answer is a draft; a person checks names, numbers, dates. The confident lie has its own story: the plausible parrot.

In one breath

LLM = a parrot that read a mountain of books and got very good at picking the next word piece. It continues rather than understands, so its answer is a draft and a person checks it.

A neural network trained on vast text to predict the next token. Size is parameter count, what it read is training data, the drill afterwards is fine-tuning and RLHF. Because it continues statistically rather than understands, it can produce confident errors (hallucinations).

When grown-ups say it

Large language model
The mountain-of-books parrot. GPT, Claude, Gemini, Llama are all this bird.
Token
One bean. The unit the parrot counts words in. → the parrot counts words in beans
Next-token prediction
Picking the next bean. The whole of what the parrot does. Sentences, poems, code all come from this.
Training data
The pile of books it read. It doesn't know what wasn't in the books — or what happened after it finished reading (knowledge cutoff).
Parameters
The parrot's size. Billions to trillions. Bigger knows more, runs slower, costs more.
Fine-tuning · RLHF
The drill after reading. Fixes its manners. → the short drill
Hallucination
The confident lie. It continues even when it doesn't know. → the plausible parrot
Prompt
The trainer's note. How you tell the parrot what to do. → the trainer's note
Next story
The Parrot Counts Words in Beans →
Want another term as a picture book? Request →