Bolds the start of each word so you can scan the text faster.
Theme
Language
App
How does the AI "understand"?
~ min read
📋30-second summary
The AI predicts the next word, one at a time. It doesn’t “understand” in the human sense.
It was trained by reading huge amounts of text. It doesn’t memorize: it learns patterns and rebuilds on the spot.
It doesn’t “know”, it estimates. When it’s wrong, it’s wrong with the same confidence as when it’s right.
The more context you give, the better it answers: context narrows the field of probable words.
It doesn’t remember you between chats. Within the same chat yes, but with a cap on how much it can reread.
You’ve opened an AI, you’ve written to it, it answered you in a sensible
way. The natural question is: how? Under the hood there is no one
thinking, no intelligence in the human sense. There is a mechanism that
is fairly easy to understand, but surprisingly powerful.
The heart of a conversational AI is one single operation: predict what
the next word will be.
When you write it a question, the AI reads everything you wrote and picks
which word to put first in the answer. Then it rereads everything (your
question plus its first word) and picks the second. Then the third. One
word at a time, until it closes the answer.
👤
You
Can you explain in two sentences what a black hole is?Can you explain in two sentences what a black hole is?
🤖
AI
A black hole is a region of spacetime where gravity is so strong that nothing, not even light, can escape it. It forms when a very massive star, at the end of its life, collapses on itself under its own weight.A black hole is a region of spacetime where gravity is so strong that nothing, not even light, can escape it. It forms when a very massive star, at the end of its life, collapses on itself under its own weight.
Seen from the outside it’s an articulate answer. Seen from the inside
it’s the same choice repeated word after word: “Given all these words,
which one makes the most sense to put now?”
To pick the right word, the AI was trained by reading enormous amounts
of text: books, articles, web pages, code, conversations. Hundreds of
billions of words. Its developers fed them to it, automatically gathering
material largely from public sources on the internet. From all that
material it picked up patterns: how English sentences are built, which
words typically appear next to each other, how arguments are made on a
topic, how to answer politely, what the most frequent answers to a math
question are.
It hasn’t memorized the texts: they are not inside it like in a database.
To put it concretely: it doesn’t have the Wikipedia page about black
holes filed away. It learned how an encyclopedia about black holes is
written, and when you ask, it rebuilds one on the spot. This difference
explains why it often nails the general content but can get a specific
detail wrong (we’ll come back to this in the next lesson).
Knowing what’s behind the scenes helps you understand why it behaves the
way it does.
It doesn’t “know” anything, it estimates. When it answers, it
isn’t consulting an archive of truths. It is picking the most probable
words given your question and what it has learned. Often it nails it.
When it gets it wrong, it does so with the same confidence.
The more Everything the AI has in front of it while it generates the answer: your question, previous replies in the same chat, any documents. you give it, the better it answers. Every useful
piece of information you add to the What you write to the AI to get an answer. The clearer and more contextualized it is, the better the AI answers. narrows the field of
probable words and pushes it toward an answer that fits you better.
That is why a clear prompt works better than a vague one, as you saw
in the previous lesson.
It doesn’t remember you between one chat and the next. Every new
conversation starts from zero. Within the same chat, however, it
remembers everything you’ve said, because it rereads it every time as
part of the “question” it bases the next word on.
Because what it was trained for, picking the next word well, is very
close to “writing well”. And the texts it was made to read were, mostly,
written by people who had figured something out. The AI itself doesn’t
understand, but it recycles the patterns of those who had. So its
reasoning, on the surface, has the shape of solid reasoning.
Sometimes it really is. Sometimes it is words that link together nicely
but that, if you check, turn out to be wrong. Telling the two cases
apart is the topic of the next lesson.
The conversational AIs we’ve described are called, in jargon, LLMs,
short for Large Language Model. “Model” here means the mathematical
program (a huge collection of formulas) that powers them. “Large”
because it has read a vast amount of text during training and because
that program is genuinely huge. When you hear “LLM” referring to ChatGPT
or Claude or Gemini, this is what they mean.