How AI Actually "Thinks" — It's Fancy Autocomplete, Not Magic

Part 2 of the series "AI, Without the Hype". Start at Part 1.

AI does not "think." A large language model predicts the next word — really, the next token — over and over, based on patterns it learned from huge amounts of text. That is the whole engine. There is no understanding inside, no reasoning, no little mind in the box. It is autocomplete, scaled up to a size that feels like magic but isn't.

I build these systems in Ahmedabad, and I have shipped them to real people — including SmartON, an assistive AI used by more than 17,000 blind and low-vision users. The question I get most, usually from smart people who feel slightly silly asking, is: "But how does it actually think?" The honest answer is calming once you hear it. So here it is, plainly.

The honest one-line answer

When you type a question, the model reads your words and asks itself one question: given everything before this, what word is most likely to come next? It picks a word. Then it adds that word to the sentence and asks the same question again. And again. Hundreds of times, faster than you can blink.

That is it. An essay, a poem, a block of working code — all built one most-likely-next-word at a time. There is no plan, no idea it is trying to express. Only prediction, repeated until your answer is complete.

The reason this is hard to believe is that the output is so good. But "good prediction" and "thinking" are not the same thing. A weather model predicts tomorrow's rain without understanding the sky. This is the same trick, pointed at language.

What "training on the internet" actually did

You have heard that these models were "trained on the internet." Here is what that means, concretely.

The model was shown an enormous amount of human writing — books, articles, code, conversations. The words were hidden one at a time, and it was asked to guess the missing word. Wrong guess, small correction. Right guess, small reinforcement. Repeat this billions of times.

After enough rounds, something useful emerges. To predict the next word well across that much text, the model absorbs the patterns of how we write:

• Grammar and spelling, because that is how words follow words.
• Facts that appear often, because "The capital of Gujarat is ___" almost always ends one way.
• Reasoning shapes, because explanations and arguments have a predictable rhythm.
• Tone and style, because a polite email and an angry tweet look different.

So the model is not a database of facts. It is a giant statistical map of how humans tend to continue a sentence. That one distinction explains almost everything that follows.

Why it feels like understanding (and why it isn't)

When an AI answers fluently, your brain does what brains do — it assumes a mind is behind the words. For all of human history, fluent language meant a person was there. The model breaks that assumption, and our instinct hasn't caught up.

But fluency is exactly what the model was built to produce. It learned the shape of a good answer without learning the meaning of it. Think of a brilliant student who has read every textbook and can finish any sentence a teacher starts — but has never once been outside the classroom. The words are right. The grounding is missing.

This is why an AI can write a confident, well-structured paragraph about a court case that never happened. It is not lying in the human sense. It is producing text that looks exactly like a true answer would. The packaging is perfect; whether the contents are real is a question the model never asked.

What a "token" is, simply

I keep saying "token" instead of "word," so let me make it concrete, because this one idea quietly explains a lot.

A token is a chunk of text — sometimes a whole word, often a piece of one. "Ahmedabad" might be split into "Ahmed" and "abad." Spaces and punctuation are tokens too. The model does not see letters or sentences the way you do; it sees a stream of these chunks, and it predicts the next chunk.

Why care? Because two everyday mysteries dissolve once you know this:

• Why AI miscounts letters or fumbles a name. It never saw the individual letters — it saw chunks. Asking it to count the r's in a word is like asking someone to count bricks in a wall they only ever saw as finished rooms.
• Why long documents cost more and sometimes get "forgotten." Models have a limit on how many tokens they can hold at once — their working memory. Feed in a hundred-page PDF and you may exceed it, so the earliest parts fall out of view. Getting the right information in front of the model at the right time is its own skill, which I cover in context engineering.

Why this single idea explains both its brilliance and its mistakes

Hold on to one sentence: it predicts the most likely next token. Almost every strength and every flaw falls out of it.

• It explains why AI is brilliant. Prediction trained on humanity's collective writing is shockingly powerful. Most useful answers genuinely are the most likely continuation, so most of the time you get gold.
• It explains why it makes things up. If a confident, wrong answer happens to be the statistically likely continuation, the model produces it without hesitation. It has no separate "is this true?" check.
• It explains why prompting works. Better input narrows what counts as a likely next word. Tell it who it is and what you want, and you steer the prediction toward the answer you need — which is the whole point of learning to prompt well.

This is also why I treat every model I deploy as a brilliant, fast, slightly overconfident intern. Astonishing range, great work in seconds — and the occasional invented fact delivered with a straight face. You would never let an intern send a client email unchecked. Same rule here.

What this means for how much you should trust it

Trust it the way you would trust that intern: for drafts, ideas, summaries, and anything where you can quickly see if it went wrong. Be slower to trust it on exact facts, fresh news, numbers, and citations — where "sounds right" and "is right" drift apart.

The good news: you do not need the maths to use it well. You need this one mental model. Once "it predicts the next word" is in your head, AI stops being a wizard or a threat, and becomes a tool you can reason about. That is the entire purpose of this series.

If you remember one line from Part 2, make it this: AI does not know things — it predicts what a knowing answer would sound like, and your job is to tell the difference.

Next up — Part 3: Why AI confidently lies, and how to catch it.