Key Takeaways

  • A large language model is trained to do one thing — predict the next token — billions of times, until useful abilities appear as a side effect.
  • GPT-3, the 2020 model that started the modern era, had 175 billion parameters and trained on 300 billion tokens.
  • Prediction is also the weakness: models "guess when uncertain," producing plausible but wrong answers instead of admitting they do not know.

The most powerful software of the decade is, at its core, an autocomplete

Strip away the chat windows, the voices and the headlines. What's left?

A machine that predicts what comes next.

That's essentially what a large language model (LLM) does. Give it a sequence of text, and it predicts the next piece. Then the next. And the next.

It does this so well, and at such enormous scale, that the result starts looking like reasoning, writing and conversation.

But how does predicting text turn into something that can explain quantum physics, write computer code or solve mathematical problems?

The answer starts with a surprisingly simple process.

Step one: text becomes tokens

Before an AI model can work with your words, it has to break them into pieces.

These pieces are called tokens. In OpenAI's definition, "a token can represent a character, part of a word, a whole word, or punctuation."

The English rule of thumb: one token is roughly three-quarters of a word. That means 100 tokens are about 75 words.

But don't assume every word gets treated equally. Unusual spellings, code and other languages can be broken into pieces differently.

Once the text is split, each token is converted into a list of numbers called an embedding. Think of it as giving every piece of text a mathematical address in a space of meanings, where related tokens sit near each other.

The model doesn't see words the way you do. It sees numbers representing pieces of language.

And that's only the beginning.

Step two: attention decides what matters

The breakthrough that made modern LLMs possible was the transformer, introduced in a 2017 research paper with a title that says almost everything: "Attention Is All You Need."

Its authors built a network "based solely on attention mechanisms, dispensing with recurrence and convolutions entirely."

So what exactly is attention?

Imagine reading the sentence "the river bank."

Your brain understands that bank refers to the side of a river, not a financial institution.

Now change the sentence to "the bank account." Suddenly, bank means something entirely different.

Attention helps the model make that distinction.

As it processes each token, it weighs which other tokens in the surrounding context matter most to understanding it.

That happens across stacks of attention layers, interleaved with ordinary computation. Together, they allow the model to build a rich internal picture of the text: who did what to whom, what a pronoun refers to and what a question is really asking.

In other words, the model doesn't simply look at the last word and guess blindly. It uses the surrounding context to make its prediction.

Step three: training is prediction, repeated 300 billion times

Now comes the part that makes the whole system powerful: training.

Imagine showing a model a sentence with its final token hidden.

It makes a guess. If the guess is wrong, its internal settings — called parameters — are adjusted by a small amount.

Then it tries again.

And again.

This prediction-and-correction process is repeated across enormous amounts of text.

GPT-3, the 2020 model documented in its own research paper, had 175 billion parameters, "10x more than any previous non-sparse language model."

All its variants were "trained for a total of 300 billion tokens."

The training mix included filtered Common Crawl web text (60 per cent of the mix), curated web text, two book collections, and Wikipedia.

Here's the fascinating part: nobody programmed grammar, facts or arithmetic directly into those 175 billion numbers.

Instead, the paper's own summary says that pre-training develops "a broad set of skills and pattern recognition abilities" in the model.

Why?

Because predicting language well requires more than memorising which words tend to follow others.

To predict what comes next in a conversation about gravity, the model benefits from learning patterns about gravity. To predict the next step in a mathematical solution, it benefits from learning how mathematical solutions work.

Prediction at this scale forces the model to absorb patterns from the world described by the text.

And that's where things get interesting.

Why a predictor can look like a thinker

If an LLM only predicts the next token, how can it solve problems that require several steps?

One important answer is chain-of-thought.

A 2022 paper showed that when a model generates "a series of intermediate reasoning steps" — essentially writing out its working before answering — its performance on complex reasoning improves significantly.

The researchers also found that these abilities "emerge naturally in sufficiently large language models."

A companion paper defines emergence precisely: an ability that is "not present in smaller models but is present in larger models."

Think about what that means.

Instead of jumping straight to an answer, the model produces intermediate steps. Each step becomes part of the context used to predict the next one.

Show the working, in other words, and the next-token predictor becomes a passable mathematician.

Not because it was taught maths as a subject, but because text full of worked solutions taught it the shape of solving.

But predicting text well isn't enough to make an AI assistant useful. It also needs to learn how people want it to respond.

That's where alignment comes in.

After pre-training comes alignment. The InstructGPT work showed the recipe: fine-tune on demonstrations written by human labelers, then optimise against human rankings of the model's outputs (RLHF).

The result was striking. The paper's famous finding was that outputs from the 1.3-billion-parameter tuned model "are preferred to outputs from the 175B GPT-3, despite having 100x fewer parameters."

Raw prediction makes a model capable. Human feedback makes it usable.

The flaw built into the trick

Here's the catch: the same mechanism that makes LLMs impressive also makes them unreliable.

Remember, the model is trained to predict what comes next. It isn't automatically trained to stop and admit when it doesn't know something.

An OpenAI research paper on hallucination puts it plainly: language models "guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty."

The paper also argues that training and evaluation systems "reward guessing over acknowledging uncertainty."

That creates a fundamental problem.

A model rewarded for always answering will often produce an answer, even when it doesn't have the information needed to get it right.

And because it has learned to produce fluent language, an invented answer can sound just as convincing as a correct one.

There are also limits to what outsiders actually know about today's most advanced models.

The labs do not publish parameter counts or training sizes for current frontier models. That means every circulating figure for them is rumour.

And there's another important distinction: "emergence" describes observed behaviour, not a settled account of what is happening inside.

We can observe what these models do. That doesn't mean we fully understand why every capability appears.

The Twikup Take

An LLM is a prediction engine that became so good at predicting text that it absorbed the patterns text contains — grammar, facts, arguments and worked solutions.

That's genuinely remarkable. But it's also the source of the problem.

The same system that produces an impressive explanation can confidently invent an answer. Fluency and fabrication come from the same underlying mechanism.

The smartest way to use AI is to appreciate what it can do without confusing a convincing answer with a correct one.

Image: Illustrative image.

Sources