What is a Large Language Model (LLM)? The Technology Behind ChatGPT Explained

A large language model is a neural network trained to do one thing: predict the next fragment of text.

Everything else — writing code, arguing a position, translating, summarising, holding a conversation — falls out of doing that one thing extremely well across an enormous amount of text. Nobody built in the ability to write a function or structure an argument. Those emerged, because both are useful for predicting what comes next in text written by people who could do them.

That gap between the simplicity of the objective and the range of the result is the genuinely surprising thing about this technology, and it is worth understanding properly.

What "Large" Refers To

Two things, both of which grew by orders of magnitude in a few years.

Parameters — the adjustable numbers inside the network, tuned during training. Early language models had millions. Current frontier models have hundreds of billions. These are where everything the model learned is stored, distributed across the whole network in a form nobody can read directly.

Training data — the volume of text processed, now measured in trillions of words. Effectively a large fraction of the accessible written internet, plus books, code, and reference material.

Both mattered, and the second may have mattered more than was initially appreciated. Some of the biggest gains have come from better curation rather than more scale.

What Attention Changed

Before transformers, models processed text sequentially, one word after another, carrying forward a compressed summary of what came before. That summary degraded with distance. By the end of a long paragraph the beginning had largely faded, which made long-range dependencies — a pronoun referring back three sentences, a clause that only makes sense given the opening — genuinely hard.

Attention removed the sequence constraint. When processing any given word, the model can weigh how relevant every other word in the context is, regardless of how far away it sits. "It" in a long sentence can attend directly to the noun it refers to.

Two consequences followed, and the second is why this became an industry.

First, long-range context stopped degrading, which made coherent long-form output possible.

Second — and this is the part that changed the economics — attention is parallelisable in a way sequential processing is not. Every position can be computed simultaneously. That made training on internet-scale data feasible on GPU hardware, and without it the data volumes these models need would simply have been unreachable.

The architecture is a specific application of deep learning, and the general mechanism of prediction is covered in how AI actually works.

Tokens, Briefly

Models do not read words. They read tokens — usually word fragments. "Unbelievable" might arrive as three pieces.

This matters for two practical reasons. Context windows are measured in tokens, so the amount you can put into a conversation is bounded in these units rather than in words. And API pricing is per token, so cost estimates need the conversion. The working rule of thumb is roughly three-quarters of a word per token, meaning a thousand tokens is around 750 words.

It also explains a class of odd failures. Asking a model to count letters in a word or reverse a string often goes wrong, because it never saw individual letters — it saw fragments.

What Emerged, And What Did Not

The capabilities that appeared without being designed for are worth listing, because the pattern in them is informative.

It writes fluently across registers, because the training text spanned all of them. It translates, because parallel text appears throughout the internet. It writes and debugs code, because code is text with unusually consistent structure — which is also why coding is where these models are strongest. It follows multi-step reasoning to a degree, because explanations of reasoning appear extensively in written material.

What did not emerge is equally consistent: anything requiring a model of the world rather than a model of text about the world.

It has no reliable sense of its own uncertainty. It cannot verify a claim. It does not know what any word refers to in the physical sense. And it cannot learn from your conversation — weights are frozen after training, which is why every session starts from the same base regardless of what you told it yesterday.

The Limitation That Matters Most

Generating a true statement and generating a convincing false one are the same operation.

There is no internal truth check. When the model lacks solid grounding — an obscure fact, a specific number, a citation — it does not stop or hedge. It produces what a plausible answer would look like, in the same fluent, confident register as everything else.

This is not a defect awaiting a fix. It is what the objective produces, and it is why the single most useful working rule is: transformation is reliable, recall is not. Summarising a document you supplied is safe because everything needed is present. Asking for a statistic requires verification, every time. Why AI fails covers this and the other failure modes.

Scale Stopped Being The Story

For several years the field ran on a simple assumption: bigger models, more data, better results. That held remarkably well, and then flattened.

What replaced it is less dramatic and more consequential. Training quality, data curation, and efficiency improvements now produce more per unit of compute than raw scale does. Smaller models match much older larger ones on most real work, which changes the economics of building anything — a shift covered in AI trends 2026.

The practical effect is that model choice is now less about capability ceilings and more about cost, latency, and whether you need to run it on your own hardware. That last consideration is why open-weight models matter, and DeepSeek is the clearest example of the trade.

The Major Ones

Rather than a comparison table that will be wrong in three months: they genuinely differ in writing style, willingness to take a position, long-document handling, and tooling around them, and those differences are stable enough to have a preference about.

ChatGPT vs Claude covers the two most commonly compared, ChatGPT alternatives covers the wider field, and AI personal assistants covers where these sit against the voice assistants they are often confused with.

Why This Is Worth Knowing

Not to use the tools — you can use them fine without any of it.

It is worth knowing because it tells you what to trust. A system that predicts likely text is excellent at anything where the material is in front of it and unreliable at anything it must retrieve. It is agreeable because it was trained on human preference ratings, so you should explicitly ask it to argue against itself. It has a knowledge cutoff, so recent events are a blind spot unless the product adds search.

Every one of those follows from the architecture rather than from any product decision — which is why understanding the architecture, even at this depth, is more useful than any list of tips. AI vs machine learning vs deep learning places all of this within the wider field if you want the map.