How these things are made

Training, in plain language — and why it explains both the brilliance and the failures.

There is no maths here. Understanding the process in outline is enough to predict most model behaviour, including the parts that seem inexplicable.

The process

Step one: gather text. Enormous quantities — books, websites, code, articles, forums. Effectively a large fraction of what has been written down and digitised.

Step two: guess the next piece. The model is shown text with the end hidden, and asked to predict what comes next.

Step three: correct it. Its guess is compared with the real ending, and every number inside the model is nudged — very slightly — in the direction that would have made the right answer more likely.

Step four: repeat. Billions of times.

That is the whole thing. Everything else is scale, engineering and refinement.

Why this produces something that seems to understand

Because predicting the next word well, across everything humans have written, requires understanding an enormous amount — grammar, facts, style, reasoning patterns, how arguments are structured, what usually follows what.

To predict the end of a sentence about gravity, you need something that behaves rather like knowing about gravity. The capability is a side effect of the task, and it turned out to be a very large side effect.

Why it also explains the failures

The model learned what text tends to look like, not what is true.

Most of the time those coincide — most written sentences are accurate, so imitating them usually produces accurate text. But when the training data is thin on a topic, the model does not become silent. It produces the kind of text that would go there.

That is hallucination, mechanically. Not a lie, not a bug: the same process that makes it useful, applied where it has nothing to draw on.

Where the cutoff comes from

Training takes months and costs a great deal, so it is not continuous. The model is trained on text up to a certain date and then frozen. It knows nothing after that, and — importantly — it usually will not tell you that.

The refinement that changed everything

Raw prediction gives you a model that continues text. It does not give you one that answers.

So a second stage follows: humans rate which responses are better, and the model is tuned toward the preferred ones. This is what turns a text-continuation engine into something that answers questions, follows instructions, and declines requests politely.

It also introduces a subtlety worth knowing: the model is now optimised to produce answers people rate highly. That is not identical to producing answers that are true. Usually the two agree. Occasionally they do not, and the difference is where confident nonsense comes from.

Why all models feel similar

They are all made this way. The differences come from four things: how much text, how much compute, what was filtered out, and how the refinement stage was done.

There is no secret ingredient. The field is unusually open about the recipe, which is why progress has been so fast and why so many organisations can compete.

What this means for you

Three practical consequences, and they are the spine of this whole section:

1. It compresses patterns, not facts. Think of it as a very good impression of knowledgeable writing. 2. It cannot know what it was not trained on — and will not reliably admit it. 3. It is tuned to be liked, which is why it is so agreeable and so fluent, and why you must check it.