News

How large language models actually work, no maths required

Ask an AI chatbot to explain how it works and it will happily produce three confident paragraphs, at least one of which is subtly wrong. We’ve tested this enough times to find it funny rather than alarming. It’s actually a decent entry point into understanding how large language models, the technology behind ChatGPT, Google’s Gemini and Microsoft Copilot, actually operate. They’re not thinking. They’re not looking anything up in real time by default. They’re doing something stranger and, once you see it, a lot less mysterious.

Our take: the “it’s just autocomplete” explanation you’ll see shared around is technically defensible and still manages to mislead people, because it undersells how much useful behaviour falls out of doing autocomplete extremely well, at enormous scale. Let’s get into how large language models work without a single equation, because the maths was never the part that mattered to a reader deciding whether to trust one with their business email or their kid’s homework help.

What a language model is actually doing

Strip away the branding and a large language model is a very sophisticated next-word predictor. You type “the capital of France is” and it calculates, from everything it absorbed during training, that “Paris” is overwhelmingly the most likely next word. Then it does that again for the next word, and the next, one token at a time, each guess informed by everything written so far.

A token isn’t quite a word, by the way. It’s a chunk of text, sometimes a whole word, sometimes a fragment like “ing” or “un”, that the model treats as a unit. That’s a detail worth knowing only because it explains why these tools occasionally mangle unusual names or do oddly well with made-up words. They’re not reading like we do. They’re assembling.

The “training” part is where the scale comes in. These systems are shown an enormous amount of text, scraped from books, websites, forums, code repositories, the lot, and adjusted repeatedly until their predictions get closer to what actually appears in that text. Nobody sat down and taught the model grammar rules or facts about geography. It picked up the statistical shadow of those things by being shown enough examples that the patterns became unavoidable.

Why this explains both the magic and the mistakes

This is also why a model can write a passable sonnet and then confidently invent a court case that never happened. Both behaviours come from the same mechanism: predicting plausible next words. Sonnets have strong, learnable patterns. Legal citations look structurally plausible even when the actual content is fabricated, because the model is optimising for “sounds right” rather than “is true”. It has no separate fact-checking step bolted on, unless the product wraps one around it.

Short version: it’s not lying, because lying requires knowing the truth and choosing otherwise. It’s guessing, very fluently.

Where the building blocks came from

The architecture underneath most modern models is called a transformer, a design Google researchers published in 2017. Its trick is something called “attention”, which lets the model weigh which earlier words in a sentence matter most to the next word it’s predicting, rather than just reading left to right in a straight line like older systems did.

If that sounds abstract, think of reading the sentence “the trophy didn’t fit in the suitcase because it was too big.” Working out what “it” refers to requires weighing other words in the sentence against each other, not just marching through in order. Attention gives the model a mechanism to do that weighing, at scale, across thousands of words of context.

It’s worth sitting with how recent this all is. The core idea is well under a decade old. The jump from “interesting research paper” to “tool my accountant in Ballarat uses to draft emails” happened inside about seven years, which is a genuinely unusual pace even by tech’s normal standards of hype compressing timelines.

Training versus chatting: two different jobs

There’s a useful split between training a model and using one, because conflating the two is where a lot of public confusion sits.

Training happens once, in a data centre, over weeks or months, and costs the company doing it a genuinely large amount of money in computing time. This is the expensive, slow phase where the model absorbs patterns from its training data. Once that’s done, the model is effectively frozen. It doesn’t learn from your conversation with it the way a person would remember a chat tomorrow.

Using the model, what the industry calls “inference”, is the quick bit: you type a prompt, it predicts tokens one after another, and you get a reply in seconds. Every conversation starts from the same frozen starting point, dressed up with whatever context you’ve fed it in that session, plus occasionally some retrieved information bolted on through search.

That frozen quality explains why a model can be confidently out of date. If it was trained on data up to a certain point, events after that point simply aren’t in there unless the product has layered a web-search feature on top, which is a separate system entirely.

Why more data doesn’t mean more truth

Here’s where we’ll have a crack at the mildly contrarian bit. The industry narrative has settled on “bigger models, more data, better results,” and for a long while that held up reasonably well. But we reckon the public conversation still badly conflates “more fluent” with “more accurate,” and those are not the same axis at all.

A model trained on more text gets better at sounding like a confident expert in more domains. It does not automatically get better at knowing when it doesn’t know something. That gap, between fluency and calibrated honesty, is the actual unsolved problem, and it’s one that scaling up the training data doesn’t obviously fix on its own. The major labs building these systems publish their own safety and evaluation research acknowledging versions of this, which is a useful tell that it’s not just sceptics on the sidelines saying it.

Verdict: the size of the model is not a proxy for how much you should trust what it tells you, and anyone spruiking parameter counts as a quality signal is selling you something.

What this means for everyday use in Australia

For most Australian readers, the practical upshot is narrower than the hype suggests. These tools are genuinely good at restructuring text you already trust, drafting a first pass of an email, summarising a document you’ll still read yourself, or explaining a concept you can then verify. They’re much shakier as a source of facts you haven’t checked, particularly anything involving numbers, dates, legal thresholds or specific product details.

If you’re using one of the AI-infused browsers or assistants that have popped up this year, the same caution applies, arguably more so, because the slick interface makes the output feel more authoritative than a plain chat window does. We’ve covered how those stack up in our look at AI browsers compared across Australia, and the gap between the best and worst of them is bigger than most people expect.

There’s also a quieter trend worth flagging: AI “agents” that don’t just answer questions but go off and perform multi-step tasks on your behalf, booking things, filling in forms, chaining actions together. That’s a meaningfully different capability built on the same underlying prediction engine, and we’ve broken down what they can and can’t actually be trusted with in our explainer on what AI agents actually do in 2026.

The privacy question nobody asks early enough

One thing that gets skipped in most “how AI works” explainers: what happens to what you type in. Depending on the service and your account settings, your prompts may be used to further train or fine-tune future models, unless you’ve opted out or you’re on a business tier with different terms. That matters more than it sounds, because it means pasting in a client’s personal details or a draft contract isn’t a private act the way scribbling in a notebook is.

The OAIC, Australia’s privacy regulator, has published guidance on AI and privacy obligations for organisations handling personal information, and it’s worth a read if you’re using these tools for anything work-related rather than just drafting a birthday message. The short version is the same one we’d give for any cloud service: assume anything you type could be seen by someone at the company running it, and act accordingly.

This isn’t unique to AI, incidentally. It’s the same lesson Australians had to learn about cloud storage a decade ago, and the same one that eventually produced sensible habits around organising and backing up digital photos rather than trusting one service blindly. New technology, old lesson.

The hardware arms race behind the curtain

None of this runs on a laptop’s regular processor. Training and running these models at scale needs specialised chips, mostly graphics processors originally designed for rendering video games, because it turns out the maths involved in both jobs is similar enough that GPUs became the default choice for AI workloads. That’s part of why Nvidia became one of the most valuable companies on earth almost overnight, and why global GPU supply has been tight enough to affect gamers trying to buy a graphics card for entirely unrelated reasons. We go through the state of that market in our rundown of the Nvidia RTX 50 series in Australia, which is as much a story about AI demand as it is about gaming.

The chips powering your phone are a different story again, tuned for efficiency rather than raw throughput, which is why on-device AI features behave so differently to the cloud-based chatbots. Our comparison of Snapdragon, MediaTek and Apple’s mobile chips covers that split in more detail, if the hardware side interests you as much as the software side does. The National Institute of Standards and Technology in the United States has also started publishing frameworks for evaluating AI system risk, via its AI Risk Management Framework, which gives a sense of just how seriously governments are now taking the reliability question.

So, should you trust it?

Trust the fluency, question the facts, is about as close as we can get to a one-line rule. These models are a genuinely useful piece of writing and reasoning scaffolding, not an oracle, and not a search engine either, however much the interface might nudge you to treat it as both. The most useful mental model we’ve landed on is to treat an AI chatbot the way you’d treat a very well-read colleague who’s extremely confident, occasionally brilliant, and has never once said “I’m not sure” even when they should have.

That’s not a reason to avoid them. It’s a reason to keep a hand on the wheel, particularly for anything involving money, health or legal obligations, where Australia’s consumer protections don’t bend just because an algorithm was involved in the mistake.

— Dale Whitfield, Founding Editor, Tech Geek

Featured image: AI-generated illustration.

Dale Whitfield

Dale Whitfield is the Founding Editor of Tech Geek. He set the brief the site still runs on — clear, useful technology journalism with no hype — and writes its bigger news analysis pieces. Dale has spent years following how technology companies behave and translating their announcements into what they actually mean for readers.

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Back to top button