The mechanism is genuinely simpler than the results suggest, and knowing it changes how you work with these tools. A language model does one thing: given some text, it produces the next small piece of text that best fits. Then it does that again, and again, until it has an answer. That’s it. The reason this sounds too thin to explain a system that can draft a contract or debug a spreadsheet is that we underestimate how much understanding is required to genuinely predict well — you cannot reliably finish the sentence “the reason the invoice was rejected is” without having absorbed a great deal about invoices.
The ability came from learning, not from rules. Nobody sat down and wrote out grammar, or the structure of a good email, or how a balance sheet works. The model was shown an enormous quantity of text and adjusted itself, over and over, until its predictions got good. This is what “machine learning” means in practice, and it explains both the strengths and the shape of the gaps: the model is excellent at things that were well represented in what it read, and unreliable at things that weren’t.
That material — the training data — is a snapshot, and it’s frozen. The model does not consult it afterwards the way you’d search a folder; it absorbed the patterns once and the file cabinet was then thrown away. This is the honest explanation for the knowledge cutoff. Ask about something that happened after training and the model has no way to know it’s missing, so it will often produce a confident, plausible, wrong answer rather than an empty one. The fix is never to hope for better recall — it’s to hand it the current material yourself, which is what context is about.
Underneath the words, the model works in tokens — chunks roughly three-quarters of a word long. This is a plumbing detail that leaks into your day surprisingly often. It’s why a model can be brilliant at reasoning and bad at counting the letters in a word: it never saw the letters. It’s why limits and prices are quoted in numbers that don’t match any word count you’d recognise. And it’s why “how much can it hold at once” has a hard, countable ceiling rather than a vague one.
There is a useful distinction between the two times a model does work. Training happened once, took months, and cost an enormous amount. Inference is every individual run afterwards — the few seconds when your specific question goes in and an answer comes out. Almost everything you experience as “using AI” is inference, and it is what you pay for by the use. It also means each answer is computed fresh rather than looked up, which is why the same question can come back worded differently twice in a row.
A few dials come up in conversation and are worth recognising even though you’ll rarely touch them. Temperature controls how adventurous the wording is — low for anything where correctness matters, higher when you want range from a brainstorm. Fine-tuning is taking a finished model and training it a little further on a narrow set of examples to specialise it; it’s a real technique and almost never the answer to “my results aren’t good enough.” Multimodal just means the model takes images and PDFs alongside text, which is why pasting a screenshot of an error works.
Finally, the word that distorts every conversation: AGI. It doesn’t exist. Today’s models are extraordinary in some directions and have flat blind spots in others, and the gap between those two facts is where most disappointment lives. The practical takeaway from everything above is smaller and more useful than the headlines: the model’s raw ability is fixed, so your leverage is entirely in what you put in front of it and how clearly you ask. That’s prompting and context — and knowing when not to believe the answer is trust and accuracy.
The words
- AI artificial intelligence
- Software that does things we used to assume needed a human: understanding language, recognizing images, making decisions. It's a broad umbrella — today's most useful kind for everyday work is the
LLMbehind tools like Claude. - LLM large language model
- A Large Language Model — the kind of AI behind Claude. It was trained on a huge amount of text until it became very good at one thing: predicting what word should come next. At enough scale, that single skill is enough to hold a conversation, write code, and explain almost anything.
- Machine learning ML
- A way of building software that learns patterns from examples instead of being handed step-by-step rules. Show it thousands of cat photos and it learns what a cat looks like — nobody wrote down "a cat has whiskers." It's the engine under most modern AI.
- Training data
- The huge pile of examples a model learns from — for an
LLM, an enormous amount of text. The model never "looks things up" in it later; it absorbed the patterns once, during training, and that's frozen in place. - Inference
- The moment the model actually runs to produce an answer — its 'thinking' in response to your input. It's the opposite end from training: training is the long, one-time process where the model learned, and inference is every quick run afterward when it puts that learning to use on your specific request.
- AGI Artificial general intelligence
- A hypothetical AI that could match a capable human across basically any task — learning, reasoning, and adapting broadly — rather than being strong at a narrow set of things. Today's models, impressive as they are, are not AGI: they're remarkably good in specific domains but still have real gaps and blind spots. AGI is a goal and a topic of debate, not a product you can buy.
- Fine-tuning
- Taking an already-trained model and training it a little more on a narrow set of examples, so it gets better at one specific job — a certain writing style, a particular kind of task. A tune-up, not a rebuild.
- Multimodal
- A model that handles more than just text — it can take in images, screenshots, PDFs, diagrams, and sometimes audio, all in the same conversation. 'Modal' refers to a mode of input; multimodal simply means it understands several at once, so you can mix a picture and a question freely.
- Temperature
- A dial that controls how random or creative the model's output is. Turn it low and the model plays it safe — focused, predictable, likely to give the same answer twice. Turn it high and it gets more varied and surprising, reaching for less obvious wording and ideas.