How Machines Learn

Before a model can answer anything, it has to be trained. The loop is simpler than you'd expect, and it's worth knowing because it explains where a model's strengths, blind spots and personality all come from.

Take a model that starts out useless — its parameters are random numbers, so its predictions are noise. Show it a piece of real text with the last word hidden. It guesses. Compare its guess to the actual word and measure how wrong it was. Then nudge every parameter a tiny amount in the direction that would have made the guess less wrong. That's one step. Now do it again, trillions of times, across a substantial fraction of the public internet.

Nobody writes the final behaviour. Grammar, tone, arithmetic, the ability to write a limerick about a tax return — none of it is programmed. It emerges from the pressure of getting better at prediction on a very large amount of text. This is the single most counterintuitive fact about modern AI, and it's why even the people who build these systems can't fully explain any individual answer.

That first phase is called pre-training, and it produces something that completes text but isn't yet an assistant. Ask a raw pre-trained model a question and it might reply with another question, because on the internet questions are often followed by more questions. Turning it into something helpful takes two more phases.

Fine-tuning comes next: further training on curated examples of the behaviour you want — questions followed by good answers, in the right tone. Then a preference stage, often called RLHF (reinforcement learning from human feedback), where humans compare pairs of responses and the model is tuned toward the preferred one. This is where helpfulness, the house style, and the refusals come from. When a model declines something, that's not a rule in a config file — it's a learned tendency, which is exactly why refusals are sometimes inconsistent.

Now the practical consequences. The first is a knowledge cutoff. Training ends on a particular date, so anything after it simply isn't in the model. Ask about last week's news and you'll get either an admission of ignorance or, worse, a confident guess. Many products now paper over this with live web search — and when they do, they cite sources. If an answer about recent events has no citations, treat it as a guess.

The second is bias, and it's less mysterious than it sounds. A model trained mostly on English text from certain parts of the internet will be strongest on those topics, in that language, reflecting those assumptions. Ask for "a photo of a doctor" and older image models skewed heavily male, because that's what their training images skewed toward. It isn't malice and it isn't a bug in the usual sense. It's a mirror, and mirrors reflect whatever you point them at.

The third catches people out constantly: your chats do not teach the model. The model's parameters are frozen when training ends. If it remembers your name within a conversation, that's the context window, not learning. If it remembers across conversations, the product is saving notes and feeding them back in. Some providers may use conversations to train future versions, which is a privacy question worth checking in the settings — but it is not the model learning from you in real time.

Put those together and you get a genuinely useful instinct. Before trusting an answer, ask: was this likely well represented in the training data, and is it stable over time? Grammar rules, established science, how a common programming language works — well represented, stable, generally reliable. This year's pricing, a niche local regulation, last month's release — thin, volatile, verify it.