Table of Contents
What Learning Actually Means for a Machine
A spam filter gets good at catching phishing emails because it has already seen a few million messages that people flagged as junk. Nobody handed it a grammar book or a lecture on social engineering. It absorbed a pattern from examples, and now it applies that pattern to mail it has never encountered before.
That is the whole trick, repeated at enormous scale. An AI model is a large collection of numbers, sometimes hundreds of billions of them, arranged so that data going in one end produces an answer at the other. Learning is the process of nudging those numbers until the answers stop being wrong quite so often.
It sits closer to curve fitting than to teaching. Show a model two million credit card transactions, half of them fraudulent, and it will draw a boundary between the two groups. The boundary is not understanding. It is arithmetic, tuned until the error rate drops.
Parameters are where the knowledge lives
Every weight and bias inside a neural network is a parameter. A small image classifier might carry a few hundred thousand of them. GPT-3 carried 175 billion. Training changes nothing except those values, usually in batches of 32 to 1,024 examples at a time, looped for hours or weeks.
The vocabulary around this gets muddy fast. If the difference between a model, an algorithm and a training pipeline keeps sliding past you, this explainer on artificial intelligence vs. machine learning is a decent place to anchor yourself before going deeper.
The Three Ways Models Learn From Data
Supervised learning: an answer key for every question
Most of the machine learning doing real work inside companies right now is supervised. You feed the model an input and the correct output, side by side. A photo and the label cat. An email and the label spam. A loan application and the label repaid. The model predicts, compares its guess against the label, and adjusts.
Do that half a million times and it starts getting the answer right on photos it has never seen. The catch is cost. A radiologist labelling 50,000 chest X-rays for pneumonia takes months, and every labelling mistake becomes a mistake the model learns to repeat with confidence.
Unsupervised learning: structure without an answer key
No labels here. The model receives raw data and hunts for clusters, anomalies or compressed representations of what it is seeing. Retailers use it to segment customers by behaviour nobody thought to define in advance. Banks use it to surface transactions that do not resemble anything in the recent history.
Self-supervised learning: the trick behind modern language models
This is the hybrid that made large language models possible, and the idea is almost cheeky. The data labels itself. Take a sentence, hide the final word, and ask the model to guess it. The right answer was already sitting in the text, so you get supervision for free, at internet scale, without hiring a single annotator.
Reinforcement learning: learning from consequences
No fixed answer key at all. An agent takes actions in an environment and receives a score. AlphaGo learned by playing millions of games against itself, rewarded only for winning. The same loop now tunes chatbots, where human raters score responses and the model drifts toward the answers people preferred. That reward-driven cycle is also what powers the decision-making inside an intelligent agent in AI, which picks its next action rather than answering a single prompt and stopping.
Inside One Training Step
Here is what actually happens when a model learns from a single batch of data.
- Forward pass. Data flows through the network and produces a prediction. The model says, in effect, 82% confident this is a cat.
- Loss calculation. A loss function measures the gap between that prediction and the truth. If the real label is cat, the loss is small. Had the model answered 12%, the loss would be brutal.
- Backpropagation. The network runs backwards, calculating how much each individual parameter contributed to the error.
- Gradient descent. Every parameter shifts slightly in the direction that reduces the loss. The learning rate, often around 0.001, controls how big those shifts are.
Loop that across the full dataset, once per epoch, and the loss curve should fall. When it flattens, the model has squeezed out roughly what the data contains. Push much further and it starts memorising instead of learning.
Deep networks with dozens or hundreds of layers run this same loop at a scale shallow models cannot match. It is roughly the point where machine learning and deep learning stop being the same conversation.
Data Quality Beats Data Volume
Teams love to brag about dataset size. The problems that kill models are almost never about size.
- Mislabeled examples. Audits of big public datasets have found label error rates in the low single digits. Three percent of a million rows is 30,000 wrong answers teaching bad habits.
- Duplicates. Scraped data is full of near-identical rows. If the same 400 images appear in both training and test sets, your 97% accuracy is measuring memorisation, not skill.
- Historical bias. A hiring model trained on a decade of human decisions absorbs the biases in those decisions and then applies them at machine speed, consistently, to everyone.
- Distribution shift. A fraud model trained on 2019 spending behaviour met March 2020 and discovered the world had moved without telling it.
Overfitting and the Discipline of Generalisation
A model with 99.4% accuracy on its training data can be worse than useless in production. Overfitting happens when the network learns the noise in your sample rather than the pattern behind it, and it gets worse with model size and thin data.
The defence is boring and effective. Split your data, typically 80% to train, 10% to validate while tuning, and 10% held back until the very end. Add regularisation or dropout to make the network less able to lean on any single pathway. Stop training when validation loss starts climbing even as training loss keeps falling. That moment of divergence is the earliest honest signal that the model has stopped learning and started reciting.
How Large Language Models Learn From Text
Pretraining runs the self-supervised trick at a scale that would have seemed absurd a decade ago. Trillions of tokens pass through the model, each one a prediction exercise: given everything so far, what comes next? The loss is cross-entropy, punishing a confident wrong guess far harder than a hesitant one.
That phase teaches grammar, facts, reasoning patterns and a great deal of statistical texture, but it also teaches the model to imitate whatever appeared in the corpus. Fine-tuning on curated examples comes next, then preference tuning with human rankings. The difference between that and a rules-based system classifying support tickets is enormous, and this comparison of generative AI vs. traditional artificial intelligence lays out why the budgets diverge so sharply.
Learning Never Really Stops
Deployment is where a model meets reality, and reality keeps moving. A model trained on last year’s language will start misreading this year’s slang. Customer behaviour shifts, prices change, new fraud patterns appear. None of that requires the model to have been wrong at training time; the world simply left the training distribution behind.
Serious teams monitor input distributions, track accuracy on fresh labelled samples, and schedule retraining before the drift becomes visible in revenue. Some go further and build feedback loops, where user corrections flow back into the next training run. That cycle is one reason the shift toward agentic AI systems changes the operational picture: an agent that acts and observes generates its own stream of training signal.
Questions Worth Asking About Any Model
When someone shows you a model and a headline accuracy number, the useful work starts with a short list of questions.
- What data was it trained on, and does that data resemble the situation I am about to use it in?
- When was that data collected, and how much has changed since?
- What was held out, and was the test set truly untouched until the end?
- Who labelled the examples, and what instructions did they work from?
- How is performance being tracked now, months after the launch demo?
A model is a compressed opinion about the past, produced by a specific dataset, a specific loss function and a specific set of human choices. It will keep behaving like that opinion until someone changes the data underneath it. Knowing how the learning actually happens turns those questions from scepticism into engineering, and it is the difference between adopting a tool and inheriting a liability.
