Every AI system — whether it recommends a song, detects fraud, or generates text — is built on the same foundational ideas. Understanding these core concepts gives you a mental model that applies across all of AI, from a simple email classifier to a massive language model.
These six concepts are not just vocabulary — they describe the actual mechanics of how modern AI systems are built and used. Once you understand how they fit together, you can reason about any AI system you encounter.
A model is the core component of any AI system. Technically, a model is a mathematical function with millions (or billions) of adjustable numerical values called parameters or weights. During training, these parameters are tuned so that for any given input, the model produces the correct output.
A helpful analogy: think of a model as a very complex dial panel. Training turns all the dials to just the right positions so that when you feed in a new input, the correct output comes out the other side. After training, the dial positions are fixed — that is your trained model.
Different types of models are designed for different tasks. A classification model sorts inputs into categories (spam vs. not spam). A regression model predicts a numerical value (tomorrow's temperature). A language model generates text one token at a time by predicting what should come next.
Data is what AI learns from. Without data, an AI system cannot learn anything — it would have no examples to generalize from. The saying in machine learning is "garbage in, garbage out." The quality, quantity, and representativeness of your training data determines the ceiling of what your model can achieve.
Quality matters because a model learns from every example it sees. If your data contains labeling errors, the model learns incorrect patterns. If your data is missing important types of examples, the model will perform poorly on exactly those cases — often the ones that matter most. A facial recognition system trained mostly on photos of one demographic group will be demonstrably less accurate on others.
Training is the process by which a model learns. The model is shown a training example, it makes a prediction, and that prediction is compared to the correct answer. The difference between the prediction and the correct answer is called the loss or error. An optimization algorithm called gradient descent then adjusts the model's parameters slightly in the direction that would have reduced that error.
This process repeats across thousands or millions of examples. Each complete pass through the entire training dataset is called an epoch. Training a small model might take seconds. Training a large language model can take months on thousands of specialized chips, costing millions of dollars in compute.
Provide a training example — show the model an input and the correct output it should produce
Model makes a prediction — the current parameters produce an output for this input
Calculate the loss — measure how far the prediction is from the correct answer
Backpropagate the error — calculate how each parameter contributed to the mistake
Adjust the parameters — nudge each weight slightly in the direction that reduces the error
Repeat across all examples — continue for many epochs until accuracy is acceptable
A prediction is what a trained model outputs when given an input. Depending on the type of model, a prediction might be a category label (classification), a number (regression), a probability score, a generated sentence, a recommended item, or a detected object in an image.
Predictions are never guaranteed to be correct — they are the model's best estimate based on patterns it learned during training. The confidence of a prediction is often expressed as a probability. For example, a spam filter might output "98% spam" rather than simply "spam" — giving the system or operator the chance to decide how to act on that confidence level.
Customer purchased running shoes and a water bottle.
Email with subject "You've won a prize!"
Photo of a chest X-ray.
Likely interested in fitness accessories or sportswear.
96% probability — spam.
Potential opacity in lower left lobe — flag for radiologist review.
Inference is the act of using a trained model to generate a prediction for new data. Training is the learning phase — inference is the usage phase. Every time a user interacts with an AI product, inference is happening: the model is receiving input and producing output in real time.
Inference is typically much faster and cheaper than training because the model's parameters are already fixed — no learning is happening, only forward computation. A model that took weeks to train can serve millions of inference requests per day on much cheaper hardware. This is why AI products can scale to billions of users despite the enormous cost of initial training.
Feedback is how AI systems learn to improve after they are deployed. Raw model predictions are rarely perfect from the start — feedback provides the signal needed to make them better over time.
One of the most important feedback techniques in modern AI is Reinforcement Learning from Human Feedback (RLHF). In RLHF, human raters compare pairs of model outputs and indicate which one is better. The model is then trained to produce responses more like the preferred ones. This is a key technique used to make large language models safer, more helpful, and better aligned with what users actually want.
Model makes a prediction — outputs a recommendation, answer, or classification
User or system interacts with the output — clicks, skips, rates, or flags it
Feedback signal is collected — the interaction is logged as a positive or negative signal
Model or system is updated — parameters are adjusted or training data is improved based on feedback
Future predictions improve — the updated model makes better predictions in similar situations
Imagine building an AI system that estimates house prices. This is a classic machine learning problem, and it perfectly illustrates how all six concepts work together end to end.
Data — Thousands of past house sales, each with features: square footage, number of bedrooms, neighborhood, age of the home, distance to schools, and the final sale price
Model — A regression algorithm (perhaps a gradient-boosted tree or a neural network) that learns to map house features to price
Training — The model sees each house listing, predicts a price, compares it to the actual sale price, and adjusts its parameters to reduce the error — repeated across all listings for many epochs
Prediction — Given a new house (never seen during training), the model outputs an estimated sale price
Inference — When a real estate website inputs a new property's details, the trained model computes an estimate in milliseconds
Feedback — After properties actually sell, the real sale prices are compared to the model's estimates. Large errors are investigated and used to improve future training data or retrain the model
Every AI product you will ever build or use is built on these six ideas. Whether you are working with a simple classifier, a recommendation engine, or a large language model, the same core flow applies: good data trains a model, the trained model makes predictions during inference, and feedback closes the loop to make the system better over time.
Knowing these concepts means you can ask the right questions about any AI system: What data was it trained on? How was it evaluated? Is feedback being collected? What kind of predictions is it actually making? These questions help you use AI tools wisely and understand their limitations.
Imagine you are building an AI system that recommends movies to users of a streaming platform. For each of the six core concepts, describe specifically what it would look like in your system.
Ask the AI if you need help understanding or want to dive deeper into any topic.