06. How AI Works

AI works by finding patterns in data and using those patterns to make predictions on new data.

All AI — from spam filters to GPT-4 — follows the same fundamental loop:

flowchart LR
    D["Data\nLabeled / Unlabeled\nReward signals"]
    TR["Training\nAdjust weights\nMinimize loss"]
    M["Model\nMathematical function\nFrozen weights"]
    P["Predictions\nNew unseen data\nMilliseconds"]
    L["Loss\nHow wrong?"]

    D --> TR
    TR --> M
    M --> P
    P --> L
    L -->|"Backprop\n+ Gradient descent"| TR

    style D fill:#1e1b4b,stroke:#7c3aed,color:#e2e8f0
    style TR fill:#172554,stroke:#3b82f6,color:#e2e8f0
    style M fill:#14532d,stroke:#059669,color:#e2e8f0
    style P fill:#451a03,stroke:#f59e0b,color:#e2e8f0
    style L fill:#450a0a,stroke:#ef4444,color:#e2e8f0

AI needs examples to learn from. The more, and the better quality, the better.

• Labeled data (supervised): emails tagged spam/not-spam
• Unlabeled data (unsupervised): raw customer behavior logs
• Reward signals (reinforcement): score from winning/losing a game

Without good data, AI cannot learn anything meaningful. “Garbage in, garbage out.”

The algorithm processes data repeatedly, adjusting internal parameters (weights) to minimize errors.

Imagine learning to throw darts:

1. Throw → miss → adjust aim
2. Throw → miss less → adjust again
3. Repeat until accurate

Training works the same way:

1. Feed input → model makes prediction
2. Compare to correct answer → calculate loss (how wrong it was)
3. Adjust weights using backpropagation + gradient descent
4. Repeat millions of times

After training, the model is a function:

f(input) → output

f("Is this spam?") → 0.97  (97% spam)
f("image of cat") → "cat"
f("translate to French") → "Bonjour"

The model is just a mathematical function — a very complex one with billions of parameters. It doesn’t “understand” anything. It finds the best mapping from input to output based on training.

Using the trained model to make predictions on new, unseen data.

New email arrives → model scores it → 0.89 → mark as spam

Training is expensive (days/weeks, massive compute). Inference is cheap (milliseconds, single GPU or CPU).

The model doesn’t learn facts or rules the way humans do. It learns a mathematical mapping:

P(output | input) = some very complex function of weights

When GPT writes code, it’s not “thinking about programming” — it’s computing the most likely next token given billions of patterns seen during training.

Modern AI is powerful because of scale:

GPT-4 works because: transformer architecture + trillions of tokens of text + massive compute.

Q: How does a machine learning model learn?

A: It learns by iteratively adjusting its internal parameters (weights) to minimize a loss function. In each training step: the model makes a prediction, the error is calculated, and backpropagation updates the weights via gradient descent. After millions of iterations across the training data, the model finds a parameter configuration that maps inputs to correct outputs well.

Q: What is the difference between training and inference?

A: Training is the process of learning — feeding data through the model, computing loss, and updating weights. It’s computationally expensive and happens offline. Inference is using the trained model to make predictions on new inputs. It’s fast and is what users experience in production. Once trained, the model weights are frozen — inference doesn’t change them.