Lesson 1: AI Buzzword Bingo
- What AI is.
- How Machine Learning fits inside AI.
- The difference between traditional ML and Deep Learning.
- What Foundation Models are and where LLMs and Generative AI fit.
Let’s start by mapping out the key terms and seeing how they fit together.
Artificial Intelligence (AI)
AI is the master term, the big umbrella. It is about getting machines to do things that usually require human thinking—things like understanding language, recognising images, or making decisions. It’s been around for decades.
Machine Learning (ML)
Machine Learning is a subset of AI. Instead of programming every rule by hand, we give the computer data and let it learn algorithms that can recognize patterns and make accurate inferences about new, unseen data.
The central premise is simple: if you optimize a machine's performance on a dataset (model training), that model can apply learned patterns to infer correct outputs for real-world tasks (inference).
Most machine learning is grouped into three paradigms:
1. Supervised Learning
Trains a model to predict the correct output using labeled examples (ground truth). For instance, emails labeled as "spam" or "not spam".
Types of Models:
- Regression: Predicts continuous numerical values (e.g., price prediction, temperature).
- Linear Regression: Finds the best-fit line.
- Polynomial Regression: Captures non-linear relationships.
- Classification: Predicts discrete classes.
- Binary: Fraud or legit.
- Multi-class: One of several categories.
- Multi-label: Multiple tags at once.
Modern supervised learning frequently uses ensemble methods, combining multiple models for better accuracy. Semi-supervised learning sits between supervised and unsupervised, using a small set of labeled examples to generalize over a larger pool of unlabeled data.
2. Unsupervised Learning
Uses unlabeled data to discover structure on its own.
Common Families:
- Clustering: Groups similar items together.
- K-Means: Partitions data into k groups. Useful for customer segmentation (e.g., bargain hunters vs. big spenders).
- Hierarchical: Builds a tree of similarity, useful for organizing data like IT tickets into themes.
- Dimensionality Reduction: Reduces data complexity while retaining meaningful characteristics. Common techniques include PCA (Principal Component Analysis) and encoders.
3. Reinforcement Learning (RL)
Optimizes a policy through trial and error. An agent interacts with an environment by observing its state and choosing an action. The environment provides rewards (for correct moves) or penalties (for mistakes).
Over time, the agent balances exploration (trying new actions) with exploitation (repeating what works) to maximize long-term rewards. Example: A self-driving car learns to steer and brake by receiving rewards for safe driving and penalties for collisions.
Deep Learning (DL)
Deep Learning is a subset of ML that uses neural networks with many layers to learn hierarchical representations. It is especially effective with large amounts of unstructured data like images, audio, and natural language.
Foundation Models
These are big neural networks trained on huge datasets to capture broad knowledge. Instead of starting from scratch, you fine-tune these pre-trained models for specific tasks.
Large Language Models (LLMs)
A specific type of foundation model focused on text.
- Large: Billions of parameters.
- Language: Trained to understand and generate human language.
- Model: The algorithms and parameters working together.
While modern buzz focuses on LLMs and Generative AI, these systems are built on classic ML foundations. They use transformer architectures (a newer neural network design) but rely on the same principles of training, inference, and pattern recognition. Even RLHF (Reinforcement Learning with Human Feedback) is just classic RL applied to align LLMs with human preferences.
What are parameters?
Parameters are the numeric values inside a model (weights and biases). During training, the system adjusts these numbers to match training data. A simple analogy is a wall of tiny dials; each dial controls how one signal affects another. More parameters mean more capacity to learn patterns, but also higher cost and latency.
Generative AI (GenAI)
Generative AI refers to systems that can create new content—text, images, music, code. Foundation models provide the intelligence, and generative AI turns that intelligence into tangible output.