If a model is trained to predict text rather than to be correct, what should you always verify before trusting its output?
Every machine learning system answers the same question: what counts as correct? In supervised learning, a human labels the answer. In unsupervised learning, there is no label — the machine finds structure that wasn’t defined in advance. In reinforcement learning, the machine tries things and receives rewards or penalties. These are not interchangeable. The training signal shapes everything the model can do — and everything it cannot.
| Paradigm | Training signal | Data requirement | What it learns | Where it fails |
|---|---|---|---|---|
| Supervised | Human-labelled examples (input → correct output) | Large labelled dataset — expensive and slow to create | To map inputs to known output categories or values | When labels are wrong, biased, or missing. Can only categorise what it saw in training. |
| Unsupervised | None — the machine defines its own signal by finding patterns | Large unlabelled dataset — much easier to collect | Hidden structure: clusters, patterns, anomalies, relationships | No way to verify if the patterns it found are meaningful. Human interpretation still required. |
| Reinforcement | Reward function — defined by the designer, not labelled data | An environment to act in, plus a reward signal | A policy: which action maximises cumulative reward over time | Reward hacking — optimising the signal rather than the intent. Designing the right reward is hard. |
DeepMind’s AlphaFold (2020) solved the 50-year protein folding problem. Trained on a database of ~170,000 proteins with experimentally determined structures — labelled by scientists over decades. Given an amino acid sequence, it predicts 3D structure with near-atomic accuracy.
Spotify’s Discover Weekly uses collaborative filtering — an unsupervised approach that groups users by listening behaviour without any human saying “these users are similar.” The system finds latent clusters in 600M+ users’ play histories and maps tracks to those clusters.
GPT-4’s base model was self-supervised (trained to predict the next token). RLHF — Reinforcement Learning from Human Feedback — added the alignment layer: human raters score outputs, a reward model learns those preferences, and RL fine-tunes the base model toward responses humans rate higher.
Every frontier LLM — GPT-4, Claude, Gemini — begins with self-supervised pre-training. The model is given a sequence of text with words masked out and asked to predict them. The labels come from the text itself. This scales to the entire internet.
A model can only do what its training signal rewarded. Understanding which paradigm built a system tells you exactly what it can and cannot do — and where to look when it fails.
What is the difference between supervised, unsupervised, and reinforcement learning — in terms of training signal?
Self-supervised learning is described as “the most important paradigm right now.” Why — and what does it not give you?
AlphaFold, Spotify Discover Weekly, and ChatGPT each use a different paradigm. Match them — and explain why each needed that paradigm.