⚡ Rapid revision — the 20-minute version
| Spine | The mark |
|---|---|
| Narrow vs general | Everything today is narrow — impressive surface, one domain |
| ML paradigms | Labels? supervised. No labels? unsupervised. Rewards? reinforcement |
| The pipeline | Collect → clean → train → evaluate on UNSEEN data → deploy → monitor |
| Overfitting | Train score high, test low — memorised the noise, not the pattern |
| Ethics | Bias flows from data; confident output is not true output — verify |
| Emerging tech | Blockchain = tamper-EVIDENT; quantum = experimental; IoT's weak link = device security |
1. What AI is — and what it is not
Artificial intelligence: AI = systems performing tasks that normally need HUMAN intelligence: perceiving, reasoning, learning, deciding, language. NARROW AI does ONE domain well (spam filters, chess engines, translation) — every AI in daily use today. GENERAL AI would match human ability ACROSS domains — it does not exist. The exam's trap: today's systems are narrow, however impressive; 'it is just statistics' is also wrong — it is pattern learning at scale.
Worked: a face-unlock system recognises faces (narrow, vision only); it cannot write an essay or drive a car. Chatbots sound general but predict text — narrow, with a wide surface. 'State whether this is narrow or general AI and why' — the domain-bound answer is the mark.
2. Machine learning — learn patterns from data
Machine learning: Classical programming: RULES + DATA → ANSWERS. Machine learning: DATA + ANSWERS → RULES (the model). Paradigms: SUPERVISED (labelled examples — classification: spam or not; regression: a price), UNSUPERVISED (no labels — find structure: customer CLUSTERS), REINFORCEMENT (actions with REWARDS — game agents, robotics). The paradigm question: 'does the training data carry labels?' decides everything.
Worked: 10,000 labelled emails (spam/not-spam) → supervised classification. Unlabelled purchase histories grouped into segments → unsupervised clustering. A robot rewarded for staying upright → reinforcement. Name the paradigm from the data description — labels present or not is the fork.
3. The ML pipeline — data in, decisions out
Pipeline: COLLECT data → CLEAN it (missing, wrong, duplicated values) → engineer FEATURES (the inputs the model uses) → TRAIN (fit the model on a TRAINING split) → EVALUATE (score it on a HELD-OUT TEST split it never saw) → DEPLOY → MONITOR (retrain when the world drifts). Data quality dominates: garbage in, garbage out — the model can only be as good as its data.
Worked: predicting exam scores from study hours — 80% of rows train the model, 20% held back for testing; testing on TRAINING data inflates the score (it memorised) — the held-out split is the honesty check. 'Why split the data?' — to measure performance on UNSEEN examples; that is the mark.
4. Overfitting — memorising is not learning
Overfitting: OVERFITTING: the model learns the training data's NOISE — perfect on training examples, poor on new ones (memorised, not generalised). UNDERFITTING: too simple to capture the pattern at all — poor everywhere. Fixes for overfitting: more data, simpler model, regularisation, early stopping. The symptom pair (high train score, low test score) is the exam's signature.
Worked: 98% accuracy on training data, 61% on the test set — overfitted: it memorised the training set's quirks. 55% on both — underfitted: the model is too crude. Diagnose from the TWO numbers; one number alone cannot distinguish the failure modes.
5. Neural networks — layers, weights, learning
A neural network layers simple units: each connection has a WEIGHT, each unit sums its weighted inputs and applies an ACTIVATION. Training adjusts the weights to reduce error (backpropagation + gradient descent — 'nudge weights downhill'). DEEP learning = many layers, excelling at vision, speech and language; the cost is data and compute hunger, and the model's reasoning is hard to inspect (black box).
Worked: handwriting recognition — the first layers detect strokes, deeper layers assemble shapes into letters. No layer is 'programmed' — weights start random and are tuned by examples. 'Explain how the network learns' = adjust weights to reduce error, repeatedly — the mechanics in one line is the mark.
6. Ethics and bias — the model inherits the data
| Issue | Cause | Exam response |
|---|---|---|
| Bias | Unrepresentative or historical data | The model repeats the data's unfairness — audit and diversify data |
| Privacy | Training on personal data | Consent, anonymisation, data minimisation |
| Accountability | Opaque decisions | Explainability, human oversight for high-stakes calls |
| Hallucination | Generative models predict plausible text | Verify outputs; confidence is not truth |
The core principle: a model is only as fair as its data. A hiring model trained on a biased history learns the bias — technology does not neutralise it automatically. High-stakes decisions (loans, hiring, medicine) need HUMAN oversight; generative models state falsehoods fluently — verification is never optional.
7. Emerging tech — IoT, cloud, blockchain, quantum
| Technology | What it is | Honest maturity |
|---|---|---|
| IoT | Everyday objects with sensors + connectivity, streaming data | Live — smart meters, trackers; security is the weak link |
| Cloud | Computing rented on demand (storage, servers, AI services) | Live — pay per use, scales; dependency and privacy trade-offs |
| Blockchain | Append-only distributed ledger, consensus instead of a central party | Live for crypto; enterprise uses narrower than the hype |
| Quantum computing | Qubits exploit superposition for certain problems | Experimental — real for research, not yet everyday |
Worked: a smart farm — soil sensors (IoT) stream to the cloud, models schedule irrigation. The exam wants the DEFINITION matched to the application AND the risk named: IoT devices are notoriously insecure endpoints; cloud shifts control to a provider; blockchain's guarantee is tamper-EVIDENCE, not truth of what was recorded.
8. Agents and LLMs — prediction, not understanding
A LARGE LANGUAGE MODEL predicts the next TOKEN from a prompt, trained on vast text — fluency, not comprehension. AGENTS wrap an LLM with tools and goals: plan → act (search, code, call APIs) → observe → repeat. Limits: hallucination (confident falsehoods), knowledge cut-offs, prompt injection (malicious instructions smuggled in data). The discipline: verify outputs, keep humans on the loop for consequential actions.
Worked: an agent asked to 'book the cheapest flight' plans, queries an API, compares, acts — every consequential step needs a guardrail because the model optimises the LETTER of the goal. 'How does an LLM differ from a search engine?' — it generates text from patterns; it does not look up documents (though agents can). The prediction framing is the mark.
9. Summary — the AI & emerging tech spine
| Question type | Method beat | Trap |
|---|---|---|
| Classify the AI | Narrow (all today) vs general (none) | Impressive ≠ general |
| Pick the paradigm | Labels? supervised; no labels? unsupervised; rewards? reinforcement | Regression predicts quantities, classification categories |
| Explain the split | Test on UNSEEN data | Testing on training data inflates the score |
| Diagnose the model | Train score vs test score together | Overfit: memorised noise; underfit: too simple |
| Ethics | Bias flows from data; oversight for stakes | Confident output is not true output |
| Emerging tech | Definition + application + risk | Blockchain is tamper-evident, not truthful |
Command discipline: 'explain' = mechanism in one causal line; 'discuss the ethical concerns' = issue + cause + mitigation, not a slogan list; 'justify' = the evidence from the scenario (labels, scores, stakes). AI papers pay for the REASONING from data to conclusion — buzzwords without the mechanism earn nothing.