Artificial Intelligence, Data, the Past & the Future

Module 1 · 6 lessons

The past, present and future

The past provides the data from which we learn, the present is where we observe, decide and act, and the future remains uncertain. This module introduces the central framework of the course: how observations of the world become data, how models turn that data into predictions, and how decisions made from those predictions help shape what happens next. It explores time, space, uncertainty and abstraction, establishing a crucial principle for everything that follows: models and predictions are representations of possible realities, not reality itself.

Module 2 · 13 lessons

Data: turning the world into information

Data is how we turn the world into something that can be stored, compared, analysed and used by models. This module explores what data actually is, how observations become variables, features and labels, and how different forms of data represent time, space, images, sound and language. It also examines the choices hidden inside every dataset: what was measured, who or what was included, when and where observations were taken, how accurately they were recorded, and what was left out. Students will learn why sampling, missing data, measurement error and the difference between correlation and causation matter, establishing a central principle for the rest of the course: data is not neutral raw material, but the product of a measurement and selection process that shapes everything an AI system can learn from it.

Module 3 · 13 lessons

Uncertainty, probability and inference

Uncertainty is unavoidable whenever we reason about events we cannot observe directly, especially the future. This module introduces probability as a language for representing that uncertainty, covering random variables, distributions, expected values, variance, conditional probability and Bayesian reasoning without turning the course into a statistics degree. It explores how prior beliefs change when new evidence arrives, how confidence should be calibrated, why rare events matter, and the difference between measurable risk and deeper uncertainty. The central idea is simple: a good model should not merely tell us what it thinks will happen, but also help us understand how certain it is — and where that confidence may be misplaced.

Module 4 · 16 lessons

Prediction: learning patterns from the past

Prediction is where data becomes a tool for reasoning about the unknown. This module introduces the foundations of machine learning by asking a simple question: given what has happened before, what should we expect to happen next? Students will explore regression, classification, parameters, features and how models learn from examples, before examining loss functions, gradient descent, overfitting, underfitting and generalisation. Particular attention is given to how models are trained, validated and tested, and to the choices hidden inside that process. At its heart is a fundamental idea: before a machine can learn, somebody must define what counts as being wrong. The objective we choose therefore shapes what the machine learns — connecting the mathematics of prediction to much deeper questions about optimisation, values, bias and the systems we ultimately design.

Module 5 · 14 lessons

When is a prediction good?

A prediction is only useful if we understand what it means for it to be wrong. This module explores how predictions are evaluated, moving beyond simple accuracy to measures such as absolute and squared error, precision, recall, confusion matrices, calibration and probabilistic scoring. Students will examine forecasting across time and space, distribution shift, and what happens when the future stops resembling the data on which a model learned. Crucially, different errors have different consequences: incorrectly predicting rain is not equivalent to missing a cancer diagnosis, wrongly denying someone credit, or failing to detect a pedestrian in front of an autonomous vehicle. Evaluating AI therefore requires more than measuring predictive performance — it requires deciding which errors matter, to whom they matter, and how their consequences should shape the systems we build.

Module 6 · 8 lessons

State, dynamics and decision-making

Prediction tells us what might happen; it does not tell us what to do. This module introduces state, dynamics and decision-making, beginning with the idea that an intelligent system needs a representation of where it believes the world is now before it can reason about what happens next. Students will explore state vectors, observations versus hidden state, state estimation, dynamics, actions, disturbances and feedback, before moving from prediction into objectives, constraints, optimisation, scarcity and the allocation of shared resources. Sequential decisions reveal how actions change the state from which the next decision must be made, providing the foundations for control, model predictive control, reinforcement learning, agents and robotics. The central framework becomes: the past provides evidence; state represents where we believe we are now; models describe how the world can change; predictions describe possible futures; and decisions help determine which future we actually reach.

Module 7

Designing intelligent services for shared resources

Intelligence becomes much more consequential when many people compete for resources that cannot serve everyone simultaneously. This module explores how intelligent services allocate scarce resources such as electricity, road space, healthcare, computing capacity, water, transport and housing. Students will examine capacity, queues, priorities, prices, auctions, rationing, optimisation, fairness, service levels, reliability and quality of service, alongside the trade-offs between centralised and distributed decision-making. It also introduces reflexive systems, where predictions and decisions change the very system being predicted: navigation algorithms redirect traffic, electricity prices change demand, trading models influence markets, and recommendation systems shape the behaviour that becomes their future training data. The central question is therefore not simply whether AI can predict what people will need, but how intelligent systems should use those predictions to coordinate shared resources fairly, efficiently and reliably — while continuously adapting to the consequences of their own decisions.

Lessons coming soon.

Module 8

From machine learning to modern AI

Artificial intelligence did not begin with ChatGPT. This module traces the development of AI from early attempts to encode intelligence through rules and symbolic reasoning, through expert systems and machine learning, to neural networks, deep learning, transformers, foundation models and modern generative AI. Students will explore the ideas that enabled breakthroughs in computer vision, speech and language, including backpropagation, embeddings and attention, while connecting them to the principles of data, probability and prediction developed earlier in the course. Particular attention is given to demystifying large language models: at their core, they repeatedly use context to estimate a probability distribution over possible next tokens and select what comes next. The remarkable story of modern AI is how training increasingly large models on increasingly large datasets, supported by enormous computational resources, allows extraordinarily rich capabilities to emerge from this deceptively simple predictive process.

Lessons coming soon.

Module 9

How machines learn language, images and the world

Modern AI learns representations of language, images and increasingly the wider world, then uses those representations to predict, generate, reason and act. This module looks inside contemporary AI systems, exploring tokens, embeddings, attention, transformers, pre-training, fine-tuning and reinforcement learning with human feedback before moving into multimodal and mixture-of-experts models, reasoning systems, tool use, memory, agents and world models. Students will examine how machines learn useful internal representations from enormous amounts of data and how increasingly capable systems combine those representations with context, computation and feedback. The module introduces an important progression: from models that predict, to models that generate, to systems that reason, use tools and maintain memory, and ultimately to agents and autonomous systems that can observe, decide and act. At each step, AI closes more of the loop between information and action — bringing us closer to machines that do not merely model possible futures, but can actively influence which futures occur.

Lessons coming soon.

Module 10

The physical machine behind artificial intelligence

Artificial intelligence may appear to live in the cloud, but every prediction, generated image and conversation ultimately runs on physical machines consuming physical resources. This module explores the infrastructure beneath modern AI, from transistors, integrated circuits, CPUs, GPUs and specialised AI accelerators to memory, parallel computation, semiconductor fabrication and global chip supply chains. It then moves inside the data centre, examining servers, networks, cooling, electricity, water, backup power, land and connectivity, before considering the rapidly growing resources required for AI training and inference. Students will explore power density, grid connections, geographic concentration, semiconductor capacity, critical minerals and the physical constraints that may shape the development of AI. The module ultimately returns to a familiar problem of scarcity and allocation: if machine intelligence becomes increasingly abundant in capability but remains dependent on scarce energy, compute, infrastructure and materials, who gets access to it — and how should those resources be allocated?

Lessons coming soon.

Module 11

AI, society, power and fairness

Artificial intelligence does not operate outside society: it learns from data produced by society and increasingly makes decisions that shape people's opportunities, resources and experiences. This module examines where algorithmic bias comes from, including sampling, measurement, labels, historical inequality and proxy variables, and explores how automated decisions can reinforce those patterns through feedback loops. Students will examine fairness metrics, individual and group fairness, trade-offs between accuracy and fairness, explainability, transparency, accountability and human oversight across areas such as hiring, credit, insurance, healthcare, policing, education and housing. Particular attention is given to why simply removing protected characteristics does not necessarily remove discrimination, since location, income, language, education and behaviour may encode much of the same information indirectly. The module ultimately confronts a deeper question: what should a fair algorithm optimise? Mathematics can help us measure consequences and formalise different definitions of fairness, but it cannot determine which definition society ought to choose. Designing fair AI is therefore not only a technical problem, but a problem of service design, governance, power and collective choice.

Lessons coming soon.

Module 12

Intelligence, robots and the future

Artificial intelligence becomes most consequential when it moves from predicting the world to acting within it. This final module brings the course together through robotics and autonomous systems, exploring how sensors, perception, localisation, mapping, planning, control and reinforcement learning allow machines to close the loop from observation to physical action. It then looks towards increasingly general and autonomous forms of intelligence, examining AI agents, artificial general intelligence, emergent capabilities, alignment, control, superintelligence and arguments surrounding catastrophic and existential risk. Alongside these longer-term possibilities, students will examine the much more immediate economic and social consequences of abundant machine intelligence: automation, productivity, employment, inequality, ownership of models, data and compute, education, creativity, misinformation, surveillance, privacy, warfare and political power. The course ultimately returns to its central systems question: not simply what will AI do, but what future are we trying to create? As AI progresses from prediction to decision, agency and autonomy, society must determine the objectives, constraints, rights and rules within which increasingly intelligent systems operate. The quality of the future will depend not only on how intelligent our machines become, but on how well we design the systems in which that intelligence is used.

Lessons coming soon.