Learn AI

The course

Fourteen modules over four parts, from what a trained model physically is to the law that applies to it. Modules are published one at a time, each checked against its sources on the date it shows.

Last reviewed 28 October 20261 of 14 modules published

Part 1: How the machine works

Module 1 · about 20 minutesWhat a model isThe small amount of mathematics the course uses, what a trained model physically consists of, and how weights, biases and layers turn input numbers into an output.Inside this module
  • The mathematics used in this course
  • A file of numbers and a recipe
  • The artificial neuron
  • Layers and depth
  • What a parameter is
Open the module →
Module 2 · in preparationHow a model learnsLoss, gradient descent and backpropagation, why models generalise and memorise at once, and how training compute became a regulatory yardstick.Inside this module
  • Measuring error
  • Walking downhill
  • Backpropagation
  • Learning versus memorising
  • Chips and the two phases
Module 3 · in preparationWords as numbersHow text is cut into tokens, why tokenisation causes odd failures, and how embeddings turn meaning into position.Inside this module
  • Cutting text into tokens
  • Where tokens trip models up
  • Embeddings: meaning as position
Module 4 · in preparationThe transformerAttention, feed-forward layers, the residual stream and next-token prediction, with the costs of long context.Inside this module
  • Where it came from
  • Attention
  • Feed-forward layers and stored facts
  • Depth, order and the running total
  • Predicting the next token
Module 5 · in preparationFrom predictor to assistantPretraining at scale, scaling laws, the post-training that turns a text predictor into an assistant, reasoning models, experts and distillation.Inside this module
  • Pretraining: reading the library
  • Scaling laws
  • Post-training: from document completer to assistant
  • Reasoning models
  • Experts and students
Module 6 · in preparationSampling, hallucination and memorisationHow a model chooses each word, why it invents things, and what the research shows about models reproducing their training data.Inside this module
  • Choosing the next token
  • Why models make things up
  • Memorisation

Part 2: Deployment, inspection and testing

Module 7 · in preparationThe system around the modelSystem prompts, retrieval, tools and agents, the new payment protocols for agents, and why prompt injection is so hard to fix.Inside this module
  • Model and system
  • Reading before answering
  • Tools and agents
  • Agents that pay
  • Prompt injection
Module 8 · in preparationOpen weights, images and provenanceFine-tuning and open-weight licences, how image and voice generators work, and the techniques for marking machine-made content.Inside this module
  • Customising a model
  • Open weights and licences
  • Images, video and voices
  • Marking what machines made
Module 9 · in preparationLooking inside and testingWhat interpretability research can explain so far, the legal duties to explain automated decisions, benchmarks, and the safety frameworks behind staged releases.Inside this module
  • Why the weights are hard to read
  • Duties to explain
  • Benchmarks and their limits
  • Safety frameworks and staged release

Part 3: Where the technology meets the law

Module 10 · in preparationTraining data and copyrightThe training copy, the weights and the output, and how US, UK, German and EU courts have treated each as of 1 October 2026, with a case note on each leading decision.Inside this module · EU and UK
  • The training copy, the weights and the output
  • United Kingdom
  • European Union and Germany
Module 11 · in preparationPersonal data in modelsWhether a model contains personal data, the lawful basis for training, enforcement so far, and why erasure and correction are hard.Inside this module · EU and UK
  • Does a model contain personal data?
  • Lawful basis and enforcement
  • Erasure, accuracy and unlearning
  • Automated decisions
Module 12 · in preparationRegulating the technologyThe EU AI Act after the Digital Omnibus, the Product Liability Directive, US federal and state law, and the UK and Asian approaches.Inside this module · EU and UK
  • The EU AI Act and its new timetable
  • Definitions and risk tiers
  • General-purpose models
  • Liability in the EU
  • The UK and beyond
Module 13 · in preparationLiability, agents and professional dutyWho answers for what a model says, how the law of electronic agents meets language model agents and agentic payments, and the professional duty after Ayinde.Inside this module · EU and UK
  • Who answers for what a model says
  • Agents and the law of agreements
  • Financial regulators and agents
  • Professional duty

Part 4: The road ahead

Module 14 · in preparationBuilding a leading AI practiceThe direction of AI legal work, the under-served intersection with digital assets, a twelve-month plan, hands-on exercises and a reading routine.Inside this module
  • What the work looks like
  • Where the opening is
  • Exercises
  • A twelve-month plan
  • Staying current