Lessons

47 free AI lessons, from AI fundamentals and the maths of machine learning to transformers, LLMs, agents and robotics, each with live demos and a quiz.

Lessons

• AI Fundamentals

  What AI is, where it came from, and how today’s systems actually work under the hood.

• Machine Learning

  How computers learn patterns from examples instead of following hand-written rules.

• Logistic Regression

  How a weighted sum and a sigmoid turn inputs into a probability: log loss, decision boundaries, gradient descent and thresholds.

• k-Nearest Neighbours

  Classify by asking the most similar known examples: distance, choosing k, decision regions, feature scaling and the curse of dimensionality.

• Evaluating Models

  Accuracy lies. Learn the metrics and habits that tell you whether a model really works.

• Overfitting and Regularisation

  Why models memorise noise, and how ridge, lasso, early stopping, dropout and cross-validation make them generalise.

• Neural Networks and Deep Learning

  From a single neuron to deep networks: activations, backpropagation and gradient descent.

• Backpropagation

  Computational graphs, the chain rule and the backward pass, step by step with real numbers, plus gradient checks and vanishing gradients.

• Optimisers: SGD, Momentum and Adam

  How networks move their weights: learning rates, mini-batch noise, momentum, Adam, and the schedules that keep training stable.

• Unsupervised Learning

  Finding structure in data nobody labelled: clustering, density and dimensionality reduction.

• AI Ethics and Safety

  Bias, fairness, transparency and alignment, measured on real trade-offs rather than slogans.

• Computer Vision

  How machines turn pixels into meaning, from edge filters to convolutional networks.

• Convolutional Neural Networks

  Kernels, stride, padding, channels, pooling and receptive fields, computed on images you draw, and what each CNN layer learns.

• Object Detection

  Finding and labelling every object in a scene, from R-CNN to YOLO.

• Image Segmentation

  Labelling images pixel by pixel: semantic, instance and panoptic segmentation.

• Natural Language Processing

  Teaching machines to read: tokens, embeddings, sentiment and the road to language models.

• Machine Translation

  How translation went from phrase tables to neural encoder-decoder models.

• Transformers and Attention

  The architecture behind modern AI: self-attention, multi-head attention and positional encoding.

• Generative AI

  How machines learn to create text, images, audio and video: GANs, VAEs, autoregressive models and diffusion.

• Search and Optimization

  How AI plans: graph search, heuristics, A* and local search for problems too big to brute-force.

• Reinforcement Learning

  Learning by trial and error: rewards, policies, Q-learning and the exploration trade-off.

• Large Language Models

  How models like Claude and GPT are built: pretraining, scaling laws, fine-tuning, RLHF and reasoning.

• Embeddings and Retrieval

  Turning meaning into vectors, searching by similarity, and grounding LLM answers in your own documents.

• AI Agents and Tool Use

  How language models plan, call tools and act in loops, and why reliability is the hard part.

• Decision Models and Jev

  Why a model that answers with a calibrated choice instead of text is one of 2026’s biggest stories.

• Multimodal AI

  Models that see, read and listen at once: CLIP, zero-shot classification and vision-language models.

• The Maths Behind ML

  The handful of ideas from linear algebra, calculus and probability that every model is built on, made visual.

• Forests and Boosting

  Why many weak trees beat one strong one, and why gradient boosting still wins on spreadsheets.

• Support Vector Machines

  Maximum margins, support vectors, soft margins and the kernel trick, with a real SVM that refits as you drag points.

• Naive Bayes

  Bayes’ rule, the independence assumption, Laplace smoothing and log probabilities, with a word-count spam filter you can type into.

• Recommender Systems

  How Netflix, Spotify and TikTok decide what you see next, from collaborative filtering to two-tower models.

• Robots and World Models

  How AI leaves the screen: perception, control, learning from demonstration and the vision-language-action models of 2025-26.

• Speech and Audio AI

  How machines hear and speak: spectrograms, speech recognition, voice synthesis and real-time voice assistants.

• Prompt and Context Engineering

  Getting reliable work out of language models: instructions, examples, structure, context and evaluation.

• Fine-Tuning and Efficient Models

  Adapting and shrinking big models: fine-tuning, LoRA, quantization and distillation.

• Interpretability

  Looking inside neural networks: features, circuits, probes and sparse autoencoders.

• Opus 5.5, Fable 5.1 and Looped Transformers

  How a cheaper model can match a bigger one: extra depth by looping layers versus extra size, and what is actually known about Claude Opus 5.5 and Fable 5.1.

• Tokenization

  How language models chop text into tokens, why it explains their strangest mistakes, and what it costs you per word.

• Scaling Laws and Pretraining

  Why bigger models trained on more data get predictably better, how labs plan billion-dollar training runs, and where the curves bend.

• Post-Training and Alignment

  How a raw text predictor becomes a helpful assistant: instruction tuning, RLHF, DPO, constitutional AI and rewards you can verify.

• Reasoning Models

  Why models that think before answering got so much better at maths and code: chain of thought, test-time compute and learning from verifiable rewards.

• Mixture of Experts

  How frontier models get huge without getting slow: routing each token to a few specialist sub-networks.

• GPT-6.1 Sol, Gemini 4 Argon and DeepSeek V4.1 Flash

  September 2026’s three big model releases, what is actually new in each, and the ideas behind them: sparse experts, KV-cache compression and million-token outputs.

• Claude Haiku 5.5, Muse Wonder and the price of a token

  Claude Haiku 5.5 and Meta’s Muse Wonder explained: a new tokenizer, the 100K price line, prompt caching and what an agent run costs.

• Muse Spark Open, parallel agents and a frontier model on your desk

  Meta’s 816B open-weight Muse Spark Open and ChatGPT’s parallel agents explained: active parameters, memory, quantization and speculative decoding.

• GPT-6.1 Terra, Ultrafast and agents that use a computer

  OpenAI’s GPT-6.1 Terra and Ultrafast explained: tokens per second, why speed costs more, computer-use agents and how errors compound.

• GPT-6 Rosalind, robot brains and AI that leaves the chat

  OpenAI’s GPT-6 Rosalind and Gemini Robotics on Atlas explained: DNA as tokens, robot control loops, step reliability and science benchmarks.