What do you want to be able to do?

Understand what a model is, how it learns from examples, why it makes mistakes — and how to use AI responsibly at school.

Investigator · 9 knowledge nodes

How does a computer recognise a cat, understand your voice and create images from text? Build and test it yourself — and learn where it goes wrong.

Builder · 25 knowledge nodes

Take the step from block programming to text: programs that talk to the user, draw, use randomness, and get planned before they are written.

Builder · 21 knowledge nodes

Where AI turns up in everyday life, how it changes work, what responsible use means, and what disinformation, bias and manipulation look like.

Builder · 58 knowledge nodes

From vectors and neural networks to attention and transformers — you should be able to explain what happens inside a language model as it writes the next word.

AI developer · 35 knowledge nodes

Build and train a small neural network in PyTorch, and understand every step: forward pass, loss, gradients and updating the weights.

AI developer · 30 knowledge nodes

NumPy, the terminal, regular expressions, modules and virtual environments, notebooks, debugging, code style and typing — the tools you use every day in an AI project.

AI developer · 32 knowledge nodes

Logistic regression, trees, kNN and clustering — with cross-validation, the right metrics and bias–variance as the frame. Often better than a neural network on tabular data.

AI developer · 43 knowledge nodes

From observation to table, via strange values, licences and text preprocessing, to a documented dataset without leakage.

AI developer · 40 knowledge nodes

How the internet fits together, how data is compressed, why sorting pays off, and what big-O notation means in practice.

AI developer · 16 knowledge nodes

The summation symbol, sets and logic, quadratic and exponential functions, logarithms and softmax — the notation and the functions every machine-learning text assumes you already know.

AI developer · 16 knowledge nodes

Implement self-attention, multi-head attention and positional encoding yourself, and put them together into a working small transformer.

University · 49 knowledge nodes

University level. From statistical learning and transformers to a served text system with evals, tests and documentation.

University · 47 knowledge nodes

Build a CNN, understand convolution and receptive fields, and look inside the network to see what it has actually learned.

University · 48 knowledge nodes

From bandits to policy gradient: build an environment, let an agent learn from reward, and understand why RL is hard.

University · 31 knowledge nodes

Frame the problem correctly, take the prototype to a product with tests and evals, build in degraded modes, and defend against prompt injection.

University · 51 knowledge nodes

Partial derivatives, gradients, integrals, linear maps, convexity and graphs — what is actually used when a model is trained.

University · 39 knowledge nodes

Optimisers, regularisation, normalisation, initialisation, dataloaders and transfer learning — everything that separates a training run that works from one that does not.

University · 86 knowledge nodes

Conditional probability, simulation, A/B testing and variance between runs — so that you can tell a real improvement from chance.

University · 69 knowledge nodes

Chunking, retrieval evaluation, citation and grounding — the whole chain from document to answer with traceable sources.

University · 77 knowledge nodes

Tokenisation, sampling, instruction models, structured output, hallucinations and Swedish — what you need in order to build with an LLM for real.

University · 70 knowledge nodes

Images as matrices, filters that find edges, a trained image classifier, and a model that tells two sounds apart.

University · 32 knowledge nodes

Observability, prompts as code, agent loops and tool use — how an AI feature is run once it actually has users.

University · 87 knowledge nodes

Processes, memory, parallelism, containers, cryptography and code review — the computer science underneath the model, which decides whether the system can be run and trusted.

University · 43 knowledge nodes

Margins, outliers, recommendations and systematic hyperparameter search — the methods that still solve most problems more cheaply than a neural network.

University · 64 knowledge nodes

Model cards, copyright in training data and the ethics of generative images — what has to be in place before an AI system meets real users.

University · 54 knowledge nodes

Understand why low-rank adapters work, choose target modules and rank, build a dataset, and evaluate the result with evals.

AI engineering · 36 knowledge nodes

Episodic and semantic memory, retrieval and long-horizon evaluation — a working memory system that makes an agent better over time.

AI engineering · 68 knowledge nodes

Shrink a model to a quarter of its size, measure what is lost, and choose the right format for your runtime.

AI engineering · 59 knowledge nodes

PEFT and QLoRA, dataset design, distributed training, catastrophic forgetting, export, and a capstone where you fine-tune and evaluate for real.

AI engineering · 89 knowledge nodes

Architecture, planning, observability, safety and evaluation — and the coding agent that gets objective feedback from tests.

AI engineering · 77 knowledge nodes

LLM as judge, regression tests, statistical significance and how to value negative results — measurement that holds between releases.

AI engineering · 62 knowledge nodes

Red teaming, jailbreaks, adversarial examples, preference learning and reward hacking — what can be defended against and what cannot.

AI engineering · 65 knowledge nodes

Diffusion, GANs and how you actually evaluate a model that creates content — including checking for memorisation.

AI engineering · 63 knowledge nodes

Serving, scaling, distributed systems, operations and calibrated uncertainty — plus a project where you deliver a service that holds up.

AI engineering · 80 knowledge nodes

Attention patterns, the logit lens and explanations for end users — how to see what a model is doing, and where the limits of what an explanation proves lie.

AI engineering · 66 knowledge nodes

From waveform to spectrogram and mel features, on to speech recognition and speech synthesis — and the privacy questions that voice data always brings with it.

AI engineering · 46 knowledge nodes

Policy gradient and REINFORCE, DQN with a replay buffer and a target network, actor–critic and PPO — the methods behind everything from Atari to RLHF, and the faults that stop them learning.

AI engineering · 52 knowledge nodes

Read a research paper, state the hypothesis, implement the method, run ablations, and compare with the published results.

Frontier Lab · 43 knowledge nodes

Read and reproduce a paper, run ablations, inspect the model mechanistically, and write a report someone else can reproduce.

Frontier Lab · 87 knowledge nodes

Shared vector spaces, vision–language models, document understanding, multimodal retrieval, video, and agents that see the screen — with hallucination and safety measurements that hold.

Frontier Lab · 49 knowledge nodes