Level 3
Write a design someone could actually build, while respecting the constraints you are given.
Who can apply
Open to applicants aged 18 and over.
Required programming language
Required at every level, from Level 5 to Level 1.
Shell Scripting
The single most important language for working with AI.
The is useful for running , managing files, and automating development. can also run through , apps, notebooks, and . A is one useful interface, not a requirement for every AI user.
Learn shell programmingThe one question, at every level
You want to build an artificial intelligence completely from scratch. Design how you would build it.
Every level from 5 to 1 answers this same question. What changes is the depth expected of the answer, and the constraints attached to it.
What to learn
- Design a structure that fits the given platform and language constraints.
- Explain and as two separate processes.
- State the limits and risks of your own design.
How you are marked
- Did you stay within the constraints you were given?
- Could the structure you describe actually run?
- Did you state the limits honestly instead of hiding them?
Exam format
Stage 1 is a written paper at the venue; Stage 2 is an interview followed by a practical build. Both stages use the same question.
Held at a designated exam venue. You write your answers on the exam paper by hand. This stage tests the design you can put into words.
An examiner interviews you about your design, then you build the for real. This stage tests whether the plan on paper actually works in your hands.
Practice questions
The official exam gives you just the one common question above. The 20 questions below are practice questions that check the words, names and facts you need to know before you write that design. Every one is a short-answer question with a single correct answer: read the description, write your answer, then open the answer to check. They are not official questions.
How training works
-
1
What do you call, all together, the values a model adjusts by itself while it learns from data, such as the weights and biases of a neural network?
Show answer
Answer Parameters
In "a 70B model", B stands for billion: the model has 70 billion parameters.
-
2
What do you call settings such as the learning rate or the number of layers, which training does not decide by itself but people set before training?
Show answer
Answer Hyperparameters
Training sets the parameters; people set the hyperparameters, comparing choices on validation data.
-
3
What is the optimization method that changes the weights little by little in the direction that lowers the loss (opposite to the gradient)?
Show answer
Answer Gradient descent
It is like walking down a mountain blindfolded, always stepping toward the lower ground under your feet.
-
4
What is the method that passes the error at the network's output backwards, from the last layers to the first, to compute the gradient of every weight?
Show answer
Answer Backpropagation
Backpropagation finds the gradients; gradient descent then changes the weights by those gradients.
-
5
In gradient descent, what is the hyperparameter that decides how big each change to the weights is?
Show answer
Answer Learning rate
Too large and the loss jumps up and down; too small and training is very slow.
Inside a language model
-
6
In a transformer, what is the core mechanism that works out, for each token in a sentence, how much weight to give each of the other tokens?
Show answer
Answer Attention (self-attention)
The paper that first presented the transformer is titled "Attention Is All You Need".
-
7
What do you call the maximum number of tokens a language model can read and keep in mind at once?
Show answer
Answer Context window
The model cannot see anything earlier than this length. A design that handles long documents must consider it first.
-
8
Which generation setting makes the model always pick the most likely next token when it is low, and pick more varied tokens when it is high?
Show answer
Answer Temperature
It is often set low for factual answers and high for brainstorming.
-
9
What do you call training an already-trained model a bit more on a small set of data for a specific purpose?
Show answer
Answer Fine-tuning
It needs far less data and computation than training from scratch.
-
10
Write the three-letter abbreviation for the method where the model first looks up related documents, puts their contents into the prompt, and then answers.
Show answer
Answer RAG (Retrieval-Augmented Generation)
It lets a model answer from up-to-date or in-house documents without retraining it.
Platforms and constraints
-
11
What do you call running a model directly on a device such as a phone or laptop, without going through an internet server?
Show answer
Answer On-device AI
Data never leaves the device and it works offline, but the model must be small enough for the device's memory and computing power.
-
12
What is the method that makes a model smaller and lighter on memory by storing its numbers (weights) in fewer bits, such as 8 or 4 bits instead of 32?
Show answer
Answer Quantization
It may cost a little accuracy, but it lets a large model run on a small device.
-
13
Write the name of the binary format that lets programs written in languages like C or Rust run fast inside a web browser. This site's Python practice also runs on it.
Show answer
Answer WebAssembly (Wasm)
If the platform constraint is "the browser", you can design around running Python and the model inside the browser with WebAssembly.
-
14
Write the three-letter abbreviation for the chip that was made for graphics but handles the huge matrix calculations of neural network training many at a time.
Show answer
Answer GPU (Graphics Processing Unit)
It runs thousands of small calculators at once, which suits deep learning, where the same calculation repeats.
-
15
Write the three-letter abbreviation for the gateway a program opens so other programs can use its features through agreed rules (request and response formats). You use one to call a language model from your app.
Show answer
Answer API (Application Programming Interface)
Borrowing someone else's model through an API is easy, but you must also weigh cost, speed and data leaks.
Limits and risks
-
16
What do you call the problem where training data is lopsided, so the model gives unfair results to certain groups?
Show answer
Answer Bias (data bias)
For example, a face recognition model trained on few photos of one group makes more mistakes on that group.
-
17
What do you call the attack that hides secret instructions inside a prompt or a document the model reads, so the language model breaks its original instructions?
Show answer
Answer Prompt injection
If your design reads outside documents, it must stop the model from obeying instructions found inside them.
-
18
What is the word, made from deep learning and fake, for fake video or audio that imitates a real person's face or voice so it looks real?
Show answer
Answer Deepfake
The same technology can become a crime depending on its purpose. Never use someone's face or voice without consent.
-
19
Write the three-letter abbreviation for the field of research that tries to explain, in a way people can understand, why a model gave its answer.
Show answer
Answer XAI (Explainable AI)
It matters especially in fields like medicine and finance, where the reason behind a decision is important.
-
20
Who made the world's first comprehensive AI law, which came into force in August 2024 and regulates AI according to its level of risk?
Show answer
Answer The European Union (EU)
It is called the EU AI Act. The riskier the use, the heavier the duties, and its rules apply in stages.