Level 1
Complete a specialised from the ground up, and leave it documented so another person can rebuild it.
Who can apply
Open one year after obtaining the Level 2 certificate.
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. AI 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
- Work the demands of a specialist field — mathematical computation, music composition — into the design.
- Build the core yourself rather than leaning on ready-made tools.
- Write documentation someone else could carry on from.
How you are marked
- Does it genuinely meet the demands of the specialist field?
- Did you build the core yourself, from scratch?
- Could another person reproduce it from your documents alone?
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.
From scratch — the math
-
1
What is the function that turns a neural network's list of scores into probabilities that are each between 0 and 1 and add up to 1?
Show answer
Answer Softmax
A language model uses softmax at the end to give a probability for each candidate next token.
-
2
In a classification model, what is the loss function equal to the negative log of the probability given to the correct answer, which grows larger the lower that probability is?
Show answer
Answer Cross-entropy
If the correct answer gets probability 1, the loss is 0; as it nears 0, the loss becomes huge.
-
3
What is the differentiation rule that finds the gradient of a composite function by multiplying the gradients of the outer and inner functions — the mathematical basis of backpropagation?
Show answer
Answer Chain rule
A neural network is layers composed like functions, so the chain rule is applied layer after layer to get the gradients.
-
4
In a transformer, what do you call the values added to the embeddings to give the model information about the order of the tokens?
Show answer
Answer Positional encoding
Attention alone cannot tell the order of tokens, so a number for each position is added.
-
5
What is the method against overfitting that randomly switches off some neurons during training, so the model does not rely too much on any one neuron?
Show answer
Answer Dropout
Neurons are switched off only during training; at inference time all of them are used.
Specialty — math computation
-
6
What do you call the error that appears because decimals are stored in a limited number of bits, so that 0.1 + 0.2 does not come out as exactly 0.3?
Show answer
Answer Floating-point error (rounding error)
An AI specialized in math must design its order of calculation and data types so these errors do not pile up.
-
7
Write the name (number) of the international standard that sets how computers store real numbers as a sign, an exponent and a significand.
Show answer
Answer IEEE 754
Most computers and programming languages store real numbers according to this standard.
-
8
What do you call letting a language model call an outside function, such as a calculator or code execution, to get an exact value instead of doing the arithmetic in its head?
Show answer
Answer Tool calling (function calling)
A language model only predicts the next token, so it easily gets big calculations wrong. Hand calculations that must be exact to a tool.
-
9
What do you call the matrix you get by swapping the rows and columns of a matrix (written Aᵀ)?
Show answer
Answer Transpose (transposed matrix)
Transposing an m × n matrix gives an n × m matrix. In neural network calculations it is often used to make the shapes fit for matrix multiplication.
-
10
What is the prompting technique that improves a language model's reasoning and calculation accuracy by having it write out its working step by step?
Show answer
Answer Chain-of-thought (CoT)
Writing the steps makes the model keep intermediate results as tokens while it computes, so it makes fewer mistakes.
Specialty — music
-
11
Write the abbreviation for the electronic instrument communication standard from 1983 that records performance information such as pitch, length and loudness instead of the sound itself.
Show answer
Answer MIDI (Musical Instrument Digital Interface)
Music AI sometimes learns from note data such as MIDI instead of sound waves.
-
12
What do you call the number of times per second a sound is measured and recorded as a number? For a CD it is 44,100 Hz.
Show answer
Answer Sampling rate
Sounds higher than half the sampling rate cannot be recorded (the Nyquist frequency).
-
13
Write the three-letter abbreviation for the fast method of the transform that splits a sound signal into frequency components, showing how much of each pitch is mixed in.
Show answer
Answer FFT (Fast Fourier Transform)
It is an algorithm that computes the Fourier transform quickly. It is used in almost every sound AI preprocessing step.
-
14
What do you call the picture of how the strength of each frequency changes over time, often used as the input to sound AI?
Show answer
Answer Spectrogram
Time runs along the horizontal axis, frequency up the vertical axis, and brightness shows strength. It lets you handle sound like an image.
-
15
In standard Western pitch, what is the frequency in Hz of the reference note A4 (the A above middle C)?
Show answer
Answer 440 Hz
Going up one octave doubles the frequency, so A5 is 880 Hz.
Reproducibility and documents
-
16
What do you call the starting value fixed in advance so that training that uses random numbers gives the same result when it is run again?
Show answer
Answer Seed (random seed)
Only if you write the seed in your document can others run the same experiment again.
-
17
What do you call the document released with a model that sets out its intended uses, training data, performance and limits?
Show answer
Answer Model card
A good model card also lists the uses to avoid and the known limits.
-
18
Write the name of the distributed version control tool that records the history of code changes so anyone can get the exact same version again. This site teaches it too.
Show answer
Answer Git
A single commit hash points to the exact code at that moment, which is essential for reproducing work.
-
19
In a Python project, write the conventional name of the file that lists the needed libraries and their versions so they can all be installed at once with pip install -r.
Show answer
Answer requirements.txt
Without pinned versions, the same code may give different results months later, or not run at all.
-
20
What do you call the property that lets other people rerun an experiment with the same data, code and settings and get the same result?
Show answer
Answer Reproducibility
At grade 1, others must be able to rebuild your work from the documents alone. Write down the data source, code version, seed and settings.