Level 3
Alongside creation and design explanations, explain commands to install, configure and update tools. Choose one example and explain your design and reasons out loud.
Who can apply
Open to applicants aged 8 and over.
Screen-sharing interview
Share your entire screen through SONUB video, then carry out the tasks.
Level 3: example interview tasks
- When asked to create a , start typing immediately and finish within 30 seconds. Then verify that it actually works within 60 seconds after creation. To prevent AI-assisted cheating, skill creation is limited to 30 seconds, with an additional 60 seconds to demonstrate that it works.
- Questions cover commands for installing, configuring and updating tools such as , , , , and . Without hesitation, enter the commands yourself to perform installation and configuration.
- When asked how you would create an for a specific topic, start answering without hesitation and finish explaining the entire process within 3 minutes.
The 60-second verification period starts after creation is complete; it is separate from the 30-second creation period.
Several tasks of this difficulty are given. This is the entry-level standard for level 3.
Common to every level
Example interview question: an AI that guesses the next word
Examiner Build an that guesses the next word while a student types study notes. Explain, step by step, how you would make it and how you would check that it works.
Answer in these four steps. Each step shows what to say, with a short example.
-
1. Goal
Say what the AI does and for whom — what goes in and what comes out. Example: “When a student types ‘machine’, it suggests ‘learning’.”
-
2. Data
Say what it learns from, how much, and what you keep aside for testing. Example: “2,000 sentences from my own notes — 80% to learn from, 20% kept for testing.”
-
3. How it works
Say how it works inside and what you build it with. Example: “A counts which word comes after which and suggests the most frequent next word.”
-
4. Check
Say how you know it works: one number you measure and one limit. Example: “On the test sentences, I count how often the real next word is among the top three suggestions. It looks only one word back, so long sentences go wrong.”
The examiner may change the topic — for example, a filter for spam messages or a lunch-menu recommender — but you answer in the same four steps. Higher levels explain each step in more detail, and the interview also has other questions and tasks for each level.
What to learn
- Start typing immediately and create the requested within 30 seconds.
- Verify that the skill works within 60 seconds after creation.
- Start explaining a topic-specific AI design immediately and cover the entire creation process within 3 minutes.
How you are marked
- Was the skill completed within 30 seconds?
- Was its actual operation verified within 60 seconds after creation?
- Were several questions at this level answered promptly and independently?
Practice interview questions
These 11 questions are examples of what the examiner may ask in the interview. Do not try to hit a single correct word — explain your answer out loud. Open each one to see a sample explanation and a follow-up question. They are not official exam questions.
AI names and companies
-
1
Examiner Explain what stands for and what each of its three words — G, P and T — means.
Show a sample explanation
Sample explanation GPT stands for Generative Pre-trained Transformer. Generative means it creates new text, Pre-trained means it was trained in advance on a huge amount of text, and is the name of the design first published in 2017 in the Google researchers' paper " Is All You Need". 's names (GPT-4, GPT-5 and so on) and the GPT in come from this.
Follow-up question Are ChatGPT and GPT the same thing or different things?
-
2
Examiner Explain who makes by listing at least five AI developers and the main AI of each.
Show a sample explanation
Sample explanation For example: — , and ; — ; Google () — ; Meta — ; xAI — ; — DeepSeek; Mistral AI — Mistral and Le Chat; Alibaba — . A company often gives its and the app that uses the model different names (OpenAI's GPT models and the ChatGPT app).
Follow-up question Why can AIs from different companies give different answers to the same question?
Learning, data and vectors
-
3
Examiner Explain how and differ.
Show a sample explanation
Sample explanation Machine learning is the whole family of methods where, instead of a person writing every rule, the computer is shown many examples () and finds the rules itself. Deep learning is one branch of it that learns with stacked many layers deep. Classic machine learning often needs a person to pick which to look at, while deep learning learns the features itself from raw material such as image or text — at the cost of far more data and computation.
Follow-up question Is today's machine learning or deep learning? If both, why?
-
4
Examiner Explain how an differs from a , using "puppy" and "dog" as an example.
Show a sample explanation
Sample explanation A vector is just a list of numbers in a fixed order, like [0.2, -1.5, 3.0], and has no meaning on its own. An embedding is a way of turning the meaning of a word, sentence or picture into such a vector (or the vector made that way); it is learned so that words with similar meanings, like "puppy" and "dog", end up as nearby numbers. So every embedding is a vector, but not every vector is an embedding. Search, recommendations and use embeddings to compare meanings.
Follow-up question Why can embeddings find texts with the same meaning even when the words differ?
-
5
Examiner Explain what is and why good matters.
Show a sample explanation
Sample explanation Training data is the many examples (photos, text, sounds and so on) shown to an so it can learn. An AI's answers resemble its training data, so if it learns from wrong data it gives wrong answers.
Follow-up question If you were building an AI that recognises cats, what photos would you collect?
Prompts and agents
-
6
Examiner Explain how a differs from a .
Show a sample explanation
Sample explanation A prompt is the question or instruction a user types in each time. A system prompt is an instruction the app or developer puts in before the conversation starts; it sets the 's role, tone and rules (for example, "You are a kind English teacher. Answer in Korean.") and applies for the whole conversation. Users usually do not see it, and when it conflicts with a user prompt, the is usually trained to follow the system prompt first.
Follow-up question If you built an English conversation for students, what would you write in its system prompt?
-
7
Examiner Explain how an differs from a .
Show a sample explanation
Sample explanation An agent is an system that takes a goal, makes its own plan, uses tools (reading files, running commands, searching) and checks the results until the job is done. A harness is the outer that wraps a and makes it run as an agent: it adds the , connects and actually runs the tools, feeds the results back to the in a loop, and manages and . Put simply, agent = language model + harness; and are examples of harnesses. The same model can make a stronger or weaker agent depending on its harness.
Follow-up question Why can two agents that use the same model differ in skill?
-
8
Examiner Explain how an differs from a .
Show a sample explanation
Sample explanation A language model is the itself: it takes text in and continues it with the next text (). On its own it cannot open files or run commands, and it answers once per question. An agent uses a language model as its brain and adds tool use, repeated steps and result checks, so it can take a goal like "fix this " and act many times on its own until it is done. For example, is a language model, and is an agent that uses Claude.
Follow-up question Why might a language model get "today's weather" wrong, and how could an agent solve that?
Neural networks, language models and context
-
9
Examiner Explain what a is and how it is like and unlike the human brain.
Show a sample explanation
Sample explanation A neural network (artificial neural network) is a inspired by the nerve cells () of the human brain. It connects simple calculating units in layers: each unit multiplies its inputs by , adds them up and passes the result through an . adjusts those weights from so the network learns patterns. It is inspired by the brain, but it does not work the way a real brain does.
Follow-up question What does a "layer" of a neural network mean?
-
10
Examiner Explain, one by one, whether , and are too, given that is an LLM.
Show a sample explanation
Sample explanation Gemini (Google), Claude () and Grok (xAI) are all names of LLM (Large Language Model) families, and each also has a app of the same name. Codex is different: today Codex is 's (Codex CLI and app). It is not an LLM itself but an whose brain is a GPT-family , including coding-tuned models such as GPT-5-Codex. In 2021 there was also a code-generating model named Codex. You need to tell whether a name means a model, an app or an agent.
Follow-up question Is an LLM or an app? What runs inside it?
-
11
Examiner Explain what changes when an that supports a 1M-token is cut down to a 400K-token context.
Show a sample explanation
Sample explanation The is how much text (how many ) the can see at once. At 400K, the conversation, files read and tool results that fit at once shrink to 40%, so a large codebase or long document no longer fits in one go. In long tasks the space fills faster, so earlier content is summarized (compact) or cut more often, and earlier instructions or details can be forgotten along the way. Reading only the files you need, search (), note files and splitting the work become more important. On the other hand, fewer tokens per request can lower cost and response time. The model itself does not become less intelligent.
Follow-up question How would you split up a long task for an agent with a smaller context?