Quiz: How LLMs Work
Test your understanding of how Large Language Models learn and generate text.
30 questions - Click on each question to reveal the correct answer.
Question 1
What is the primary mechanism LLMs use to generate text?
A) Looking up answers in a database
B) Predicting the next word/token
C) Copying from training data
D) Using internet search
Show Answer
Correct Answer: B
LLMs generate text by predicting what word or token should come next, one at a time, based on everything that came before.
Question 2
What is a "token" in the context of LLMs?
A) A type of cryptocurrency
B) A piece of text (usually a word or word part)
C) A login credential
D) A hardware component
Show Answer
Correct Answer: B
Tokens are the basic units that LLMs process—usually parts of words. For example, "understanding" might be split into "under-stand-ing" (3 tokens).
Question 3
What is the context window?
A) A graphical display of the AI
B) The maximum text the model can see at once
C) A setting for screen brightness
D) The model's long-term memory
Show Answer
Correct Answer: B
The context window is the maximum amount of text the model can "see" and consider when generating a response—like its short-term memory limit.
Question 4
What happens when a conversation exceeds the context window?
A) The model crashes
B) Oldest messages may be lost or summarized
C) The model automatically upgrades
D) Nothing—context is unlimited
Show Answer
Correct Answer: B
When context is exceeded, oldest messages typically get truncated or summarized, potentially causing the model to "forget" early conversation.
Question 5
What does "temperature" control in LLM generation?
A) The speed of generation
B) The randomness/creativity of outputs
C) The model's power consumption
D) The color of the interface
Show Answer
Correct Answer: B
Temperature controls how random or deterministic the output is. Low temperature = predictable, high temperature = more creative but risky.
Question 6
Why does training require so much text data?
A) To fill up storage space
B) To learn patterns across diverse topics
C) To create backup copies
D) Legal requirements
Show Answer
Correct Answer: B
Training requires billions of text samples so the model can learn patterns across many topics, writing styles, and contexts.
Question 7
What is a system prompt?
A) An error message
B) Hidden instructions shaping AI behavior
C) The model's source code
D) A user's first message
Show Answer
Correct Answer: B
A system prompt is hidden instructions (set by developers) that guide how the AI should behave throughout the conversation.
Question 8
How do LLMs "remember" earlier parts of a conversation?
A) Permanent database storage
B) The entire conversation is re-sent each time
C) Magic
D) Cloud backup
Show Answer
Correct Answer: B
LLMs don't have persistent memory. The entire conversation history is included in each message, up to the context limit.
Question 9
What is meant by "autoregressive" generation?
A) Automatic car features
B) Each generated token becomes input for the next
C) Self-regulating temperature
D) Regression in statistics
Show Answer
Correct Answer: B
Autoregressive means the model's output is fed back as input for generating the next token—each word influences what comes after.
Question 10
What is the main source of LLM training data?
A) Only Wikipedia
B) The internet, books, and various text sources
C) Manually written examples
D) Newspapers only
Show Answer
Correct Answer: B
LLMs are trained on diverse sources including web pages, books, scientific papers, code repositories, and more.
Question 11
What does "emergent abilities" mean for LLMs?
A) Abilities that cost extra
B) Capabilities that appear at scale without explicit programming
C) Emergency features
D) Starting abilities
Show Answer
Correct Answer: B
Emergent abilities are capabilities (like reasoning or translation) that appear when models reach sufficient scale, even though they weren't explicitly programmed.
Question 12
Why might an LLM give outdated information?
A) It's intentionally withholding
B) Training data has a cutoff date
C) The internet doesn't work
D) It's testing you
Show Answer
Correct Answer: B
LLMs only know what was in their training data. Information that changed after the knowledge cutoff date won't be reflected.
Question 13
What is "fine-tuning"?
A) Making the model smaller
B) Adjusting a pre-trained model for specific tasks
C) Removing bugs
D) Increasing speed
Show Answer
Correct Answer: B
Fine-tuning is taking a pre-trained model and further training it on specific data to specialize it for particular tasks (like conversation or code).
Question 14
What does RLHF stand for?
A) Really Large Human Files
B) Reinforcement Learning from Human Feedback
C) Rapid Learning High Frequency
D) Regular Language Human Format
Show Answer
Correct Answer: B
RLHF (Reinforcement Learning from Human Feedback) is a training technique where humans rate outputs, and the model learns to produce higher-rated responses.
Question 15
How many tokens is a typical page of text?
A) About 10
B) About 100
C) About 500-750
D) About 10,000
Show Answer
Correct Answer: C
A typical page of text is approximately 500-750 tokens, depending on the content and tokenization method.
Question 16
Why doesn't an LLM "know" facts like a database?
A) Databases are illegal for AI
B) It learns patterns, not stores facts
C) It deletes all facts
D) Facts are too expensive
Show Answer
Correct Answer: B
LLMs learn statistical patterns from text, not verified facts. They encode associations, not a fact database.
Question 17
What is a "hallucination" in LLM context?
A) A visual display bug
B) Generating plausible but false information
C) A dreaming state
D) A hardware failure
Show Answer
Correct Answer: B
Hallucination is when an LLM generates information that sounds convincing but is actually false or made up.
Question 18
What happens at low temperature settings?
A) The model runs slower
B) Outputs are more predictable and consistent
C) The model uses less power
D) Responses are shorter
Show Answer
Correct Answer: B
Low temperature makes the model more deterministic—it tends to pick the highest-probability words consistently, making outputs more predictable.
Question 19
How does the model choose between multiple likely next words?
A) Random selection
B) Weighted random sampling based on probabilities
C) Always picks the first
D) User votes
Show Answer
Correct Answer: B
The model samples from possible next words, weighted by their probabilities. Higher-probability words are more likely to be chosen, but randomness allows variety.
Question 20
What makes Claude particularly good for long documents?
A) Faster processing speed
B) Very large context window (200K tokens)
C) Better marketing
D) Older technology
Show Answer
Correct Answer: B
Claude has a context window of up to 200K tokens, allowing it to process entire books or large codebases at once.
Question 21
Why do LLMs need massive computing power for training?
A) For cooling purposes
B) To process billions of examples and adjust billions of parameters
C) Legal requirements
D) It doesn't actually need much power
Show Answer
Correct Answer: B
Training involves processing billions of text samples and adjusting billions (or trillions) of parameters, requiring enormous computational resources.
Question 22
What is "pre-training"?
A) Practice before the real test
B) Initial training on massive general text data
C) Software installation
D) User onboarding
Show Answer
Correct Answer: B
Pre-training is the initial phase where the model learns from massive general text data before any task-specific fine-tuning.
Question 23
Why might the same prompt give different answers?
A) The model is confused
B) Random sampling in generation creates variation
C) Internet connection issues
D) Different users get different models
Show Answer
Correct Answer: B
The random sampling process (affected by temperature) means the same prompt can produce different outputs on different runs.
Question 24
What is the relationship between training and prediction?
A) They're unrelated
B) Training teaches the model to predict better
C) Prediction teaches training
D) They're the same thing
Show Answer
Correct Answer: B
Training is the process of teaching the model to make better next-word predictions by adjusting parameters based on billions of examples.
Question 25
Why are instructions sometimes "forgotten" in long conversations?
A) The AI is being difficult
B) Old instructions get pushed out of context
C) Time limits
D) Memory corruption
Show Answer
Correct Answer: B
As conversation grows, old messages (including instructions) may fall outside the context window and be effectively "forgotten."
Question 26
What is the "compression hypothesis" about LLMs?
A) Models should be made smaller
B) Good prediction requires implicit understanding
C) Data should be compressed
D) Smaller files are better
Show Answer
Correct Answer: B
The compression hypothesis suggests that to predict text well, a model must implicitly learn grammar, facts, logic, and world knowledge—compression forces understanding.
Question 27
What does a larger context window allow?
A) Faster responses
B) Processing more text at once
C) Lower costs
D) Better graphics
Show Answer
Correct Answer: B
A larger context window allows the model to "see" more text at once—enabling analysis of longer documents and longer conversation memory.
Question 28
Why do biases appear in LLM outputs?
A) Intentional programming
B) Biases in training data get learned
C) Hardware limitations
D) User preferences
Show Answer
Correct Answer: B
If training data contains biases (historical, cultural, etc.), the model learns these patterns and may reproduce them in outputs.
Question 29
What is the role of "attention" in LLMs?
A) Making the AI pay attention to users
B) Weighing the importance of different parts of input
C) Keeping users focused
D) Hardware monitoring
Show Answer
Correct Answer: B
Attention mechanisms help the model weigh which parts of the input are most relevant for predicting the next word.
Question 30
How would you describe what LLMs "understand"?
A) They understand like humans do
B) They recognize patterns without true comprehension
C) They understand everything perfectly
D) They understand nothing at all
Show Answer
Correct Answer: B
LLMs recognize and reproduce patterns from training data. Whether this constitutes "understanding" is debated, but it's clearly different from human comprehension.
Results
| Score | Level |
|---|---|
| 27-30 | Excellent - Strong understanding of how LLMs work |
| 21-26 | Good - Some review may be helpful |
| 15-20 | Average - Review the chapters again |
| Below 15 | Needs comprehensive review |