AIMentor — AI Expertise Learning Platform

Untimed. Answers are graded when you submit, with explanations and spaced-repetition scheduling.

Question 1easy
At the most fundamental level, what does a large language model do?
Looks up answers in a databasePredicts a probability distribution over the next token, then repeatsExecutes a plan it forms before answeringCompresses and decompresses text
Question 2easy
What is a token?
Always exactly one wordA single characterA chunk of text from a fixed vocabulary — often a word or sub-word piece, including spacesAn API authentication key
Question 3medium
Why do token counts matter more than word counts in practice?
They don't — words are what's billedCost, context limits, and latency are all measured in tokensTokens are only relevant for non-English textToken counts determine the model's accuracy
Question 4medium
Roughly how many characters of English text correspond to one token (rule of thumb)?
~1 character~4 characters~20 characters~100 characters
Question 5medium
What does the attention mechanism let the model do?
Store facts in a lookup tableFor each token, weigh how much every other token in the context mattersConnect to the internetCompress the context window
Question 6hard
Attention compute scales in which way with sequence length?
Constant — length doesn't matterLinearly — twice the length, twice the computeQuadratically — twice the length, roughly four times the computeLogarithmically
Question 7medium
What is the context window?
The model's permanent memory of all past conversationsThe maximum number of tokens (prompt + output) the model can attend to at onceThe time limit for a responseThe number of parameters in the model
Question 8medium
Between two separate API calls, what does a stateless LLM remember about your first call?
The full conversation, automaticallyJust a summaryNothing — you must resend the history each timeOnly the system prompt
Question 9hard
Why do LLMs hallucinate — produce confident but false statements?
A bug that a good prompt fully removesThey optimize for plausible next tokens, not truth, and lack a built-in 'I don't know'Their database is out of dateRandom hardware errors
Question 10hard
Which are TRUE about the training-vs-inference split? (Select all that apply)
Training adjusts the model's parameters over large corpora and is done once by the providerThe model's knowledge is frozen at its training cutoff dateInference is what runs on each of your API callsInference updates the model's parameters with your prompt
Select all that apply.
Question 11easy
What does 'autoregressive' generation mean?
The model generates all tokens simultaneouslyEach generated token is fed back in to predict the next oneThe model automatically regresses to older knowledgeIt regenerates the prompt before answering
Question 12medium
You need an exact token count for a Claude request. What should you use?
OpenAI's tiktoken libraryA words × 1.33 estimateThe provider's own token counter (e.g. messages.count_tokens), for the target modelCharacter count divided by 4, always
Question 13medium
Name the neural-network architecture (one word) whose attention mechanism underlies modern LLMs.
Question 14hard
One word: an LLM turns each token into a numeric ____ (vector encoding meaning) before attention operates on it.

0 / 14 answered

How LLMs Work — Practice quiz