Skip to main content

This source assessment covers both Module 6 (Generative AI) and Module 9 (Agentic AI).

Module 9 Test: Agentic AI

This is your progress tracker. 20 questions: 5 Easy, 10 Moderate, 5 Hard. First solve all the questions yourself (think about the output in your mind before running the code in Colab), then check with the Answer Key.

ScoreMeaning
18 to 20Excellent! Your grip on Modules 6 and 9 is strong.
14 to 17Very good! Read the weak chapter again.
10 to 13Okay. Run the code yourself and look again.
Below 10Don't worry. Start again from Chapters 29 to 33 and run every code
in Colab.

EASY​

Q1. What was the biggest change that the Transformer made compared to RNN/LSTM? A) It read text one word at a time

B) It worked only on images

C) It processed tokens together (in parallel) and caught relations through attention

D) It completely removed the need for training

Q2. What does Query (Q) mean in attention?

A) What I am looking for (the search)

B) The actual content/information of the token

C) The position of the token (order)

D) The value of the loss

Q3. What is the basic goal of LLM pre-training?

A) Classifying pictures

B) Cleaning data

C) Only doing translation D) Correctly guessing the next token

Q4. Which statement about the OpenAI API key is correct? A) It should be written in code and put on GitHub

B) It should be kept in an environment variable or Colab Secrets

C) It should be sent to friends on WhatsApp

D) It should be written in a website URL

Q5. Who actually runs the real function in function calling?

A) The model itself

B) The OpenAI server itself

C) Your own code D) The browser

​

MODERATE​

Q6. Why do we divide the scores by sqrt(d_k) in scaled dot-product attention?

A) So that the output becomes smaller

B) So that softmax does not saturate (one weight does not get close to 1)

C) So that memory is saved

D) So that we know the order of tokens

Q7. Why is Positional Encoding needed in a Transformer?

A) To reduce the loss

B) To increase the vocabulary

C) To decide the number of heads

D) Because attention does not know the order of tokens by itself

Q8. Which is the correct difference between BERT and GPT?

A) BERT is encoder-only (looks both ways, masked LM); GPT is decoder-only (only the previous, predicts the next token)

B) BERT is decoder-only and GPT is encoder-only

C) Both are exactly the same

D) BERT generates text, GPT only does classification

Q9. What happens in sampling when temperature is changed from 0.2 to 2.0?

A) The answer becomes more certain and the same

B) The intelligence of the model increases

C) The answers become more random and different

D) The size of the answer becomes smaller

Q10. What is the biggest benefit of RAG?

A) The model's weights change

B) The need for a GPU disappears

C) Token cost becomes zero

D) Answers come from fresh/your own documents with sources, which reduces hallucination

Q11. Giving 3 examples (input -> output) in the prompt to make the model do a new task is called what?

A) Fine-tuning

B) Few-shot prompting

C) Embedding

D) Tokenization

Q12. For which problem is fine-tuning most suitable?

A) When you need a consistent style or a strict format of answer

B) To make the model remember new facts that change every day

C) To search the internet

D) To increase the context window

Q13. What is the purpose of max_turns in the agent's run loop?

A) To make the answer shorter

B) To change the model

C) To stop the loop from running too long (saves cost and stops endless cycles)

D) To clear the memory Q14. What is the difference between handoff and agent-as-tool?

A) In a handoff the manager keeps control; in as_tool the specialist gets control

B) In a handoff control goes to the specialist; in as_tool the manager keeps control

C) Both are exactly the same

D) Handoff is only for tools

Q15. What is prompt injection?

A) The API key being very long

B) A GPU memory error

C) The tokenizer running wrongly

D) Hidden instructions in outside text (web page/email) that mislead the model

​

HARD​

Q16. What will be the output of this code?

import numpy as np
mask = np.tril(np.ones((3, 3), dtype=bool))
print(mask.sum())

A) 6 B) 3

C) 9

D) 0

Q17. In Multi-Head Attention, if d_model = 32 and h = 4 heads, what are the total parameters of Wq, Wk, Wv, Wo (without bias)?

A) 1024

B) 16384 C) 4096

D) 512

Q18. On a 1000 x 1000 weight matrix, what will be the total parameters of A and B for LoRA (rank r = 4)?

A) 4,000

B) 8,000 C) 1,000,000

D) 4,000,000

Q19. What will be the output of this code?

def cost(i, o):
return i/1_000_000*1.00 + o/1_000_000*4.00
print(round(cost(2000, 500), 3))

A) 0.002 B) 0.006 C) 0.04 D) 0.004

Q20. Each step of an agent is 90% correct. After 5 steps, what is approximately the probability that the whole task is correct?

A) 0.90 B) 0.45 C) 0.59

D) 0.50

Modules 6 and 9 Answer Key (with explanation)​

#LevelAnswerExplanation
1EasyC) Processed tokens together
(parallel) and caught relations
through attention
There is no recurrence in the
Transformer, there is
attention: all tokens are seen
together (in parallel), which
makes training fast and
makes distant relations easy.
2EasyA) What I am looking for
(search)
Query = search, Key =
identification (label), Value =
actual content. The match of
Query with Key makes the
weights.
3EasyD) Correctly guessing the next
token
An LLM is a very large
language model which learns
to guess the next token with
cross-entropy; other abilities
come from this.
4EasyB) Kept in an environment
variable or Colab Secrets
The API key is a secret. If it is
stolen, someone else uses it
at your expense, so use
environment/Secrets.
5EasyC) Your own codeThe model only writes the
function name and
arguments; your code runs
the function and gives the
result back to the model.
6ModerateB) So that softmax does not
saturate (one weight does
not get close to 1)
When d_k increases, the
spread of dot products
increases and softmax
becomes hard; dividing by
sqrt(d_k) keeps the weights
soft.
7ModerateD) Because attention does
not know the order of tokens
by itself
Self-attention is permutation-
equivariant: shuffling tokens
only changes the place of the
output. So position
information is added
separately.
8ModerateA) BERT is encoder-only
(looks both ways, masked
BERT is a bidirectional
encoder that is good at
LM); GPT is decoder-only
(onlytheprevious, predicts
understanding text; GPT is a
causal decoder that
the next token)generates text by guessing
the next token.
9ModerateC) The answers become more
random and different
Small temperature = sharp
probabilities (certain
answer); large = equal
(random). It does not change
intelligence, only
randomness.
10ModerateD) Answers come from
fresh/your own documents
with sources, which reduces
hallucination
RAG first finds the right
chunks and gives them in the
prompt, so the answer is
based on your data and can
be verified with sources.
11ModerateB) Few-shot promptingIn few-shot, the examples are
inside the prompt and the
model's weights do not
change; in fine-tuning the
weights change.
12ModerateA) When you need a
consistent style or a strict
format of answer
Fine-tuning teaches style,
format and behaviour. For
new and changing facts, RAG
is better.
13ModerateC) To stop the loop from
running too long (saves cost
and stops endless cycles)
max_turns is a safety limit: if
the agent gets stuck in a tool
cycle, a MaxTurnsExceeded
error comes and money is not
wasted.
14ModerateB) In a handoff control goes
to the specialist; in as_tool
the manager keeps control
Handoff: the new agent
handles the conversation.
Agent-as-tool: the specialist
does one helping job and the
manager combines
everything into the answer.
15ModerateD) Hidden instructions in
outside text (web
page/email) that mislead the
model
Outside text read by an agent
may contain an attacking
instruction. Treat outside text
as data, not orders, and limit
tool permissions.
16HardA) 6np.tril keeps the lower
triangle (including the
diagonal) as True: 1 + 2 + 3 =
6 True values.
17HardC) 4096Parameters = 4 x d_model x
d_model = 4 x 32 x 32 = 4096.
Heads only use these
matrices by splitting them, so
the parameters do not
increase with h.
18HardB) 8,000A: (r x d) = 4 x 1000 = 4000, B:
(d x r) = 1000 x 4 = 4000.
Total 8000, while the full
matrix has 1,000,000
parameters (only 0.8%).
19HardD) 0.004Input: 2000/1e6 x 1 = 0.002.
Output: 500/1e6 x 4 = 0.002.
Sum = 0.004 dollars.
20HardC) 0.59p^n = 0.9^5 = 0.59049, i.e.
only about 59%. Long agents
need small steps, checks and
a way back.

Your score: ______ / 20