LLM Intuition Bank · Build model intuition without any maths

[akash-coded]

LLM Intuition Bank

MCQs, diagrams, error-fixing, and a systems recap — companion to Module 01

Zero math. Builds intuition, application, and key-term fluency. TravelMind is the running example.

How to use this bank

• Attempt each section closed-book first, then check the key at the end.
• Italic Breadcrumb lines are nudges, not answers. Read them only if stuck.
• Diagrams are plain ASCII, so they render in any markdown viewer.
• The recap at the end ties every concept into one mental model. Read it last.

---

Section A: Single choice

Guidance: one right option each. Watch for the option that is almost true.

A1. An LLM produces a response by:

• a) retrieving the closest stored answer from a database
• b) predicting one token at a time, conditioned on every prior token
• c) searching the web and pasting the top result
• d) running a fixed script

A2. "Llama 4 Scout: 109B total parameters, 17B active" describes:

• a) a bug
• b) quantization
• c) a Mixture-of-Experts design
• d) fine-tuning

A3. Which nesting is correct?

• a) LLMs ⊃ Generative AI
• b) Generative AI ⊃ Foundation Models ⊃ LLMs
• c) Foundation Models ⊃ Generative AI
• d) Reasoning LLMs ⊃ LLMs

A4. One token is roughly:

• a) one letter
• b) one word
• c) about 4 characters of English
• d) one sentence

A5. temperature = 0 means:

• a) fully deterministic and bit-exact every time
• b) greedy decoding, always take the highest-probability token
• c) the model is switched off
• d) randomness is mathematically impossible

Breadcrumb: greedy and bit-exact are not the same claim.

A6. "Lost in the middle" refers to:

• a) the model dropping the system prompt
• b) models attending less to the middle of a long context
• c) tokens being deleted at random
• d) the request timing out

A7. On most models, output tokens versus input tokens are:

• a) the same price
• b) cheaper
• c) 3 to 5 times more expensive
• d) free

A8. Which component turns token IDs into vectors?

• a) tokenizer
• b) embedding layer
• c) output head
• d) stop sequence

A9. Hallucination is best described as:

• a) a rare bug removed by retrying
• b) confident, plausible output that may be false, and a structural trait
• c) a hardware fault
• d) a network error

A10. A fresh API call carries no memory of earlier calls unless:

• a) you toggle a setting
• b) you resend the prior history yourself
• c) you pay for a higher tier
• d) you pick a larger model

A11. For high-volume intent classification in TravelMind, the sensible class is:

• a) Opus class
• b) Sonnet class
• c) Haiku class
• d) a reasoning model on max effort

A12. Open-weight models are the right call mainly when you need:

• a) the highest absolute capability ceiling
• b) zero infrastructure work
• c) data sovereignty and the option to self-host
• d) guaranteed native multimodality

A13. max_tokens controls:

• a) randomness
• b) a hard cap on output length
• c) the context window size
• d) the price per token

A14. The same prompt at temperature 0.7, sent twice:

• a) always matches exactly
• b) often differs, because sampling is probabilistic
• c) returns an error the second time
• d) costs double

A15. A 1M token context window guarantees:

• a) perfect recall across all of it
• b) a ceiling on input size, not a recall guarantee
• c) faster responses
• d) a lower bill

---

Section B: Multiple choice

Guidance: select all that apply. Most have two or three correct options. The wrong option is usually an absolute claim ("always", "all", "equal").

B1. Tokens are a unit of:

• a) cost
• b) latency
• c) context budget
• d) a fixed count that is equal across all languages

B2. Two transformer LLMs can legitimately differ in:

• a) tokenizer
• b) alignment / RLHF approach
• c) dense versus Mixture-of-Experts
• d) the laws of probability they obey

B3. Ways to lower TravelMind's token bill:

• a) trim the system prompt
• b) drop to a smaller model class where quality holds
• c) cap max_tokens
• d) raise temperature

B4. Signs you should reach for a tool instead of the model's memory:

• a) you need today's live flight status
• b) you need a fact with a real citation
• c) you need a catchy tagline
• d) you need the current weather at the destination

B5. Which belong under the Foundation Model umbrella?

• a) LLMs
• b) large pre-trained image generators
• c) a hand-coded if-else rules engine
• d) reasoning LLMs

B6. True of Mixture-of-Experts models:

• a) total parameters exceed active parameters
• b) "looks huge, runs cheap"
• c) every parameter fires on every token
• d) inference cost can stay low relative to total size

B7. Which lower hallucination risk in production:

• a) retrieval / grounding
• b) guardrails
• c) evaluations
• d) raising temperature

B8. Because each API call is stateless:

• a) you must resend history for continuity
• b) history eats into the context budget
• c) longer history costs more tokens
• d) the model quietly stores your data between calls

B9. Which are inference knobs on the API surface?

• a) temperature
• b) top_p
• c) max_tokens
• d) the model's parameter count

Breadcrumb: one of these is fixed when the model is trained, not set per request.

B10. Picking a model by brand name alone is wrong because:

• a) variant differs (dense vs MoE)
• b) context windows differ
• c) cost per token varies widely
• d) all brands behave identically

---

Section C: Diagrams to decipher

Guidance: read the ASCII flow, then answer. Two diagrams have a deliberate mistake to catch.

C1. Name the two blank stages, then explain the loop.

[ Prompt text ]
      |
      v
[  BLANK 1   ]        <- name this stage
      |
      v
[ Embedding layer ] <-----+
      |                   |
      v                   |
[ Transformer blocks ]    |
      |                   |
      v                   |
[ Output head ]           |
      |                   |
      v                   |
[  BLANK 2   ]        <- name this stage
      |                   |
      v                   |
[ Append token ] ---------+    loop: the appended token re-enters here

• BLANK 1 = ____________
• BLANK 2 = ____________
• Why does the arrow loop back? ____________

C2. One node is in the wrong place. Name it and say where it belongs.

        [ Generative AI ]
               |
               v
      [ Foundation Models ]
               |
               v
           [ LLMs ]
            /      \
           v        v
[ Reasoning LLMs ]   [ Image generator ]   <- something here does not belong

• Misplaced node: ____________
• Correct placement and why: ____________

C3. This pipeline is scrambled. Rewrite it in the right order.

[ Tokenizer ] -> [ Output head ] -> [ Transformer blocks ] -> [ Embedding layer ] -> [ Probabilities ]

• Correct order: ____________

C4. Everything below draws from one shared budget.

[ System prompt ]    --\
[ Chat history ]     ---\
[ Tool definitions ] ---+--> [ Context window budget ]
[ Retrieved chunks ] ---/
[ Room for output ]  --/

• If the budget is nearly full and the answer gets cut off, which two do you trim first, and why? ____________
• Which one can you almost never drop? ____________

---

Section D: Spot the error and rectify

Guidance: every line is wrong. Rewrite the corrected version in one sentence.

D1. "temperature 0 makes the model fully deterministic and bit-exact."

D2. "A 70B model is always smarter than a 7B model."

D3. "With a 1M token window, the model recalls everything in it equally well."

D4. "The model remembers our previous conversation on its own."

D5. "Output tokens are cheaper than input, so long replies are free money."

D6. "Hallucination is a bug we can fully patch out."

D7. "The same sentence in Hindi and English costs the same number of tokens."

D8. "Generative AI is a kind of LLM."

D9. "For TravelMind's refund-amount calculation we set temperature 1.2 to keep it interesting."

D10. "To be safe, route every TravelMind message through the Opus-class model."

---

Section E: Apply it (TravelMind)

Guidance: short scenarios. Pick the option and give a one-line reason.

E1. 50,000 inbound messages a day need their intent tagged before routing.

• Class: ________ Temperature: low / high Reason: ________

E2. TravelMind drafts an empathetic apology for a 3-hour delay.

• Class: ________ Temperature: low / high Reason: ________

E3. TravelMind confidently states a refund clause that does not exist in your docs.

• What is happening: ________ Fix 1: ________ Fix 2: ________

E4. A single user query in the agent flow burned roughly 6 times the tokens you estimated. Most likely reason?

• ---

E5. Decode the PM sentence in plain English.

"Run TravelMind on Llama 4 Scout via Bedrock with a guardrail, low temperature, 200K context."

• Llama 4 Scout: ________
• Bedrock: ________
• guardrail: ________
• low temperature: ________
• 200K context: ________

---

More question types

Section F: Match the pair

Guidance: match each term on the left to its one meaning on the right. Two terms are sampling knobs, do not swap them.

Term		Meaning
F-a Tokenizer		1. caps how long the output can get
F-b Embedding layer		2. confident output that may be false
F-c Temperature		3. total params huge, active params small
F-d top_p		4. each token depends on all prior tokens
F-e max_tokens		5. randomness dial on token sampling
F-f Context window		6. filters content, PII, denied topics
F-g Hallucination		7. total text the model can hold at once
F-h Mixture-of-Experts		8. turns text into integer IDs
F-i Autoregressive		9. turns token IDs into vectors
F-j Guardrail		10. keeps the smallest set of top tokens whose probability mass clears a threshold

---

Section G: Put it in order

Guidance: write the sequence. No math, just the right chain.

G1. Order the inference stages: Output head, Tokenizer, Sample token, Embedding layer, Transformer blocks.

G2. Order the nesting from widest to narrowest: LLMs, Generative AI, Reasoning LLMs, Foundation Models.

G3. Order by cost per output token, cheapest first: Opus class, Haiku class, Sonnet class.

---

Section H: Fill the blank, key terms

Guidance: one word or short phrase per blank. Tests term recall.

H1. Splitting text into integer IDs is the job of the ____.

H2. A window of 200K ____ holds roughly 300 pages of text.

H3. ____ sampling keeps the smallest set of top tokens whose probability mass passes a threshold.

H4. A model with 109B total but 17B ____ parameters is a Mixture-of-Experts.

H5. The trait where a model answers confidently but possibly falsely is ____.

H6. Because generation is ____, each token depends on every token before it.

H7. For TravelMind to recall a passenger across turns, the ____ must resend the history.

H8. A ____ filters PII and denied topics around the model.

---

Section I: Odd one out

Guidance: pick the item that does not belong and say why in a few words.

I1. Opus class · Sonnet class · Haiku class · Tokenizer

I2. temperature · top_p · top_k · parameter count

I3. Hallucination · stale knowledge · no native memory · GPU overheating

I4. Retrieval · guardrails · evaluations · raising temperature

I5. Llama · Qwen · DeepSeek · GPT

---

Section J: Complete the analogy

Guidance: fill the blanks. Several answers can work, the key shows one good fit.

J1. Tokens are to an LLM bill what ____ are to an electricity bill.

J2. A Mixture-of-Experts model is like a ____ that only wakes the specialist it needs.

J3. Lost in the middle is like reading a long ____ and best remembering the start and the end.

J4. A stateless API call is like a ____ that has no memory of your last visit.

J5. Routing across Haiku, Sonnet, and Opus is like assigning work to a ____, a ____, and a ____ by difficulty.

---

Section K: True or false, then fix

Guidance: mark T or F. For each F, write the one-line correction.

K1. Foundation Models are a subset of LLMs.

K2. temperature 0 is bit-exact deterministic.

K3. Output tokens usually cost more than input tokens.

K4. A 1M context window guarantees uniform recall.

K5. Open-weight models let you self-host for data sovereignty.

K6. Every parameter in an MoE model fires on every token.

---

Section L: Compare and rank

Guidance: relative order only, no numbers needed.

L1. Rank by speed, fastest first: Opus · Haiku · Sonnet.

L2. Rank by cost per output token, cheapest first: Opus · Sonnet · Haiku.

L3. For the same meaning, rank token count low to high: an English sentence · the Hindi sentence.

L4. Rank by hallucination risk when ungrounded, lowest first:

• summarise a paragraph I pasted
• cite three papers from memory
• give today's flight status from memory

Breadcrumb: which task already hands the model the facts it needs?

---

Section M: Say it to a PM in one line

Guidance: plain business English, no jargon. One sentence each.

M1. Why can't we just trust the model for live flight status?

M2. Why does our agent cost more than a single chat call?

M3. Why did the same prompt give two different taglines?

M4. Why use Haiku for classification instead of Opus?

M5. Why does the Hindi version of our bot cost more per message?

---

Recap: the systems view

Read this after attempting the bank. It connects every concept into one model.

The one-paragraph mental model

An LLM turns your prompt into tokens, predicts the next token from all the prior ones, samples it, appends it, and repeats. Output is probabilistic, bounded by a context window, frozen at a training cutoff, and stateless across calls. It completes plausible patterns, so it can sound certain and still be wrong. Every weakness has a named fix.

How the pieces connect

INPUT:  prompt text  =  a sequence of tokens
            |
            v
MODEL:  tokenize -> embed -> transformer blocks -> output head
            |
            v
        sample one token (probabilistic) -> append -> loop
            |
            v
OUTPUT: tokens out   (you pay per token, longer = slower)


WHAT BITES                         WHAT YOU REACH FOR
- context window ceiling     ->    retrieval, chunking, summarise
- stale knowledge            ->    tools, live data
- no native memory           ->    resend history, memory store
- hallucination              ->    grounding + guardrails
- probabilistic variance     ->    temperature + evals / QA

The whole game is a trade among four forces: capability, cost, latency, and control. Model class sets capability and cost. Token count sets cost and latency. Knobs and grounding set control. Pick deliberately on every task, since no single model or setting wins everywhere.

Decision heuristics

• High-volume classification or routing → Haiku class, low temperature.
• Balanced production quality → Sonnet class.
• Hard multi-step reasoning, low volume, high stakes → Opus class.
• The answer must be a fact → ground it with retrieval or a tool, never trust memory.
• Numeric or policy output → low temperature.
• Creative wording → moderate temperature.
• Continuity across turns → your app resends the history.
• Agent cost exploding → check token amplification from thinking, tools, retrieval, summarising.
• Data sovereignty required → open-weight, self-host or Bedrock with full control.

Key terms in one line

• Token: the atomic unit of text, roughly 4 English characters, the thing you pay for.
• Tokenizer: turns text into integer token IDs.
• Embedding layer: turns token IDs into vectors.
• Transformer block: attention plus feed-forward, stacked many times, holds most parameters.
• Output head: produces a probability over the next possible tokens.
• Autoregressive: each token is generated using all the tokens before it.
• Parameters: the learned weights, "70B" is their count, not a quality guarantee.
• Mixture-of-Experts: huge total parameters but only a slice active per token, looks huge, runs cheap.
• Context window: how much text the model can hold at once, measured in tokens.
• Lost in the middle: weaker recall for the middle of a long context.
• Temperature: the randomness dial on sampling.
• top_p: nucleus sampling, the smallest top set of tokens whose probability mass clears a threshold.
• max_tokens: the hard cap on output length.
• Hallucination: confident, plausible, possibly false output, structural not a bug.
• Guardrail: a filter for content, PII, and denied topics around the model.
• Foundation model: a large pre-trained model adaptable to many tasks, the umbrella above LLMs.

[ksvivekanand86-source]

A1. A
A2. B
A3. B
A4. C
A5. B
A6. B
A7. C
A8. B
A9. B
A10. B
A11. C
A12. B
A13. B
A14. B
A15. B

[ksvivekanand86-source]

B1. a, c
B2. a, c, d
B3. a, b, c
B4. a, b, d
B5. a, b, d
B6. a, b, d
B7. a, b, c
B8. a, b, c
B9. a, b, c
B10. a, b, c

[NimeshNairIBS]

<style> </style>

A1	B
A2	b
A3	b
A4	c
A5	b
A6	b
A7	c
A8	b
A9	b
A10	b
A11	d
A12	a
A13	b
A14	b
A15	b
B1	a,c
B2	b,c,d
B3	a,b,c
B4	a,c,d
B5	a,b,d
B6	b,d
B7	a,c,d
B8	a,b,c
B9	a,b,c
B10	a,b,c
C1	Tokenizer Final Output
C2	Image generator LLMs are text based. Image genrator can be a foundation model
C3	Tokenizer ->Embedding layer->Transformer blocks ->Output head -Probabilities
C4	Chat history. It increased the tokens System promp
D1	temperature 0 makes the model greedy
D2	Capability per dollar per latency is the right choice
D3	With a 1M token window, a larget amount of text can be processed but cant be recalled well
D4	Long term memory to be configured for the model remembers our previous conversation on its own
D5	Output tokens are more expensive thatn input tokens and will cost more
D6	Hallucination is not a bug, it can be reduced by proper measure of ensuring checks to ensure the related sources are selected, but cant be avoided
D7	No, tokens vaires with languages. It may cost more or less
D8	LLM is a subste of generative AI
D9	temperature needd to keep to 0 to achieve close to deterministics
D10	LLM needs to be selected based on the use case. Best model are not right modles. Capability per dolloar per latency
E1	Classificaiton.Low.Classificaiton of message based on a specific rule. Needs to be deterministic as possible
E2	Text generation. High. Temperatur is high as need to have emotional message which needs creativity/randomness
E3	Halluciniation.Tighthen output check point. Improve training data
E4	max_token is not set.multiple agent calls
E5	Llama 4 Scout: AI model Bedrock: Access via AWS managed service guardrail: Ensure Defense mechanism are set low temperature: need to be close to be determinist 200K context: Should process large instructino documents
F	F-a Tokenizer :	8. turns text into integer IDs
	F-b Embedding layer :	9. turns token IDs into vectors
	F-c Temperature :	5. randomness dial on token sampling
	F-d top_p	10. keeps the smallest set of top tokens whose probability mass clears a threshold
	F-e max_tokens	1. caps how long the output can get
	F-f Context window	7. total text the model can hold at once
	F-g Hallucination	2. confident output that may be false
	F-h Mixture-of-Experts	3. total params huge, active params small
	F-i Autoregressive	4. each token depends on all prior tokens
	F-j Guardrail	6. filters content, PII, denied topics
G1	Sample token,Tokenizer,Embedding layer,Transformer blocks,Output head
G2	Generative AI,Foundation Models,LLMs,Reasoning LLMs
G3	Haiku class,Sonnet class,Opus class
H1	tokeniser
H2	Context
H3	Nucleus
H4	active
H5	Hallucintion
H6	Autoregressive
H7	Context
H8	Guardrail
I1	Tokenizer
I2	parameter count
I3	GPU overheating
I4	raising temperature
I5	GPT -Licensed
J1	Kwh
J2	GP
J3	book
J4	Alzhimer patient
J5	SE, SSE, Architect
K1	FALSE
K2	FALSE
K3	TRUE
K4	FALSE
K5	TRUE
K6	FALSE
L1	Haiku, Sonnet, Opus
L2	Haiku, Sonnet, Opus
L3	Englist, Hindi
L4	summarise a paragraph I pasted cite three papers from memory give today's flight status from memory
M1	AI model can provide wrong result as its take best probable answer
M2	Agent wil provide more input data behind the scenes than single chat cell and hence the cost
M3	AI model take best probable answer and it can vary always
M4	Classificaiton don’t need high reasoning and Haiku is good choice for this in terms of efficiency and cost
M5	When hindi language is converted to token, more tokens is created when compared to english