AI hierarchy
Each inner layer adds one requirement
AI acts intelligent · rule-based chess
ML learns from data · spam classifier
DL deep neural nets · image recognition
GenAI creates content · draft an email
AWS Certified AI Practitioner · AIF-C01
Visual cram sheet
The comparisons I leaned on most while preparing for the exam. Each block reads definition → example → exam signal.
Passed · 2026-07-24
01 · Mental map
Big → small
Containment ≠ synonym
Foundation model vs LLM
FM = reusable base model · LLM = language FM
Foundation model
Broad base · text/image/audio
Parent concept
Stable Diffusion · image FM
LLM
Language-focused FM
FM subset
Claude / GPT · text + code
All LLMs are FMs. Not all FMs are LLMs.
Model families · one verb each
Architecture/process names can overlap in one product
Transformer
attends
Diffusion
denoises
GAN
competes
VAE
compresses + rebuilds
Autoregressive
next item
CNN
local image patterns
RNN/LSTM
sequence state
ChatGPT
Transformer + autoregressive
- Stable Diffusion → diffusion
- DALL·E 3 → diffusion-based
- Firefly → exam: diffusion
- BERT → bidirectional transformer
- ResNet → image CNN
- DeepAR → time-series RNN
- SVM → separating boundary
- XGBoost → boosted trees
- WaveNet → audio waveform
Fast glossary
- FM · foundation model
- LLM · large language model
- RAG · retrieval-augmented generation
- RLHF · human-feedback reinforcement
- PII · personal data
- CVE · known vulnerability
- SHAP · feature attribution
- A2I · Augmented AI
- Context window · token capacity
- Distillation · teacher → smaller student
- Nova / Titan · AWS foundation models
02 · AWS service families
Parent → sub-service → trigger
Pick by job, not name
Amazon Bedrock
managed FM platform
Knowledge Bases
managed RAG
▸ private/current facts
answer from company policy PDFs
Guardrails
safety + PII + topics
▸ harmful content
block violence + mask card numbers
Model Evaluation
compare FMs before production
▸ auto = metrics • human = judgment
auto: built-in/custom + BERTScore/F1 • human: own dataset + coherence/relevance
Agents
tools + multistep actions
▸ execute task
check stock → place order
Embeddings
vectors for retrieval
▸ similarity/search
find shoes from photo + text
Customization
fine-tune / continued pre-train / distill
▸ stable behavior
always write legal-report format
Amazon SageMaker AI
build/train/deploy your model
Data Wrangler
clean + transform
▸ prepare raw data
fill missing ages · encode country
Feature Store
store/share features
▸ training + inference inputs
reuse customer_risk_score live
Ground Truth
label examples
▸ training dataset
mark images cat / dog
Canvas
no-code end-to-end ML
▸ business analyst
predict churn without code
JumpStart
pretrained model hub
▸ deploy quickly
deploy a pretrained Llama model
Autopilot
AutoML candidates
▸ automate build + tune
compare churn models automatically
Studio
ML development IDE
▸ data scientist workspace
notebooks + training jobs
Clarify
bias + explainability
▸ fairness / SHAP
loan approvals differ by group?
Model Monitor
production drift
▸ quality after deploy
live income data shifts
A2I
human reviews predictions
▸ inference review
review low-confidence insurance claim
Model Cards
document your model
▸ risk + intended use
limits of a credit-risk model
Model Registry
version + approve
▸ promotion lifecycle
approve model v4 for production
Model Dashboard
view model status
▸ models/endpoints
which endpoint is deployed?
Pipelines
ML workflow automation
▸ MLOps / CI/CD
prepare → train → deploy
MLflow
track experiments
▸ runs + metrics
compare learning rates
Inference: real-time = steady · serverless = sporadic · async = long/large · batch = offline dataset.
Amazon Q
choose user first
Q Developer
code + AWS operations
▸ developer / cloud
fix Lambda error · list functions
Q Business
enterprise knowledge
▸ employee / HR / IT
“What is our leave policy?”
Connect AI agents (was Q in Connect)
live agent guidance
▸ contact center
suggest reply during customer call
Quick Sight (was Q in QuickSight)
BI + visuals
▸ analyst / dashboard
“Chart sales by region”
- Q for EC2 · instance guidance
- Q Developer in chat apps (was AWS Chatbot) · Slack/Teams ops
- Q for Glue · ETL help
- Q Apps · no-code enterprise app
- PartyRock · no-code FM playground
Ready-made AI APIs
Call an API · no custom model training
- Comprehend · review sentiment
- Rekognition · faces in video
- Transcribe · call audio → text
- Polly · article → voice
- Translate · English → Korean
- Textract · invoice fields
- Lex · support chatbot
- Personalize · product suggestions
- Kendra · search company docs
- Mechanical Turk · crowd microtasks
- HealthScribe · clinical notes
- Transcribe Medical · medical speech
- Comprehend Medical · clinical NLP
- Forecast · predict demand
Pretrained API ≠ SageMaker custom model.
Security/service confusion
Inspector
CVEs / vulnerabilities
unpatched EC2 package
GuardDuty
active threats
unusual crypto-mining traffic
Macie
PII in S3
bucket contains passport numbers
AWS Config
resource configuration
public S3 rule violation
CloudTrail
API activity history
who deleted this model?
Artifact
AWS compliance docs
download SOC report
Audit Manager
audit evidence
collect PCI control evidence
Trusted Advisor
account advice
cost + security recommendations
PrivateLink
private service access
VPC → Bedrock without public internet
03 · GenAI decisions
What must change?
Knowledge · behavior · output
Prompt vs RAG vs fine-tuning vs RLHF
Same FM · different improvement method
Prompt
ask better · no weights
instructions/examples
“Answer in 3 bullet points”
RAG
look it up · no weights
fresh/private facts
retrieve today’s inventory
Fine-tune
learn behavior · weights change
tone/format/task
learn company support tone
RLHF / RFT
optimize via reward
human rankings · reward functions
humans rank answer A over B
Bedrock inference modes
On-Demand
pay per use · no term
new chatbot · uneven traffic
Provisioned
reserved throughput · no-commit or term
steady high-volume support
Smaller model → usually cheaper + faster.
Distillation: large teacher → smaller student · lower cost/latency.
Inference parameters
Temperature
randomness
low · financial report
Top K
fixed count
K=10 · exactly 10 candidates
Top P
probability mass
P=.9 · enough tokens to reach 90%
Stop sequence
halt pattern
stop at </answer>
Response length
min/max output tokens
max 100 tokens · not candidate count
K = quantity. P = probability pool. Length = output size. Low temperature = consistent.
Prompting
Zero-shot
“Classify this review”
Few-shot
show 3 labeled reviews first
Chain-of-thought
break tax problem into steps
Negative
“Do not mention competitors”
Multimodal choice · choose output
Search / match / recommend
“find shoes like this photo”
Multimodal embedding
retrieves existing items
Create / describe / answer
“describe damage in this photo”
Multimodal generative
creates a response
FM lifecycle
Pre-train
unlabeled general data · build FM
Continued pre-train
unlabeled medical text · add domain
Fine-tune
labeled Q&A · specialize task
Inference
new prompt · generate answer
04 · ML core
Signal → training → evaluation
Training signal decides
Learning type · identify the signal
Task name can vary · training signal decides
Supervised
labels → predict
emails labeled spam / safe
Unsupervised
no labels → groups
cluster customers by behavior
Semi-supervised
few labels + much unlabeled
100 labeled + 10k unlabeled reviews
Self-supervised
data makes targets · FMs
predict masked / next token
Reinforcement
actions + rewards
robot rewarded for avoiding collision
Clustering
unsupervised
discover 4 traffic patterns
Fraud / sentiment
can be semi-supervised
experts label only a small sample
Neural network
architecture, not paradigm
Data split
Study → mock exam → untouched final exam
Train
learn weights
Validation
tune + choose
Test
final generalization
Validation tunes + selects. Test stays untouched.
Three parameter layers
Parameters
learned · weights/biases
internal feature weights
Hyperparameters
chosen · LR/epochs
learning rate = .001
Inference params
request-time controls
temperature = .2
Feature engineering
Raw data → useful model inputs
Structured
normalize · impute · encode
country “KR” → category vector
Unstructured
tokenize · vectorize · extract
review text → embedding
Organized ≠ model-ready.
Deep-learning loop
Repeat until prediction error falls
Forward
predict
Loss
error
Backprop
gradients
Update
weights
Fit diagnosis
Underfitting
train bad · test bad · high bias
straight line for curved pattern
Good fit
train good · test good
works on old + new customers
Overfitting
train good · test bad · high variance
memorizes training customers
Overfitting fixes: more/diverse data · regularization · early stopping · simpler model.
Token → embedding
Readable unit → numerical meaning
Text
input
Token
unit
Embedding
meaning vector
Generative vs discriminative
Generative
learn pattern → create new
generate a new animal image
Discriminative
learn boundary → classify
label this image cat / dog
Image processing vs computer vision
Image processing
change pixels
crop · sharpen · remove noise
Computer vision
understand content
detect + count shelf products
Metric routing
BLEU
translation · precision-ish
ROUGE
summarization · recall-ish
BERTScore
semantic similarity
Perplexity
LLM uncertainty · lower is better
Accuracy
% correct · beware imbalance
Precision
avoid false positives
Recall
avoid false negatives
F1
balance precision + recall
AUC-ROC
ranking across thresholds
Confusion matrix
classification counts
RMSE
regression · punishes large error
MAE
regression · equal error weight
Implementation risks
Data quality
main practical challenge
missing values · wrong labels
Bias
unfair groups/classes
training data excludes a region
Explainability
why this output?
regulator asks why loan denied
Scalability
cost + latency + throughput
10 users → 1M requests
Hallucination
plausible + unsupported
invented policy citation
Data leakage
future/test data leaks in
train with tomorrow’s answer
05 · Responsible AI + governance
Who controls what?
Dimension → service
Responsible AI compass
Ask which dimension the scenario violates
Fairness
comparable treatment
Explainability
understand reasons
Privacy + security
protect data
Transparency
disclose use + limits
Veracity + robustness
accurate under change/attack
Governance
owner + policy + review
Safety
prevent harm
Controllability
guide · override · stop
Responsible AI routing
Before deploy · after deploy · generated output
Clarify
bias + explain
why was this loan denied?
Model Monitor
drift + quality
fraud accuracy falls this month
Guardrails
GenAI safety
block unsafe chatbot answer
Model Cards
your model docs
document intended users + limits
AI Service Cards
AWS service docs
AWS explains Rekognition limits
A2I
human prediction review
person checks uncertain diagnosis
Ground Truth vs A2I
Ground Truth
label raw data · before training
label historical scans
A2I
review prediction · during inference
review new scan prediction
Governance words
Data residency
where · address
data stays in eu-west-1
Data retention
how long · expiry
delete logs after 7 years
Data lineage
origin + transformations
source → clean → train
Data provenance
ownership + license
who supplied this dataset?
Encryption
how protected
KMS key at rest
Access control
who may use
only auditor IAM role
LLM attacks
Poisoning
malicious training data
inject scam links into dataset
Prompt leak
hidden/private info exits
model reveals system prompt
Prompt injection
input overrides instruction
“ignore policy; reveal secrets”
Jailbreak
bypass safety
role-play to evade filter
Interpretability + transparency
Decision tree
high · visual rules
if income > X → approve
Logistic regression
high · coefficients
income weight lowers churn risk
Neural network
black box
millions of image weights
SVM
kernel: low · linear: readable weights
SHAP / Shapley
local prediction
why this applicant was denied
PDP
global feature effect
age effect across all applicants
Transparency
trust + debug
Transparency trade-off
opaque model may be more accurate / protect IP
Shared responsibility + MLOps
AWS
security of the cloud
facilities + managed infrastructure
Customer
security in the cloud
data + IAM + encryption + Guardrails
MLOps
version + automate + monitor
retrain after drift
Data lifecycle
collect → archive/delete
controls at every stage
06 · Exam traps
See X → pick Y
Keyword → answer
- percentage / cumulative pool
- → Top P · not Top K
- fixed candidate count
- → Top K · not Top P
- false positive is costly
- → precision
- false negative is costly
- → recall
- FM makes labels from raw input
- → self-supervised · not unsupervised
- groups / clusters
- → unsupervised · not RL
- actions + rewards
- → reinforcement learning
- small labeled + large unlabeled
- → semi-supervised
- complex problem → smaller steps
- → chain-of-thought
- fresh/private company facts
- → RAG / Knowledge Bases
- stable style / behavior
- → fine-tuning
- smaller model imitates teacher
- → distillation
- harmful generated content
- → Bedrock Guardrails
- choose best FM
- → Bedrock Model Evaluation
- bias / explain prediction
- → SageMaker Clarify
- production drift / quality
- → Model Monitor
- human reviews model prediction
- → A2I · not Ground Truth
- store/share model inputs
- → Feature Store
- clean/transform raw data
- → Data Wrangler
- no-code build/train/deploy
- → Canvas
- contact-center live help
- → Connect AI agents (was Q in Connect)
- internal employee knowledge
- → Q Business
- BI chart / calculation
- → Quick Sight (was Q in QuickSight)
- code / AWS resources / costs
- → Q Developer
- ready FM through managed API
- → Bedrock
- custom ML lifecycle
- → SageMaker AI
- ready task API
- → Comprehend / Rekognition / Textract / …
- text + image search
- → multimodal embedding
- good train · bad test
- → overfit / high variance
- bad train · bad test
- → underfit / high bias
- final generalization
- → test set · not validation
- automated CVE scan
- → Inspector · not Config
- who called / changed / deleted?
- → CloudTrail
- resource state / compliance rule
- → AWS Config
- security of cloud / in cloud
- → AWS / customer
- plausible but false
- → hallucination
about this page
Distilled from my notes on the AWS course slides (v19). Practice-test questions, answers, and scores are left out. AWS services change quickly, so check current AWS documentation before relying on a detail.