RESPONSIBLE AI

Question 1

Remember ML AI and GenAI

Answer 1

AI > Machine Learning > Deep Learning > GenAI

Question 2

What is tokenization?

Answer 2

converting raw text into a sequence of tokens, can be word-based or sub-word based

Question 3

what is a context window?

Answer 3

the number of tokens an LLM can consider when generating text, the larger the context window, the more information and coherence (at the cost or memory and power) - first factor when choosing a FM

Question 4

what is an embedding model?

Answer 4

a model that creates a vector for each token. converting tokens into models encodes many features for one input token (information about that word) and stores it into a high-dimensionality vector (used for vector databases and RAG)

Question 5

is there a semantic relationship between tokens with similar embeddings?

Answer 5

yes, that’s why we use them. embedding models can be easily searchable so it’s a good idea to use an embeddings model to power a search application

Question 6

what are the basic components of AI?

Answer 6

Data layer (in vast ammounts)
ML framework and algorithm layers
Model layer (implement and train it)
Application layer

Question 7

what is responsible AI?

Answer 7

making sure that systems are transparent and trustworthy, mitigating potential risk and negative outcomes

Question 8

what is secure AI?

Answer 8

ensuring that confidentiality, integrity and availability are maintained on organization data and information assets

Question 9

what is AI governance?

Answer 9

ensuring that we can add value and manage risk in the operation business; clear policies, guidelines and oversight mechanisms to ensure AI systems align with legal and regulatory requirements - IMPROVE TRUST

Question 10

what is AI compliance?

Answer 10

ensuring the adherence to regulations and guidelinew, specially for sensitive domains such as healthcare, finance, and legal applications

Question 11

what are the pillars of responsible AI?

Answer 11

fairness
explainability
privacy and security
transparency
veracity and robustness (reliability)
governance (responsible AI practices)
safety
controllability (align with human values)

Question 12

what are AWS services for implementing responsible AI?

Answer 12

Bedrock (human or automatic model evaluation)
Guardrails for Bedrock
Sagemaker Clarify
Sagemaker Data Wrangler
Sagemaker Model Monitor
A2A (Amazon Augmented AI - human review of ML predictions)
and for governance: Sagemaker Role Manager, Model Cards and Model Dashboard

Question 13

what are AWS AI Service Cards?

Answer 13

responsible AI documentation

Question 14

what are the capabilities of GenAI?

Answer 14

adaptability
responsiveness
simplicity
creativity and exploration
data efficiency
personalization
scalability

Question 15

what are the challenges of GenAI?

Answer 15

regulatory violations
social risks
data security and privacy concerns
toxicity
hallucinations
interpretability
nondeterminism
plagiarism and cheating

Question 16

what are toxicity and hallucinations? and what are their mitigations?

Answer 16

toxicity: generating offensive, disturbing or innapropriate content
mitigation: curate training data - guardrail

hallucinations: assertions or claims that sound true, but are incorrect (because of next-probability sampling)
mitigation: educate users about fact checking, mark generated content as unverified

Question 17

what are the prompt misuse methods?

Answer 17

poisoning: intentional introduction of malicious or biased data intontue training dataset

hijacking and prompt injection: influencing outputs by embedding specific instructions within the prompts

exposure: of sensitive or confidential information

prompt leaking: unintenional disclosure or leakage of prompts or inputs used within a model - can expose protected data

jailbreaking: many-shot jailbreaking

Question 18

what are the AWS tools for governance?

Answer 18

aws config
aws inspector
aws audit manager
aws artifact
aws cloudtrail
aws trusted advisor