Generative AI Flashcards
A precursor of Transformers models are the encoder-decoder RNN architectures. Explain what they are.
Standard Encoder-Decoder architecture send the sequence of input tokens through a sequential RNN that processes each token sequentially, the encoder, getting to a intermediary state that is then passed to decoder model that generates the output sequentially in the same way.
What are the main disadvantages of the standard Encoder-Decoder RNN architectures?
-Since processing is sequential, it cannot be parallellized easily
-Because th final input state is a single vector, information from the start of a sentence can easily get lostW
What is the main difference between a standar Encoder-Decoder CNN and a transformer?
On a transformer, you generate a hidden state for each word that takes into account other words in the phrase and substitute RNN for Fee-forward Networks, making parallellization easier
True or False: Self-attention is a mechanism that produces a weighted average of token embeddings by computing the weights trough the analysis of which tokens are more strongly related to each other
True
What matrices are trained when training a Transformers model?
-Query
-Key
-Value
How is each transformer matrix used?
-Each token embedding is multiplied by these matrices, generating new vectors
-Each Query token is dot multiplied by each Key token, generating a matrix that indicates how much each word is related to others in the sentence
-This Matrix is normalized (division by root of size and softmax)and multiplied by the Value vectors
What is Masked Self-Attention?
It is the same as self-attention, but each word can only look at words that came before it, not every word of the sentence
What is multi-headed self-attention?
It is the same as self attention, but you have multiple Query, Key and Value matrices you use to perform the attention process and then you merge all the results into a single vector.
What are some Foundation Models available on AWS?
-Claude
-Jurassic
-Stable Diffusion
-Titan
What is Sagemaker Jumpstart? can it be used for GenAI?
It is a feature from Sagemaker Studio that allows you to load up a notebook with a model ready to go. Many of those models involve GenAI
What are some types of model that can be loaded using Sagemaker Jumpstart?
-Huggingface
-Alexa
-Stable Diffusion
What is Amazon Bedrock?
An API for using Prebuilt GenAI Models
What is Amazon Code Whisperer?
It is an AI coding companion that sugests code for you in real time.
What are Amazon Code Whisperer’s features?
-Real-time code suggestion: Code is written automatically based on comments
-Security Scans: Scans your code for vulnerabilities
-Reference tracker: Flags suggestions that are similar to open source code
-Bias avoidance: Filters out code suggestions that might be biased or unfair
-AWS Code Integration: Suggests code for interfacing with AWS APIs
What programming languages are compatible with Code Whisperer Security Scan?
-Java
-Python
-JavaScript
What are Amazon Code Whisperer billing options?
-Individual: Free to use
-Professional: 19$ / user / month
What are Amazon Code Whisperer Professional’s benefits?
-Allows management which developers in an organization has access
-Authenticated with IAM Identity
Center
-Allows 500 security scans / user / month (Individual allows only 50)
What are the existing Bedrock endpoints?
-Bedrock
-Bedrock-runtime
-Bedrock-agent
-Bedrock-agent-runtime
What is the new name for Amazon Code Whisperer?
Amazon Q Developer
What is AWS HealthScribe?
It is a Speech-to-text solution that has been tuned to healthcare
Mark True or False for each statement regarding AWS HealthScribe:
-HealthScribe can generate clinical notes for the transcription
-HealthScribe is HIPAA compliant
-HealthScribe creates a new transcript when run, so it cannot reference anything from the original
-It can skip small talk and silence
-Patient recording and notes are stored on S3 by default
-True
-True
-False, it does not create a new transcript, so it can reference the original
-True
-False, only stored in the cloud if client puts it there
