AWS Cloud Experience 2023

AI & Machine Learning

• In recent years, usage of Machine Learning increased in quantity and cost.

• Service: Amazon SageMaker
• From training to production

• Top 3 challenges of AI adoption in enterprises (source: O’Reilly’s AI Adoption in the Enterprise Survey)
• Difficulty of hiring qualified people for the required roles
• Lack of quality data
• Difficulties in identifying appropriate use cases

• Major use-case groups for Machine Learning
1. Improving end user experience
2. Optimizing business
3. Accelerate innovation

• AWS CCI (Contact Center Intelligence) Solutions
• Service: AWS Transcribe
• Transforms audio into text. That can be used to run analytics or change the behavior of your system from the audio of a support call.
• Service: Amazon Personalize
• Easily segment users and improve your recommendation system based on data from your systems.

• Service: Amazon Kendra
• Search engine powered by ML that uses multiple data-sources of structured or unstructured data. Can even connect with third-party services, such as Confluence, Salesforce, Slack.

• Service: Amazon Textract
• Extracts text from printed text, handwriting, and data from any document.

• Service: Amazon Bedrock
• Lowers the complexity of starting to use Foundational Models.
• There are a handful of options, from AI startups and Amazon themselves. Amazon’s FMs are Titan Text and Titan Embeddings. The former, is a generative LLM similar to GPT. The latter, translates inputted text into embeddings for future use.
• Both Titan FMs are built to reject harmful content. Filtering both the input and output.

Generative AI

• (Should I keep this?) Currently, Machine Learning is at an inflection point.

• Some AI startups that are in partnership with AWS:
• Cohere: is focused on building ML models for enterprises. They have models for retrieving text, generating text and classifying text. The retrieving text counts with a semantic search that leverages the embedding to make search on private data.
• AI21 Labs: has two major products for helping with writing and reading text.
• Anthropic: has Claude, an AI assistant with capabilities similar to ChatGPT’s. But Cloud has some more specific features for assistance, such as how it should respond a request. Anthropic also makes a lot of their research available.
• Stability AI: focuses on open-source models, and their biggest one yet is Stable Diffusion. They are involved in many different fronts of AI research.

• AWS has a variety of foundational models that include: text-to-text, text-to-embedding, and multimodal.

• Challenges with different kinds of FMs:
• When using proprietary models (such as GPT-n): can we trust that our company data is safe? Is the owner of the model not using our data for improving themselves, and we end up losing competitively?
• When using open-source models: do we have the infrastructure to publish the model?
• When creating your own FM: do we have the staff and data to train the FM?

• When using AWS for your AI needs:
• With Amazon Bedrock, you have an easier way of creating proprietary FMs, in a safe and isolated way, and you can leverage already existing FMs created by AWS or partner startups. You can also make a fine-tuning of those models.
• AWS CodeWhisperer is a code assistant similar to GitHub Copilot. It accepts plugins to enhance its features and can run without internet access (probably with limited capabilities).

• What are the differences between Amazon SageMaker VS Amazon Bedrock?
• With SageMaker models are created in Notebooks. In Bedrock, there is a UI that abstracts this complexity. You also won’t need to worry about infrastructure management.

Databases

• Service: Amazon Aurora
• Relational database, architected by AWS for high availability.
• Automatically replicated between different AZs, even if using only one node.

• Service: DynamoDB
• Key-value database.
• Latency of 10ms, when using it the recommended way.

• Service: DocumentDB
• Similar to MongoDB.

• Service: Amazon Keyspaces
• Wide-column store, similar to Cassandra. A wide-column store uses keyspaces with flexible columns and distributes data across multiple nodes. Think of keyspaces a schema and a column as a table, in relational databases.

• Service: Amazon QLDB (Quantum Ledger Database)
• Ledger database.
• It is fully auditable and provides a transparent, immutable, and cryptographically verifiable transaction log.
• Data is never deleted, only new versions of data are created.

Networking and Global Architecture

• Important to know the basics of AWS networking:
• a zone (e.g. us-east-1) is the largest physical part.
• Inside the zones, there can be many VPC (Virtual Private Cloud), which are just logical isolations.
• Each VPC can have many subnets (logical), and those subnets belong to availability zones (physical).

• When using the Internet Gateway to access AWS resources in the same AWS Region, you won’t have additional charges. Altough, you use the NAT Gateway, there will be a data processing charge.
• The Internet Gateway is a component of the VPC used to access the Internet. Resources on the internet can initiate connection to your public subnet resource.
• The NAT Gateway is used by a private subnet to access resources outside your VPC. It only works from the subnet to the outside. An outside resource cannot make inbound requests through the NAT Gateway.

• Be extra careful of communication between services within different VPCs. If they communicate via public IP, it will be more expensive.

• Service: AWS Transit Gateway
• Connects VPCs and on-premise infrastructure.
• There is no charge for communication between nodes within a VPC and different AZs. (Since April)

• Service: AWS Site-to-Site VPN
• Creates a secure connection between your on-premise infrastructure and AWS resources.

• Feature of AWS VPC: VPC Peering
• Allow the communication of resources in one VPC with the resources in another one. The VPC can be on different Regions as well as different AWS accounts.

• Service: AWS Direct Connect
• You can create a dedicated connection within AWS’s network. While in transit, your data stays on AWS’s network.

• Service: AWS PrivateLink
• Connects VPCs to AWS resources without exposing to the internet.
• Good for accessing resources that are not in VPCs (S3, CloudWatch, ...).

• You can use the Cost Explorer service to find spots in your network’s architecture that could bring some cost reduction for your cloud expenses.

• Feature in EC2: Elastic IPs
• Reserves an static IPv4 on you AWS account. That address will never cease to be from your account. You can stop or drop the EC2 instance related to that IP that the IP will still be on your account.