Fundamentals of ML and AI
AWS Infrastructure and Technologies
AWS organizes its AI and ML offerings into three layers, from most control to most ready-made.
The AWS AI/ML stack
ML frameworks layer
- Amazon SageMaker AI (the new name for Amazon SageMaker) is a fully managed service to build, train, and deploy your own ML models.
AI/ML services layer (pre-trained, no ML expertise needed)
| Service | Use it to |
|---|---|
| Amazon Comprehend | Run NLP on text: detect language, extract key phrases and entities, analyze sentiment, and organize documents by topic |
| Amazon Translate | Translate text with neural machine translation |
| Amazon Textract | Extract text, form fields, and tables from scanned documents (goes beyond OCR) |
| Amazon Lex | Build chatbots and voice bots (IVR) with speech recognition and natural language understanding, the technology behind Alexa |
| Amazon Polly | Turn text into lifelike speech |
| Amazon Transcribe | Turn speech into text, in batch or real time, with timestamps per word |
| Amazon Rekognition | Analyze images and video: objects, people, text, scenes, faces, and inappropriate content |
| Amazon Kendra | Run intelligent enterprise search across many content repositories |
| Amazon Personalize | Create real-time, individualized recommendations from user activity |
| AWS DeepRacer | Learn reinforcement learning with a 1/18th scale autonomous race car |
Generative AI layer
- Amazon SageMaker JumpStart offers ready-made solutions and one-click deployment and fine-tuning of popular open-source models.
- Amazon Bedrock is a fully managed, serverless way to use FMs from Amazon and leading AI companies through a single API, and to customize them privately with your data.
- PartyRock is an Amazon Bedrock playground for experimenting with generative AI apps.
- Amazon Q is a generative AI assistant for work that uses your company's data.
- Amazon Q Developer generates code and code recommendations inside IDEs.
Why build AI on AWS
- Faster development: managed services handle infrastructure, training, and deployment.
- Scalability and cost: pay-as-you-go pricing and global infrastructure.
- Model choice: many FMs behind one API in Amazon Bedrock.
- Integration: AI services work with the rest of AWS through SDKs and APIs.
- Security and compliance: the shared responsibility model and compliance programs.
Cost trade-offs
- Responsiveness and availability: lower latency and multi-Region deployments cost more.
- Redundancy and Regional coverage: multiple Availability Zones or Regions add cost.
- Performance: GPUs and accelerators cost more than CPUs but can be much faster.
- Token-based pricing: Amazon Bedrock and Amazon Q Developer charge per token processed or generated.
- Provisioned throughput: paying up front for capacity costs more but gives predictable performance.
- Custom models: training and fine-tuning add compute and data costs.