Decoding generative AI
summary
Demo builds a searchable “government citizen assistant” that indexes CRA ESDC and National Student Loan Center pages, deploys a Falcon 40B on SageMaker Jumpstart endpoint in minutes, and returns sourced answers (mostly correct but missing phone/URL data until improved). Architecture uses RAG with retrieval + Bedrock for model choice and fine-tune with ~20–25 examples to fix errors (one wrong amortization answer fixed after fine-tuning). Also shows CodeWhisperer for code generation and Stability AI for on‑demand image creation.
AI-written summary · about
chapters (18)
- 0:00Introduction to generative AI
- 1:07Why GenAI matters for government
- 3:14Demo of government citizen assistant
- 8:14Use cases beyond Q&A chatbots
- 9:45What is generative AI
- 11:34Foundation models explained
- 13:02SageMaker Jumpstart overview
- 14:05Deploying Falcon 40B model demo
- 17:45Introduction to Amazon Bedrock
- 19:57Legacy app integration with GenAI
- 21:21Text summarization with Titan
- 25:16Fine-tuning models for accuracy
- 28:49RAG technique and code development
- 30:37CodeWhisperer for developers
- 35:20Connecting legacy applications
- 38:11Generating images with Stability AI
- 41:01Reflections and future of GenAI
- 43:51Q&A and closing remarks
related talks
28mImproving responses in generative AI applications
37mBuilding Generative AI Applications
19mHow to use generative AI to deliver self-service data products
48mHow to Make Smart Architecture Decisions when Building Gen AI Apps
20mExplore what’s possible with new AWS generative AI services
36mGenerative AI + Python: Unlocking Efficiency, Personalization, and Insight
12mCodeless Generative AI Pipelines: Prompt Automation
27mBuilding Generative AI features