Title

AWS re:Invent 2023 - Building applications with generative AI and live data (BOA307)

Summary

Presenters Memo Doering and Mike Chambers discussed building applications using generative AI and live data.
They showcased a demo involving a webcam, a toy elephant, Amazon Rekognition, and Amazon Bedrock, which combined multiple large language models (LLMs) to generate stories.
The demo faced challenges due to different lighting conditions, which affected the custom-trained recognition model's ability to detect the blue elephant toy.
They discussed the architecture of their application, which used AWS services like Lambda, Step Functions, and S3, and highlighted the importance of not storing sensitive data.
The session included live polling to gauge audience engagement and opinions on generative AI.
They demonstrated the use of Amazon Bedrock's playgrounds for experimenting with different models and prompts.
The presenters emphasized the importance of prompt engineering and the ability to use live data in conjunction with LLMs.
They mentioned the recent addition of direct Bedrock service calls from Step Functions, eliminating the need for intermediary Lambda functions.
The session concluded with a call for audience input on what they would like to explore next in generative AI.

The presenters highlighted the rapid pace of development in generative AI, noting that even during the week of re:Invent, new models and capabilities were released.
They emphasized the non-deterministic nature of LLMs and the importance of configuring models to balance creativity with accuracy.
The session showcased the practical application of generative AI in a hands-on, interactive demo, demonstrating the potential for engaging customer experiences without requiring technical expertise.
The architecture of their application was designed with privacy in mind, avoiding the storage of images and sensitive data.
The presenters discussed the use of AWS Step Functions for orchestrating the application flow and the benefits of the recent capability to call Amazon Bedrock services directly.
The session demonstrated the use of Amazon Bedrock's playgrounds as a starting point for developers to experiment with different models and prompts before building applications.
The presenters introduced PartyRock, a no-code platform for experimenting with generative AI, as a resource for those new to the field.
Audience engagement through live polling and the call for input on future explorations indicated a collaborative approach to learning and innovation in the generative AI space.