AWS re:Invent 2024 - AI/ML
Back to all Re:Invent 2024
Table of Contents
AI & ML
 |
This talk discusses how AWS Bedrock can help scale generative AI workloads by providing intelligent prompt routing, model distillation, and access to a marketplace of diverse models. The speakers showcase customer use cases and the benefits of these Bedrock features in optimizing for quality, cost, and speed of generative AI responses.
|
 |
The talk provides a comprehensive overview of responsible generative AI evaluation, focusing on establishing launch confidence through a structured evaluation strategy that considers specific use case risks, chooses relevant metrics, and interprets results with statistical rigor to ensure reliable deployment of generative AI systems.
|
 |
This video discusses the challenges of training large-scale AI models and how Amazon SageMaker HyperPod and training plans can help address these challenges. It also introduces SageMaker HyperPod recipes, which simplify the process of customizing and fine-tuning foundation models, and showcases how a startup, NinjaTech AI, is using these capabilities to build innovative generative AI applications.
|
 |
This presentation covers the latest advancements in Amazon's generative AI models, including the Amazon Nova Canvas for image generation and the Amazon Nova Reel for video generation. The presenters also discuss the future roadmap for the Nova family of models, which will soon include speech capabilities and the ability to handle multi-modal inputs and outputs.
|
 |
The video introduces the new Amazon Nova understanding models, which offer state-of-the-art performance, speed, and cost-effectiveness for a range of tasks. The models showcase impressive capabilities in areas like visual understanding, instruction-following, and agent-based automation, with potential applications across various industries.
|
 |
This session discusses how AWS SageMaker HyperPod, a distributed training service, is helping companies like Hugging Face, Writers, and HOPPR build and scale their foundation models. The panelists share their experiences on how HyperPod's features like resiliency, flexibility, and visibility have enabled them to innovate and deliver better AI solutions to their customers.
|
Compute
 |
The video discusses the launch of Amazon SageMaker HyperPod Task Governance, which enables dynamic allocation of compute resources, task prioritization, and real-time monitoring to maximize utilization and reduce costs for AI workloads. The solution addresses challenges faced by companies like Amazon and Articul8 in efficiently managing and allocating GPU resources for their growing generative AI initiatives.
|
 |
AWS re:Invent 2024 features a session on training large models on Amazon SageMaker for scale and performance. The session covers how SageMaker HyperPod can reduce training time by up to 40%, provide resiliency features, and offer flexibility for customizing the computing environment.
|
 |
The presentation discusses the challenges faced by the Amazon AGI team in developing large-scale foundation models, such as the memory and compute walls, and how they have addressed these challenges by leveraging parallelism strategies and building resilient infrastructure using Amazon SageMaker HyperPod. The presentation also highlights how SageMaker HyperPod can help customers reduce the time and effort required to manage infrastructure and build their own foundation models.
|
 |
This session discusses how Amazon Bedrock can help enterprises deploy cost-optimized and scalable generative AI workloads. The presentation covers key considerations for enterprise readiness, including compliance and governance, cost and performance management, and operational excellence, with detailed examples of how Bedrock addresses these requirements.
|
 |
This session explores how Amazon SageMaker can help reduce deployment costs and latency for foundational machine learning models. The presenters discuss features like multi-model endpoints, speculative decoding, and quantization that can optimize performance and cost, and they also showcase a customer success story from Capital One.
|
Databases
 |
The presentation discusses how Amazon Bedrock Knowledge Bases can help businesses unlock the power of structured data by providing a fully-managed service for retrieval-augmented generation (RAG) workflows. It highlights the challenges of text-to-SQL conversion and how Bedrock Knowledge Bases addresses them through personalization, context-awareness, and security features.
|
Developer Experience
 |
This session explores the latest advancements in Amazon Q Business, a generative AI-powered assistant designed to enhance productivity, efficiency, and decision-making for enterprise customers. The presentation covers how Amazon Q Business can streamline workflows, automate repetitive tasks, and provide data-driven insights, while maintaining robust security and privacy controls to meet the needs of regulated industries.
|
 |
The presentation introduces automated reasoning checks in Amazon Bedrock Guardrails, a tool that helps build accurate and transparent AI applications by validating the factual accuracy of language model outputs. The speakers discuss the challenges of hallucinations in large language models, the principles of automated reasoning, and how it can be used to improve the reliability of AI-powered applications.
|
 |
The video discusses how software companies can leverage generative AI to improve operational efficiency and embed it into their products to drive differentiation and growth. Key themes include balancing accuracy, cost, and performance, mastering data management, and building the right organizational capabilities to succeed with generative AI.
|
 |
This presentation covers the new evaluation features in Amazon Bedrock, including LLM-as-a-judge for model evaluation and RAG (Retrieval Augmented Generation) evaluation for knowledge bases. The speakers demonstrate how these tools can help streamline the evaluation process and provide detailed insights to improve the quality and responsible AI aspects of generative AI applications.
|
 |
Principal Financial Group's journey with Amazon Q Business showcases how the company leveraged generative AI to improve productivity and customer experiences by addressing challenges around fragmented data sources, legacy systems, and user education. The presentation highlights Principal's approach to deploying Q Business, overcoming obstacles, and driving quantifiable benefits through user feedback and a strong governance framework.
|
 |
The video discusses the importance of responsible AI practices, including addressing issues like bias, fairness, explainability, and safety. It outlines the various tools and methods available on AWS, such as SageMaker Clarify, Bedrock Guardrails, and Automated Reasoning Checks, to help organizations implement responsible AI in their applications.
|
 |
This talk introduces Amazon Q, a set of features and APIs that empower software vendors to enhance their applications with generative AI capabilities that leverage cross-application data. The presentation showcases how vendors can embed the Amazon Q experience in their apps or leverage the Amazon Q index to build richer, more contextual experiences for their users.
|
 |
This session showcases how KONE, a leading elevator and escalator company, has partnered with AWS to develop a generative AI-powered Technician Assistant application that helps field service engineers quickly diagnose and resolve equipment issues. The presentation highlights the use of Amazon Bedrock Guardrails to ensure the safety and reliability of the AI-powered application, addressing challenges such as content filtering, prompt attacks, and hallucination detection.
|
 |
The video showcases the features and capabilities of AWS App Studio, a low-code platform that leverages generative AI to enable rapid application development. It highlights how the NFL used App Studio to streamline their player headshot management process and demonstrates the platform's ability to quickly generate and customize applications with minimal coding.
|
 |
This video demonstrates how Amazon SageMaker Studio can accelerate machine learning workflows by providing a unified interface for data preparation, model customization, evaluation, deployment, and monitoring. The presentation covers key capabilities such as data discovery, Jupyter Notebook integration, JumpStart model hub, MLflow integration, and the newly announced SageMaker Partner AI Apps, which enable secure and private access to industry-leading AI and ML development tools.
|
 |
The presentation discusses the challenges of large-scale distributed model training and introduces Amazon SageMaker HyperPod, a managed infrastructure that provides optimized performance and resilience for training foundational models. The speakers also demonstrate SageMaker HyperPod Recipes, a set of pre-configured training configurations that simplify the process of pre-training and fine-tuning large language models on AWS.
|
 |
The video discusses the benefits of fine-tuning language models, such as improving performance on specific tasks, reducing token usage, and adapting the model to the desired use case. It showcases examples of fine-tuning Anthropic's Claude 3 Haiku and Meta's Llama models, highlighting the importance of data preparation, hyperparameter tuning, and the features offered by Amazon Bedrock for model customization.
|
 |
The session discusses the security and privacy considerations when building generative AI applications on Amazon Bedrock, a fully managed service. The presenters cover various security features of Bedrock, including data protection, connectivity, access control, and observability, as well as patterns and best practices for building secure generative AI applications.
|
 |
This video introduces Amazon Bedrock Agents, a fully managed service that enables the creation and deployment of intelligent software agents. The presentation showcases the capabilities of Bedrock Agents, including the new multi-agent collaboration feature, and demonstrates how it can be used to unify customer experiences and automate complex business processes.
|
 |
The presentation covers the journey of building scalable Retrieve-Augment-Generate (RAG) applications using Amazon Bedrock Knowledge Bases, highlighting best practices, challenges, and new features like structured data retrieval and GraphRAG. The presenters emphasize the importance of starting small, continuous improvement, and partnering with AWS to shape the product through feedback and new use cases.
|
 |
This AWS re:Invent 2024 session discusses how Amazon SageMaker Canvas can accelerate machine learning innovation without writing code, enabling organizations to build high-quality models and realize tangible business benefits, as demonstrated by the success story of Gosoft Thailand's demand forecasting solution.
|
 |
The video showcases the launch of Amazon Nova, a new foundational AI model from Amazon, and its various capabilities across understanding and generation tasks. The presenters demonstrate how Amazon Nova is being leveraged by different Amazon teams, such as AWS Support, Prime Video, and Amazon Ads, to drive efficiencies and enhance customer experiences.
|
 |
The talk discusses how to power a cost-effective Retrieval Augmented Generation (RAG) solution using Amazon Titan Embeddings Text (AIM358). It covers the introduction to embeddings, the benefits of using Titan Embeddings, and an architectural approach to leverage binary embeddings for efficient storage and performance optimization.
|
 |
The video discusses how AWS Bedrock is helping companies like Forcura and Cencosud build and scale generative AI applications. It highlights the key features of Bedrock, such as model choice, corporate data integration, responsible AI, and developer experience, and showcases real-world use cases and results from the two companies.
|
 |
The presentation explores the evolution of large language models (LLMs) and their ability to automate work using function calling and agents. It discusses the origins of function calling, best practices for tool use, and the complexities around defining agents, highlighting the trade-offs between agency, effect, and orchestration styles.
|
 |
This session explores the return on investment (ROI) of generative AI, with a focus on how companies are measuring and achieving success with this technology. Executives from various organizations share their strategies, challenges, and insights on balancing the initial investment with long-term gains, choosing between custom and pre-trained models, and organizing stakeholders to ensure alignment and strong ROI.
|
 |
This session covers how to accelerate production for generative AI using AWS SageMaker MLOps and FMOps capabilities. It showcases how Rocket Mortgage has successfully leveraged these capabilities to streamline their ML workflows, increase model deployment velocity, and enhance customer experiences through conversational AI.
|
 |
This session explores how AWS Generative AI, particularly Amazon BedRock, can boost productivity and empower revenue teams. The presentation covers technical and business challenges, best practices, and a real-world example of Showpad's successful integration of Generative AI to enhance user experiences and streamline operations.
|
 |
This video showcases how Volkswagen Group of America leveraged Amazon Q Apps, a capability within Amazon Q Business, to automate a critical job mapping task for a global HR system implementation. The video highlights the ease, speed, and high return on investment Volkswagen achieved by using generative AI to address their challenge, demonstrating the transformative potential of Q Apps for enterprises.
|
 |
The session discusses how the NFL leveraged Amazon Q Business, a generative AI-powered assistant, to streamline content production workflows and knowledge management, leading to significant time savings and improved productivity for their content creators. The presentation highlights the ease of deployment, flexibility in data source integration, and the importance of user training in driving successful adoption of the solution.
|
 |
Amazon Q Business is a generative AI-powered assistant that helps make organizational data more accessible, enabling users to get answers to questions, generate content, and take actions in third-party applications. The session explores how Amazon Q Business can boost employee productivity by uniting data sources, delivering quick and accurate answers, executing actions on behalf of users, and streamlining daily tasks with user-created lightweight applications.
|
 |
The video discusses how AWS customers can customize foundation models using Amazon SageMaker, including when to fine-tune models, how to prepare data, and the capabilities SageMaker offers to simplify the process. The presenters also provide a demo and discuss Intuit's use case of fine-tuning foundation models to improve the accuracy and efficiency of their transaction categorization for small businesses.
|
Operations
 |
Jabil, a leading global manufacturing company, is revolutionizing its supply chain and operations through the strategic use of generative AI, including Amazon Q Business, to improve productivity, efficiency, and decision-making across its vast network of 400 customers, 38,000 suppliers, and 140,000 employees worldwide.
|