AWS has announced the availability of the new Well-Architected Generative AI Lens, which focuses on providing best practices for designing and operating generative AI workloads. The lens is aimed at business leaders, data scientists, architects, and engineers responsible for delivering robust and cost-effective solutions using generative AI. The document offers cloud-agnostic best practices, implementation guidance, and links to additional resources.
The Generative IA Lens addresses responsible AI, acknowledging the new challenges presented by the emergence of AI-driven capabilities and outlining a set of considerations for customers to review and address. The paper highlights the need for ensuring veracity and robustness (i.e., achieving correct system outputs, even with unexpected or adversarial inputs) as particularly important, compared to traditional machine learning solutions.
The lens promotes an iterative process for the design, delivery, and operation of generative AI solutions. The six phases of the generative AI lifecycle include scoping the impact, selecting and customizing the model, integrating the model into existing applications, and deploying the new AI-driven capability. The final phase involves iterating and improving the capability, closing the iterative loop.
The six phases of the generative AI lifecycle (Source: AWS Architecture Blog)
The paper also covers additional challenges to the data architecture posed by delivering generative AI solutions. The document focuses on three primary use cases: model pre-training, model fine-tuning, and retrieval-augmented generation (RAG). Each of these use cases has different requirements, but overall, they demand mature and adaptable approaches that can support large datasets and complex infrastructure footprints.
The authors of the announcement highlight the value offered by the lens:
The Generative AI Lens provides a consistent approach for customers to evaluate architectures that use large language models (LLMs) to achieve their business goals. This lens addresses common considerations relevant to model selection, prompt engineering, model customization, workload integration, and continuous improvement.
The document covers all six pillars of the Well-Architected Framework and discusses many areas specific to delivering generative AI solutions. It also offers a set of design principles for generative AI workflows created on AWS, specifically highlighting the need for controlled autonomy, which is particularly relevant to AI workloads.
Danilo Poccia, Chief Evangelist (EMEA) at AWS, summarized the announcement in his X post:
The lens emphasizes responsible AI practices with clear dimensions for fairness, explainability, privacy, safety, and transparency, recognizing the shared responsibilities between model producers, providers, and consumers.
