Foundational models (FMs) are large-language models (LLMs) and multi-modal models that enable various capabilities, such as code generation and image creation from natural language descriptions. They can be easily customized for specific use cases like summarization and translation without needing to build a new machine learning model from scratch. FMs are pre-trained using vast amounts of unlabeled data, allowing them to deliver strong performance across multiple tasks.

Organizations can customize foundational models through several methods: zero-shot learning, where users provide prompts to interact with the model; in-context learning, where examples are included in prompts to enhance output relevance; and fine-tuning, which involves training the model with a small number of labeled examples to tailor it for specific tasks. This approach is efficient and requires significantly less labeled data than developing a model from scratch.

AWS provides several benefits for building generative AI applications, including enterprise-grade security and privacy, access to leading foundational models, and a data-first approach. Additionally, AWS offers a highly performant and cost-effective infrastructure, enabling organizations to train their models and run inferences at scale. This allows businesses to leverage their data as a strategic asset to create customized and differentiated generative AI solutions.