Foundation Model

A Foundation Model is a large-scale artificial intelligence model—typically a neural network with billions of parameters—that is trained on a massive quantity of unlabeled data at scale. Unlike narrow AI models designed to do one specific thing (like identifying cats in photos or predicting housing prices), foundation models are generalists. Because they have developed a deep, underlying representation of human language and logic during their initial training, they can be adapted (e.g., through prompting or fine-tuning) to perform a wide variety of downstream tasks they were not explicitly programmed to do.

Prominent examples of foundation models include OpenAI's GPT-4, Anthropic's Claude 3, and Meta's Llama 3. For the enterprise, foundation models represent a massive paradigm shift. Instead of building a separate machine learning model for sentiment analysis, another for translation, and another for summarization, an enterprise can route all of these tasks through a single foundation model via an API. This drastically reduces the time-to-market for new AI capabilities.

However, relying on foundation models introduces vendor lock-in and data privacy concerns. Because training a foundation model costs tens of millions of dollars, most enterprises must rent access to them via cloud APIs. A robust Model Governance strategy is required to ensure that the enterprise can dynamically route traffic between different foundation models based on cost, capability, and changing Terms of Service, preventing reliance on a single vendor.