Building truly impactful artificial intelligence systems extends far beyond merely choosing a large language model. It demands a sophisticated understanding and deliberate construction of a multi-layered technology stack, as illuminated by Lauren McHugh Olende, Program Director at IBM, in her recent presentation on "What Is an AI Stack?" Her insights reveal that the efficacy, cost, speed, and safety of an AI solution are inextricably linked to the strategic choices made across its entire architectural foundation.
McHugh Olende frames the AI stack as five interconnected layers essential for turning raw AI into real-world applications, using an AI drug discovery tool as an example.
The Five Layers of the AI Stack
- 1. Infrastructure: The foundational layer, demanding AI-specific hardware (GPUs). Choices range from on-premise to cloud or local execution, impacting cost, accessibility, and processing power.
- 2. Model: The core AI engine. Builders must choose based on open/proprietary status and size, selecting from large, general LLMs or smaller, specialized models tailored for tasks like reasoning or code generation.
- 3. Data: The system's intellectual fuel, often requiring external data sources and pipelines to overcome model knowledge cut-offs. Crucially, vector databases and RAG (Retrieval Augmented Generation) ground the model in current, relevant data, minimizing hallucinations.
- 4. Orchestration: The layer for complex problems, involving breaking user queries into iterative steps like "thinking" (planning), "execution" (tool/function calling), and "reviewing" (self-critique). This iterative process enhances reliability.
- 5. Application: The user-facing interface. It defines human interaction, accommodating diverse inputs (text, image, audio) and critical features like revisions and citations for trust. Integrations are vital for embedding the AI into existing workflows.
McHugh Olende stresses that a holistic understanding of this entire stack is critical for builders to make practical, informed choices and design reliable, effective AI systems aligned with real-world needs.
