Practical AI hosts terminology primer on models, agents, and multi-agent systems

In this episode of the Practical AI Podcast, hosts Daniel Whitenack and Chris Benson tackle the evolving terminology of AI to help listeners stay current with rapid industry changes. Whitenack, CEO at Prediction Guard, and Benson, a principal AI and autonomy research engineer at Lockheed Martin, decided to pause their usual guest-interview format for several recent episodes to provide educational deep-dives on concepts that are shifting in meaning as the field accelerates.

The first major concept is the AI model itself. Whitenack frames it as a data transformation or software function: you provide an input and receive an output. A large language model takes text and generates text; an object recognition model takes an image and identifies what is in it; an image generation model takes text and produces an image. Models can combine different input and output types—text plus image, for instance—and can be composed of various internal architectures. The key insight is that regardless of the technical machinery inside, a model is fundamentally a transformation of data.

Benson adds an important architectural context: models have grown more complex over the years, but they remain neural network-based architectures at their core. Keeping in mind that models are functions—not magic—is essential, especially as mass media hype surrounds AI. The function takes an input and produces an output called an inference. Understanding this prevents misplaced expectations about what AI systems can and cannot do.

A critical distinction is whether a model is open-weight or closed. Whitenack explains that running a model requires two components: software that executes the function, and weights (numerical parameters that configure how the software behaves). With an open-weight model like Gemma, both components are released publicly. Users can download the weights from platforms like Hugging Face, use open-source software like the Transformers library, and run the model locally on their own laptop or cloud environment. Closed models—such as those from major US providers—keep either the software or the weights (or both) proprietary. Users access them only through a managed API or chat interface sitting behind the company's proprietary infrastructure.

Benson highlights a geopolitical and economic dimension: frontier models—the largest and most sophisticated—require enormous data centers with racks of expensive hardware. US-based cloud providers and model creators have made these investments and maintain leadership in frontier models accessed via API. However, if an organization wants to run a sophisticated model on its own servers without paying for cloud infrastructure, open-weight models are the path. Notably, Chinese models are currently leading in the open-weight space and represent the best option for on-premises deployment, while American models still lead through cloud APIs. Whitenack agrees: "They're the best right now." This reflects a current market reality where openness and local runnable capability versus frontier sophistication and cloud access represent different regional strengths.

The hosts indicate they will continue the discussion into agents and agent harnesses—the next layer of abstraction built on top of models—as well as multi-agent systems, enterprise AI architectures, and the practical ways organizations can begin adopting agentic AI while avoiding vendor lock-in. The episode is framed as a catch-up primer for anyone who feels left behind by the acceleration of the agentic AI revolution.