AI Models

What Is an AI Model?

At its core, an AI model is a mathematical and computational system trained on data to recognize patterns and generate outputs. Think of it as a sophisticated engine that takes in information—text, images, sensor data—and processes it to produce predictions, answers, or creative content. While the term is often used interchangeably with specific tools like ChatGPT, an AI model is the underlying engine that powers those experiences.

Foundation Models: The Versatile Engines

Not all models are built for a single task. A foundation model is a class of AI trained on vast, diverse datasets using self-supervised learning at scale. The term was coined in August 2021 by the Stanford Center for Research on Foundation Models (CRFM) to describe systems that are not limited to one narrow job. Instead, they are adaptable to a wide range of downstream tasks. Their scope is broad, encompassing language, image generation (like DALL-E or Stable Diffusion), audio (such as MusicGen), and even robotic control (like RT-2).

Large Language Models (LLMs)

A Large Language Model is the most prominent subset of foundation models—specifically trained on text. LLMs excel at natural language processing by predicting the next token (a word or part of a word) in a sequence. This capability enables question answering, summarization, translation, and code generation. Familiar examples include GPT-4, Claude, Gemini, and Llama.

Parameters: The Model’s Internal Knobs

To understand how models work, you need to understand parameters—the learned weights inside a model that determine how inputs are converted into outputs. During training, the model adjusts these weights to improve its accuracy. The scale has grown exponentially: GPT-1 started with 117 million parameters, GPT-2 had 1.5 billion, and modern frontier models reportedly contain over a trillion. You can think of parameters as thousands of tiny knobs the model tunes to refine its understanding of the world.

The Chef Analogy

Imagine a chef learning to cook. Training is the long, intensive process of studying thousands of recipes, learning which ingredients pair well, and practicing techniques until they become second nature. The parameters are the chef’s internalized knowledge—the refined intuition of how much salt to add or how long to sear a steak.

Inference is the moment the chef is in the kitchen, taking a specific order and using that trained knowledge to prepare a meal. The chef is not learning to cook anymore—they are applying what they already know to produce a result. When you type a prompt into ChatGPT, you are triggering inference: the model is using its trained parameters to generate a response to your specific input.

Training vs. Inference

Training is the process of adjusting a model’s parameters using optimization techniques over massive datasets. It is computationally expensive, requiring clusters of GPUs running for weeks or months inside data centers.

Inference is the process of running a trained model to produce outputs from new inputs. Because inference is what happens every time a user interacts with an AI system, engineers use techniques like compression, distillation, and distributed processing to make it faster and more efficient at scale.

Adaptation Paths: Making Models Useful

Once a model is trained, it can be adapted to specific needs through several paths:

• Prompting — the cheapest method. You provide instructions to the model without changing its internal weights. The model responds based on its existing training.
• Fine-tuning — retrains the model’s weights on a smaller, specialized dataset to make it an expert in a specific domain (legal, medical, financial).
• RAG (Retrieval-Augmented Generation) — connects the model to external or real-time data at runtime, allowing it to answer questions about information it was never trained on without retraining.

Why It Matters for the Agentic Economy

Understanding AI models is essential for navigating the agentic economy. As we move from simple chatbots to autonomous AI agents—systems capable of taking action on your behalf—the underlying model acts as the decision-making core. The model’s architecture, training data, and adaptation method determine its reliability, capability, and risk profile, including its susceptibility to hallucination. Knowing how models are built, adapted, and deployed helps us better anticipate their potential and manage their limitations.

Sources

• [1] Stanford Center for Research on Foundation Models (CRFM), “On the Opportunities and Risks of Foundation Models,” https://crfm.stanford.edu/2021/08/03/foundation-models.html (August 2021)
• [2] Bommasani, R. et al., “On the Opportunities and Risks of Foundation Models,” https://arxiv.org/abs/2108.07258 (August 2021)
• [3] U.S. Executive Order 14110, “Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence,” https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/ (October 2023)