What Does "AI Model" Mean? A Non-Technical Glossary Entry for Executives

In AI, a model is a mathematical system that has been trained to perform a specific task. The training process involves exposing the model to large amounts of data and adjusting its internal parameters (billions of numbers) until it performs the task well on the training data and generalises to new examples it has not seen before.

A language model is trained to predict and generate language. An image classification model is trained to identify what is in images. A recommendation model is trained to predict what content or products a user might engage with. In each case, the 'model' is the trained mathematical system that performs the task, encapsulated in a set of parameters that can be saved, loaded, and run on computer hardware.

Two AI models trained for the same task but trained on different data, with different architectures, or with different training objectives will behave differently. This is why GPT-4, Claude, and Gemini give different responses to the same question, even though all three are large language models designed for similar purposes.

The differences reflect choices made during training: what data was included, what values were embedded through reinforcement learning from human feedback, what safety constraints were applied, and what tasks the training emphasised. These choices produce models with different capabilities, different tendencies, and different failure modes.

Understanding that different models have different characteristics helps you make better AI procurement and deployment decisions.

Different models are better at different tasks. Claude has been noted for strong performance on long document analysis and following nuanced instructions. GPT-4 has been noted for coding ability and broad general knowledge. Gemini has been noted for multimodal tasks. These are generalisations that change as models are updated, but the principle holds: choose your model based on what your use case requires, not just on vendor familiarity.

Different models have different safety and reliability characteristics. Anthropic designs Claude with safety as a primary design objective. OpenAI applies specific safety training to GPT-4. Google applies its own responsible AI standards to Gemini. These differences have practical implications for regulated use cases where the AI's reliability and refusal to produce harmful outputs is important.

Model versions matter: GPT-4 and GPT-3.5 are different models with significantly different capabilities. Claude 3 Opus and Claude 3 Haiku are different models with different performance and cost characteristics. When evaluating AI performance or cost, specifying the exact model version is important for comparable assessment.