“I Trained My AI Model”: A Reality Check

What “Training a Model” Really Means

Training a generative AI model from scratch, often called pre-training, means teaching a neural network from a very large dataset using deep learning methods.

During training, the model processes a very large number of examples. For each one, it predicts an output, compares it with the expected result and adjusts its internal values to reduce the error. For a language model, the examples are mostly text, and the task is to predict the next part of a sequence.

These internal values are the model's “weights” or “parameters”, and recent large models have billions of them. The adjustment is repeated over a very large number of steps, until the model produces coherent results.

This requires a large amount of computing power, typically provided by specialized processors such as GPUs, which are suited to intensive calculations on large volumes of data. The scale of well-known models gives an idea of the effort involved:

Pre-training a large general-purpose model requires substantial data, computing infrastructure and specialized expertise. In practice, the most widely used foundation models are developed by a small number of large technology companies and well-funded research organizations.

Pre-Training, Fine-Tuning, RAG and Workflows

Adapting AI to a specific business does not necessarily mean training a model. Four approaches are often confused, and only the first two change the model itself:

Fine-tuning is a legitimate approach for a specific problem, such as a task a general model does not perform well even with good instructions. Many marketing tools described as “AI-powered” combine an existing model with instructions, reference material and workflows. These approaches are not mutually exclusive: a product can use a fine-tuned model, retrieve reference material and run inside a workflow.

What the Claim Can Mean in Practice

When a tool or a service is described as “trained”, it may involve fine-tuning, one or more of the following approaches, or a combination. Each of these approaches can be useful; none of them creates a new model:

1. Detailed instructions: The work consists of writing detailed prompts. This is a valuable skill: it instructs an existing model rather than building a new one, like directing a film rather than building the studio.
2. A no-code or low-code workflow: Tools such as N8N or Make chain together calls to existing models. This can work well for well-defined tasks. Without validation and handling of unexpected inputs, such workflows can be fragile when cases differ from those they were designed for.
3. A model combined with templates: A generative model fills in a library of pre-made templates for ad copy, blog posts or emails. This is useful in some situations, within the limits of the templates.
4. Reference material retrieved at run time: The model receives relevant documents from a defined source when it answers (retrieval-augmented generation). The model itself is not retrained.

The Bottom Line: Ask Which Approach Is Used

New general-purpose foundation models usually come from large technology companies or well-funded, research-driven organizations. For digital marketing teams, useful progress is likely to come from two directions:

• Improvements in the foundation models released by these organizations.
• Specialized solutions built on top of these models, combining domain expertise, reliable data, evaluation and integration into existing processes.

When a tool or a service is described as “trained”, ask which approach it uses: pre-training, fine-tuning, retrieval of reference material, a workflow or a combination. Then ask how it was evaluated on tasks similar to yours, and what it takes to run and maintain. The answers say more about its value than the word “trained”.