I have always found explanations of how modern AI models get built to either be a dense math lesson or oversimplified to the point of inaccuracy. I won’t make either of those mistakes, so here’s an attempt to split the difference and give a realistic overview without requiring knowledge in statistics.
Data is Key!
AI models rely on data, usually in the form of large datasets. For a language model, large datasets mean large text collections. In image models, it means a large collection of labelled photographs. With heavily simplified terminology, you can think of these models as being empty, and then taught by example. Like anything that is self taught, the data dictates the constraints on the model’s knowledge and ability. Blanks in data become the model’s blind spots. Because of this, a model that has largely been trained in one domain will be weak outside of it, just like a person who solely reads one genre of literature will struggle to write in other genres.
Most issues in AI, as we know them, become apparent at this stage. The biases in the data become apparent in the outputs when a model “learns” the biases. That’s why data collection is just as important as anything that comes after it, including building and training the model.
Training: pattern-finding at enormous scale
The model looks at data, adjusts itself, then looks at the data again. For a language model, this means predicting single words in a string of text and learning to do that for as many instances as data allows, developing a framework or scaffolding of grammatical structure and common linguistic patterns.
Statistical pattern recognition is not the same as understanding. The model does not analyze, or reason, like a person. The model makes textual predictions about the output that it statistically expects to come next given the data it has seen. The model does not make logical analysis to predict what is “factually correct,” and therefore, it is perfectly reasonable for people to misunderstand its outputs.
Fine-tuning: shaping general capability toward a specific use
Following the initial general training, models are often subjected to additional training called fine-tuning, where the model is shown a smaller corpus normed to the intended domain, and the model’s behavior is further constrained. This is generally where a language model’s ability to continue text in the most statistically likely way is replaced by the ability to follow instructions, converse, and produce safe outputs.
This is one of the main reasons two models built on top of similar foundational technology can appear quite different when you use them. Base training enables a model to learn general skills, but it is the refining that uniquely shapes their personalities and patterns experienced when interacting them.
Evaluation: testing whether it actually works
Models are evaluated against test sets and examples prior to launch. These evaluations demonstrate how well models perform Goal-Directed tasks like answering question and interpreter instructions, as well as avoiding harmful and biased outputs. This phase presents a legitimate challenge because quantifying how good a model is at abstract or broad tasks like reasoning will not necessarily simplify to a clearly discernible number the way it might for a task that is more defined. Different model evaluations can paint different pictures of the same model. This is one of the reasons why comparisons of models built by different entities often include more depth than what is communicated by marketing.
Why this matters if you’re just using these tools
Acquiring some understanding of how these models are constructed will help you use these models within the bounds of their capabilities. Understanding that these models are predicting text instead of constructing linguistic structures will provide you with explanation for the confidence that these models possess, with respect to outputs that are often very wrong. It will help you appreciate the bounds of these technologies and use them better.