Back to the future: training models
The funny thing about RAG is that the basic concept has been around for as long as machine learning has. Long time readers will recall that back in the day, I studied Data Science in undergrad. “Old School” machine learning, before everyone was calling it AI, was entirely predicated on training a new model for every problem.
How old school ML worked: custom models
Imagine you’re a Data Scientist tasked with understanding and predicting customer churn (your employer has a big churn problem). Your goal is to be able to predict a brand new customer’s chances of churning, so your marketing team can give them discounts and winback offers before they leave. Here’s what you might do:
- You gather a curated data set
You spend time gathering all of the historical data your company has about churn: who churned, when, what characteristics they had when they did, and anything they did beforehand. Each customer gets a label: churned, or didn’t churn.
- You train a model on the data set
Using either simple linear regression or something as complicated as deep learning with neural networks, your model goes through the data and tries to find patterns. It eventually learns (or tries to learn) which characteristics tend to lead to a customer churning, and which don’t.
- You test the model on new data
To make sure the model isn’t just spitting your data back at you, you test it on new data and see how it performs. The model needs to generalize, meaning perform well on new data that doesn’t look exactly like the data you trained it on. Models that are trained too well on the training data and don’t generalize well are called “overfit.”
The important theme here is that each model – whether you trained it from scratch, or took an existing model off the shelf – needed to be customized to your data set. This was how everyone thought about machine learning when I was doing it professionally. Everyone has different problems, so everyone needs different models.
Generative AI: not customized by default
The Generative AI that we use today, like ChatGPT or Claude, isn’t like this at all. Instead, it’s trained on one, colossally large data set – all of the internet – that isn’t curated and doesn’t belong to your business at all. You prompt the model to focus it on a specific problem, say, generating an outreach email, and it outputs something that represents the data it was trained on.
The broad strokes, non-customized nature of these GenAI models is fine for some use cases. But for many, especially business use cases, you need models to be aware of your data and give you those kinds of tailored responses. You can’t build a customer support chatbot that doesn’t know anything about your product .
So how do you customize GenAI models to work with your data?
