Is it still worth learning to build a model from scratch?

Training a model from scratch seems like a way to master AI. In practice, there are cheaper ways to gain the same understanding and get the result you want.

Have a reason for the training project. A small model for a specific task is very different from trying to create yet another general-purpose model.

For most people, the more useful path is learning to adapt existing models and get the behavior you need from them.

And be skeptical of claims about models that “improve on their own”: not every claim like that proves it already happens the way people imagine.

Understanding is different from building

Building a small model in a course with little text and limited hardware teaches you a lot: what tokens (pieces of text) and parameters (the numbers the model learns) are, and why the model makes mistakes.

You can do that on an ordinary computer for little cost. Training a useful general-purpose model takes a lot of data, hardware, and time, and puts you in competition with companies that do only that.

When training makes sense

A custom model may be worthwhile for a narrow task, such as classifying documents in an industry, running on a small device, or working with data that can’t leave the company.

Even in those cases, people almost always start from an existing small model, not from scratch.

The steps for adapting a model

Start with the simplest and least expensive option. First, write good instructions and include examples in the request. Next, give the model your documents at response time, a technique called RAG (searching your own knowledge base).

Last comes fine-tuning, which trains an existing model a little further with your examples. Many problems can be solved in the first two steps.

Improve on your own, with evidence

When someone says a model improves on its own, ask how they measured it, on which task, and compared with what. Without a before-and-after test, it’s just a claim. The same applies to your project: define the metric before training.

1. Write one sentence describing the task the model needs to do and how you’ll measure accuracy.
2. Set aside a set of real examples with the correct answers for testing.
3. Test an existing model with only instructions and examples in the request.
4. If it lacks company knowledge, add a search of your documents.
5. Consider fine-tuning only if the previous steps don’t meet the goal.
6. To learn how it works, do a small study project without expecting production use.

• Starting with training: it costs time and money before you know whether an existing model will work.
• Training without a test set: there’s no way to know whether it got better.
• Believing in automatic improvement without numbers: impressions of improvement can be misleading.

• Train from scratch only for a clear reason.
• Adapting an existing model solves more problems.
• “Improves on its own” needs evidence.

• CourseKit Local — the local labBefore training, understand what ready-made models can already do on your machine.
• ProjectQual modelo de IA usarBefore training, see which ready-made model can already handle your task.

Some of these materials are in Portuguese.