Should we train our own model?

Training a model of your own has something reassuring about it: it would be adapted to your business, fed with your data, and belong to you alone. Some suppliers encourage the idea, and projects get started on that basis.

In the vast majority of cases, it is not necessary. Existing models can already read, write, summarise and reason. What they lack is your knowledge: your procedures, your documents, your rules. You give it to them by connecting them to your own sources, without retraining them. It is faster, cheaper, and keeps up with changes to your documents without starting over.

• A trained model ages quickly.General models improve every few months, and yours has to be redone to keep up.
• Training requires a lot of well-prepared data.Far more than most companies have for a given use.
• Connecting a model to your sources covers most needs.The model relies on your up-to-date documents, and you know where its answer comes from.

1. 01Before considering training, test the use with an existing model connected to your documents.
2. 02Note precisely what is missing from the answers you get, and what that costs.
3. 03Only commit to training if that gap is measured, significant and lasting.

• ‘The bigger the model, the better the result.’→
• Should we host our own models or use a major provider?→
• Is our data ready for AI?→
• How do you keep up with a technology that changes every month?→