‘The bigger the model, the better the result.’

Every new model arrives with its benchmark scores, and it is readily assumed that the most powerful will be best for everything. Projects then choose the model at the top of the rankings, at the highest price.

For many business uses, that is not the case. Classifying requests, extracting information from a document, answering from a knowledge base: a smaller model, well chosen and connected to your data, often does just as well, faster and for much less. It can also be hosted more easily in Europe or on your premises. The right model is chosen by testing it on your real cases.

• Rankings measure general performance.Your specific use does not feature in them, and the gaps there are often different.
• A bigger model costs more for every request.At high volume, the difference becomes a significant budget item.
• Quality depends mostly on what you give the model.Clear instructions and reliable sources count for more than size.

1. 01Build a set of around fifty real cases, with the expected answer.
2. 02Compare several models of different sizes on that set, measuring quality, speed and cost.
3. 03Choose the smallest one that gives a satisfactory result, and repeat the test regularly.

• Should we train our own model?→
• Should we host our own models or use a major provider?→
• How do you keep up with a technology that changes every month?→
• How do we avoid becoming dependent on an AI supplier?→