From the editor: One of our most credentialed AI experts shared some essential perspectives for companies betting on AI. This speaker has a PhD in Brain and Cognitive Sciences, and has been working on machine learning models for over 15 years. As always, ON_Discourse operates under the Chatham House Rule—no attribution of perspectives without explicit consent.

You need an
enormous amount

This is one of the AI era’s most
under-appreciated assumptions.

In the AI era, everyone seems to have convinced themselves that their company has the best data. You likely believe that not only is your data just the best, but it also happens to be just so unique that the resulting AI model is going to supercharge your whole business.

This is wrong, for two reasons:

1. You have too little data. You need billions or even trillions of heterogeneous observations to be able to do anything interesting with AI. There are only a handful of companies today that possess that quantity.
2. Your data is of poor quality. Even most large enterprises do not have a dataset rich enough to make a meaningful AI play. They rely on buying data from multiple sources or stealing it from internet users under the guise of fair use.

Many companies are opting for a shortcut into AI. Any company today can implement base GPT models. That works for experimentation, but it won’t drive revenue or improve efficiency. Why? AI is not trustworthy. You need to hire experienced teams to manage its bad behaviors.

For almost every single company, trying to build something bespoke that will move the needle is simply not worth the investment.