The Probabilistic Product

Chapter 2. How AI Actually Works (Without the Hype)

A Note for Early Release Readers

With Early Release ebooks, you get books in their earliest form—the author’s raw and unedited content as they write—so you can take advantage of these technologies long before the official release of these titles.

This will be the 2nd chapter of the final book.

If you’d like to be actively involved in reviewing and commenting on this draft, please reach out to the editor at [e-mail].

Early in my career, my team had a model in production. It powered a care management tool built to help prioritize which patients needed intervention most—which care gaps to close and in what order, calibrated per person. It worked, for a while. Then, quietly, the data science team’s metrics started sliding. Not crashing—just degrading, slowly enough that you could convince yourself it was noise until it wasn’t. This was machine learning (ML) in its earlier days of production deployment. The frameworks for keeping models honest at scale were still being figured out industry-wide. So we figured it out ourselves: retrain every three to six months, build out better performance tracking, measure model drift—the decay of predictive power due to changes in real-world data— explicitly. The model stayed live. We just learned that shipping it was never the finish line.

Traditional software is deterministic; you define the rules and get the output. AI is something else entirely. Machine learning shifted engineering ...