Operating AI

Foreword

The first time I met Ulrika, I knew she was a passionate and enthusiastic leader who cared immensely about operationalizing and productizing AI systems at scale. Although many books do a superb job of describing the wealth of AI technologies and parallelized algorithms, only a few describe the challenges of MLOps and the life cycle management of AI systems end to end. Widespread adoption of AI is not hindered today by any specific technology, but by the lack of understanding, trust, explainability, and experience in deploying these systems in the field. Many companies claim today to be using AI systems, but in reality, only a few of those systems are completely deployed and unleashing real business value.

I spent 30 years of my career in telecommunications on research & development of machine learning and AI systems, and I experienced firsthand multiple challenges that have hindered the massive deployments of AI. Let me state a few:

• Significant investment and multiple skill sets are required to build the underlying AI software and compute infrastructure, to create and test machine learning models, and to deploy and operationalize those models in the field. In a typical large company, this is an investment in the hundreds of employees across multiple organizations (operation engineers, architects, software developers, data scientists, researchers, product managers, marketers, to name a few), and millions of dollars in licensing, building, and operating the software ...