I have been immersed in the rapidly evolving world of artificial intelligence and machine learning (AI/ML) for almost two decades with a focus on technologies that can transform markets by solving hard(er) problems. Several of these instances involved semiconductor-based hardware systems, while others were pure software to make at-scale compute more efficient and more performant. 



Over the past decade or more, I’ve had the opportunity to assist with many software companies in the AI/ML space. Whether using AI/ML to optimize genetics for biological factories, providing a platform for AI/ML workloads across CPU, FPU, and FPGA hardware accelerators, or tripling optimized performance for AI/ML inference, the technologies applied within AI/ML neural networks and processing systems have not only made great strides in data set size and processing, but also some significant advances in formulation of problems. 

The first neural network I personally constructed was very, very limited and was based on a switched capacitor structure back in my university days and only mimicked a couple of synapses. The advances since have been enormous, and it is not hyperbole to say exponential with hundreds of thousands of times the compute available and models that are millions, even billions of times larger today.  

However, even as these advances have occurred and we are seeing broader applicability and adoption of AI/ML, we still face several monumental challenges that have not been addressed because ultimately AI/ML systems rely on content, data, and appropriate training and governance of operation.