How do you find the jagged edge between AI performance and hallucinations?

About ten years ago, I was in Silicon Valley at an auto technology show.  These were days before OpenAI and LLM’s became household names.  A researcher was demonstrating AI on an edge device.  He used an Android smartphone and a simple classifer application to test cards showing a hand, a cloud, and a cat.  With each, the classifier quickly identified the subject of the test cards.

I knew from my own research back in the 80’s that researchers would boost results by using training data for testing.  I challenged the researcher, “Well, that’s great.  But what if we use a novel image, like my hand?”  I put my hand in view of the phone’s camera.  On the display, I saw a picture of my hand, and the caption, “Hand.”

AI has grown in leaps and bounds since that day.  A fundamental question is: “How do I successfully use AI?”  In this blog, we explore the jagged, cutting edge of AI, to find where it is useful, and where it is not.