Modern AI

I’m excited, and concerned, to see Artificial Intelligence become capable after so many years of failed promise. Back in the late 1970s I took a college¹ course surveyed the state of the art in AI. At that time there was very little practical application as the technology was weak and the hardware was limited. The course comprised only readings and we did not write or run any code.

In the 1990s I wrote a simple rule-based application intended to help software developers in a large organization² satisfy a set of process-management requirements. The rules were expressed in a declarative format much like a decision tree, and my program³ converted those rules to operational code in a simple survey dialog program. This was perhaps my favorite project in my career.

Not long after that I took a new job with a company that sold a development tool⁴ to generate applications using model-based reasoning with automated forward and backward-chaining over rules on an object-oriented framework. Sadly, that company pivoted almost immediately to an entirely different product space and they gave up any new development and most support of their model-based reasoning product.

Since then I’ve not worked on anything even AI-adjacent – to my regret. It was a shock then in 2017 to learn how AI technology had advanced to where AlphaGo⁵ was able to train itself to play Go at a world-master level in just a matter of days, simply by playing games against itself.

And now we have ChatGPT, Midjourney, and similar large language models proving capable of elaborate responses to simple prompts. The neural network architecture that was so long dimissed in the heydey of model-based reasoning has come back with a vengence, and the results are astounding.