Over the past 5 years, I’ve had a bit more than a passing interest in the intersection of machine learning models and how they can be applied to tasks in the physical world. In university, I initially focused on projects centered around how ML can be beneficial to design-type classification (text first, then images). 

In retrospect, it was incredibly fun. But, I think these efforts tried to solve problems that were 1. based in 100 percent ground truth with no risk tolerance and 2. don’t leverage the the capabilities that AI solutions can offer in areas where there is ambiguity. 

The problem I tried to solve with the image classification model had one clear solution, so whose to say that couldn’t have been done by creating a statistical model? This also points us to a question - where does using AI actually justify all the associated added costs?

Like many other people online, I spent a good chunk of the last year trying to answer that question by creating too many GPT 3.5 wrapper scripts, and seeing what works well (and what doesn’t). 

I don’t have an answer yet, but I can say I’ve learned a lot about issues humans have with processing text and issues machines have processing text. And at the intersection of those two issue sets is a really cool opportunity space that I hope to build products in.