Some Thoughts on AI in Manufacturing

Intro

AI is an umbrella term for many different technologies, and has been a research topic for decades. The defining feature of modern AI is its flexibility. Before the ChatGPT era, AI often meant spending millions of dollars to train a very specific system that was good at one task, like detecting scratches. Researchers discovered that with enough compute and data, a single model can form the foundation for many different tasks with little or no task-specific training. That is why we call these "foundation models".

Foundation model technical capabilities have been accelerating. An application that either wasn't possible or wasn't economical 6 months ago may be both today.

We're going to highlight some capabilities for the manufacturing world that are valuable, interesting, and have just become viable in the past year or two. We're going to skip the stuff that floods LinkedIn like workflow automation around text documents. However that stuff is valuable and we're happy to chat about it too.

Hoping you find some nuggets here.

Flexible computer vision

The 2010s saw the rise of computer vision (CV) technology based on deep learning. This approach was more flexible than "traditional" CV, but it still required lots of data and training for each use case, and it couldn't apply human-like judgement to new scenarios. These approaches worked well for tasks in controlled environments like looking for scratches on mass-produced cans.