Smarter and Dumber

While I was running errands the other day, I caught part of a This American Life piece on NPR about the development of GPT-4 at Microsoft. The episode included interviews with several engineers and developers who were involved in the project, and their thoughts (and surprise) at just how advanced GPT-4 became.

Some of these folks were AI skeptics, to some degree. They were not sure if something like ChatGPT was really learning and expressing some sort of intelligence, or if it was just BSing its way through questions and tasks.

As they continued pushing and testing the system, they were shocked and sometimes disturbed by the responses they got to questions they posed to the system—things that were not likely to have been something in the training data. The system was making connections and extrapolating correct answers in ways it was not originally designed to do. It could create rough images, even though it was not an image generator.

At one point, Peter Lee, a computer scientist and head of Microsoft Research, said something pretty profound. Instead of being in awe of how an AI system can mimic human intelligence, maybe we are overestimating the complexity of our own brains. As he put it, “It does make me wonder how much of our intelligence is truly complicated and special.”

Of course, we are probably also overestimating AI, since it has been shown over and over again that AI systems can generate garbage answers. Again, as Lee said in the interview about GPT-4, “The thing is both smarter and dumber than any person you have ever met.”

I was thinking about all of this because we are seeing more and more development of AI-based tools for simulation, for CAD, for creating manufacturing instructions in CAM—all over the place. At the recent NAFEMS Americas conference in Louisville, Kentucky, the AI-focused conference sessions were filled to capacity, and most of the simulation software vendors present had an AI story to tell.

On Oct. 31, our fourth annual Digital Engineering Design & Simulation Summit will lead off with a keynote panel discussing AI in design—the possibilities, the challenges, and what AI can and cannot do (or be trusted to do) in an engineering workflow.

Our panel of experts will discuss how AI, machine learning and reduced order models can optimize simulation workflows. They will take a look at how well AI can model against uncertainty, what data might be needed to train the models, and how you can prevent AI data hallucination or model collapse.

Panelists include Alexander Lavin, founder & CEO of Pasteur Labs and the Institute for Simulation Intelligence (ISI), as well as AI expert Astrid Walle, principal of Astrid Walle CFDSolutions, and Sandeepak Natu, co-director of the simulation-driven systems development practice at CIMdata.

The summit will also feature sessions on real-world digital twin use cases in design, engineering workstation configuration tips, optimizing design for additive manufacturing, and how to determine the business benefits of simulation-driven design.

You can learn more about the show and register on our website.