In this wide-ranging conversation, we look at various challenges posed by the different types of construction AI applications coming to market now for construction business and financial software, construction equipment automation, machine control and more.
IRONPROS: Well, you know, if you look at an AI application built into a safety system that would be installed on an excavator, that might use machine vision to detect a person on an intercept course with that machine and slow and stop the machine before it hits them. It also may be able to stop the tour it before it swings the bucket over and decapitates someone. These are things where there is not a human in the loop. And there are some cases where I know some types of machine vision have been found to have some inherent biases and the people creating the safety systems need to have had to make sure Well, gee, we're going into a new geography. And people look different or they dress different or there's smaller, we need to get a good sample and make sure they're represented. Where do you see like that bias problem? You know, factoring in with large language models?
SEATON: Yeah, it's a great question. I think bias is less a problem, then. Well, bias leads to a problem. That is true in all AI. I think the way it shows up in large language models is less about bias, and more about incompleteness, but it but in both cases, it's we don't know why it's making the decisions it's making until we get to look in the model and understand why right, like, why is it not recognizing this person? Or, you know, there's a story that I like to tell from like 2017, this team that made a machine vision model that could recognize dog breeds, and it kept mistaking Huskies for wolves. And the reason it did ...
STREAM OUR FULL CONVERSATION FOR MORE
