Does "Good AI" Exist?

I’m working on a piece that roughly is “The Fermi Paradox, but for AI.” Has anyone seen actual examples of generative AI positively transforming an organization or process? What has it made a lot better? I want the best exemplars of this thing—if they exist—with data to back up the claims of transformation. Anything you’ve heard of would be appreciated.

I do need to do more research, but there are two areas that come to mind. The first one is using it on medical data to find anomalies, but I’m sure that’s something that you’ll do deeper research on. I’ve only heard rumblings of it: How Artificial Intelligence is Disrupting Medicine and What it Means for Physicians | Harvard Medical School Professional, Corporate, and Continuing Education

The second is related to VFX. I’m not here to say generating concept art or assets is good, however, from following Corridor Crew over the last few years, they’ve experimented with AI in their workflows for a while now. They don’t seem to be as public about it anymore (likely due to backlash from their anime projects), but the stuff they’ve shown with AI/ML-powered rotoscoping tools has been quite compelling and seems healthy as someone who is a fan of movies but does not work on them. I can’t find clips of specific examples, but it’s something.

1 Like

For the healthcare component, I’m close enough to tell you that the actual benefits seen are almost entirely in the predictive space. Example: a model that predicts whether an admitted patient is a high risk for falls, and automatically tags that patient for additional observation and support. This has had real benefits to safety in hospitals. Another one looks for signs of sepsis before it’s officially diagnosed. These are huge wins. I am a believer in well-applied machine learning.

The generative side…I dunno. The proposals I’ve seen pretty much come down to summarization. Doctors will tell you, and I believe them, that it would be useful to have an accurate summary of a patient’s chart in the five minutes they get before seeing them for the first time about a new complaint. Patient histories can be the equivalent of hundreds of pages. I understand wanting a distillation.

But I have also read of what happens when the models hallucinate issues and the doctors do not confirm on the chart itself.

I’m willing to believe there are really domain-specific applications like VFX. That makes sense to me, although my guess is that the good stuff is a lot less generative (i.e. transformer/diffusion architecture) and a lot more ML + procedural to make quick work of hard jobs like, I dunno, hair.