Nicholas James

Unfortunately, I think that in the near future, not using LLMs to write for you will be like someone refusing to use Google Maps for directions in a new city. A bizarre idiosyncratic choice that's just completely incomprehensible to the vast majority of people.

I usually agree with Dr. Locasale. But not on this. A $100 full-body ultrasound on asymptomatic people, sold as a wellness service, will open a massive can of financial worms for individuals. You’ll pay around $100 for the scan, there’s a 90% chance they find something you’ll worry about, your insurance won’t cover the expensive follow-up they recommend, so you’ll be deciding whether to spend thousands to relieve anxiety about something with a 98% chance of being nothing. How’d I get there? Enough screening full body MRIs have been done to give us a good idea of the size of the can and what’ll be in it. Roughly 90% of asymptomatic people who get a screening full body MRI have an abnormality. We have decades of diagnostic imaging with millions and millions of followed-up findings that allow 2/3 of those to be classified as incidental findings that definitely don’t need further work up. Of the roughly 1/3 they can’t be sure about, after work up, 10% of them (4% of people who got a scan) end up with something important, and about 1.5% of people have a cancer detected. And the existing evidence says finding those lesions doesn’t improve overall mortality. I know that’s hard to believe, but it’s true. Look up Korean thyroid ultrasounds or German melanoma screening. Now extend it to a modality we have no baseline data on follow-up of lesions. Midjourney has been explicit that the early scans train the model and diagnostic accuracy comes later. For the first year or two, it’s reasonable to say 90% of people who get a scan will have an abnormality they can’t rule out as concerning, and it’ll be up to the individual to pay to train Midjourney’s model. Insurance probably won’t cover the follow-up scan or biopsy with an established modality, because there’s no evidence saying it needs follow-up. So 90% of people who get scanned will be worried and facing the decision of whether to pay out of pocket for the expensive follow-up. A small percentage of those will have side effects and complications from the follow-up testing. And remember, 98% of the findings are nothing. You can argue affluent people could get one, chase down all the abnormals, then repeat a scam (sorry, meant scan) every 3 months to look for changes. But that just leads to over-diagnosis with no mortality benefit. So, rich people get scanned every 3 months with subsequent overtesting and overdiagnosis, normal people get anxious over things that are 98% likely to be fine, nobody has a survival benefit and Midjourney gets their model trained for free. I wonder who’s coming out ahead here?

This book from 1993 was really way ahead of its time, explaining the concept of what we'd now call a "digital twin," a computer model of a real-world system that is kept in constant sync with the underlying real data. This approach is more relevant than ever now that agents can study documents, plans, and data and use them to automatically design the models. And then other agents can automatically keep the models in sync with reality by querying APIs, parsing documents, checking cameras, monitoring transactions, etc. This lets you understand what's happening in real time, make predictions (and ground them in logic and causality), detect anomalies and explain them, perform hypothetical what-if simulations, and generally apply the scientific method in a rigorous way to test your understanding of the underlying reality. While digital twins used to be so much work to build and update that they only made sense for big, high-stakes projects like NASA space missions, now you can model all sorts of everyday systems in this way, like a business or school. We know that the scientific method is fabulously, "unreasonably" effective in explaining the rich complexity of the natural world. It makes you wonder what becomes possible when we start applying that same lens and approach, guided by super-smart AI agent practitioners, to realms that until now have mostly been governed by simple heuristics and institutional inertia.