Pour yourself a drink, as I have a longish story that might be a useful metaphor.

Back in the day, there were more or less two consumer flight sims: MS Flight Simulator and XPlane. MSFS was and has always been the much prettier one, much easier to work with; xplane is kludgy, very old-school *NIX, and chonky in terms of resource usage. I was doing some work integrating flight systems data (FDAU/FDR outputs) into a cheaper flight re-creation tool, since the aircraft OEM's tool cost more than my annual salary. Hmm, actually, ten years of my salary.

So why use xplane at all, then?

The difference was that MSFS flight dynamics was driven from a model using table-based lookup that reproduced performance characteristics for a given airframe, whereas xplane (as you might be able to tell from the company name, Laminar Research) does fluid and gas simulation over the actual skin of the airframe, and then does the physics for the forces and masses and such.

I caught some flack for going with xplane: "Why not MSFS!? It's so much prettier!"

Unless the airframe is in a state that is near-equivalent with tabular lookup model, the actual flight is not going to be actually re-created. A plane in distress is very often in a boundary state- at best. OR you might be flying a plane that doesn't really have a model, like, say, a brand new planform (like the company was trying to develop). Without the aerodynamic fundamentals, the further away you get from the model represented by the tabular lookups, the greater the risk gets.

And how does this relate?

Those fundamentals- aerodynamic or mathematical or electrical- will be able to deal with a much broader range than models trained on existing data, regardless of whether or not they are LLMs or tabular lookups. If we rely on LLMs for aerodynamics, for chemistry, for electrical engineering, we are setting ourselves up for something like the 2008 Econopalypse except now it affects ALL the physical sciences; a Black Swan event that breaks reality.

I am genuinely worried we're working outselves into just such an event, where the fundamentals are all but forgotten, and a new phenomenon simply breaks the nuts and bolts of the applied sciences.

As for my xplane selection, it helped in other ways. Because often the FDR data is just plain wrong, but with xplane you could actually tell, because a control surface sticking out one way, while the flight instruments say another, lights up a "YOU GOT PROBLEMS" light in the cockpit as the aircraft inexplicably lurches to the right.

Ah, but it's no longer "inexplicably!"