Could Path Dependence Thwart AI Advancement?

I have no idea why I woke up thinking about this this morning, but a few hours later I still find the question interesting.

The old story, which you've probably heard, bullet-pointed:

• Roman war chariot axles were built wide enough to be stable when pulled by horses -- about 4'8".
• Roman roads were built wide enough to accommodate Roman war chariots and developed ruts approximately that far apart.
• Everyone else building a wagon or cart used the same axle width so that the wheels would track along the ruts.
• English mine tram rails just adoped the same wagon width, because that's the way things were.
• The Stockton and Darlington Railway (the first steam locomotive railway) stuck with that gauge because that's the gauge craftsmen were used to working with, and it became the standard rail gauge everywhere.
• Railroad tunnels were slightly wider than that gauge to accommodate both tracks and cars.
• The space shuttle's solid rocket boosters had to fit on those railway cars and get through those tunnels, so in a very real way the size of space rockets was dictated by the width of a Roman chariot-pulling horse's ass.

That's the go-to example of "path dependence" -- past choices tend to constrain future actions. It's a form of inertia: Here's the direction I've been going in, so here's the direction I'll keep going in.

The big deal with artificial intelligence models is getting to the point where one model designs the next, better, model.

But what if there are conventions built in from the get-go that each future model continues adhering to when designing its successor model because reasons, and what if those conventions mean that some leap, jump, or left turn doesn't get made even though it would produce a breakthrough while the constrained path eventually hits a dead end?