There are three absolutely key differences here.

The first is that, if you get a four-year college degree in mining or civil engineering, you will not spend much of those four years studying management practices; you will spend it studying geology, the mechanical properties of rocks and soil, hydrology (how water flows underground), and existing designs that are known to work well. You probably will not build a mine or a highway, but you will design many of them, and your designs will be evaluated by people who have built mines and highways.

The second is related to why you will not build a mine or highway in those four years: those are inherently large projects that require a lot of capital, a lot of people, and at least months and often decades. Mining companies don't have to worry about getting outcompeted by someone digging a mine in their basement; even for-profit toll highway operators similarly don't have to worry about some midnight engineer beating them to market with a hobby highway he built on the weekends. Consequently, it never happens that the company has built two highways already by the time there is an official project plan, and I am reliably informed that it doesn't happen much with mines either.

The third is that the value produced by mining operations and highways are relatively predictable, as measured by revenue, even if profits are not guaranteed to exist at all. I don't want to overstate this; it's common for mineral commodity prices and traffic patterns to vary by factors of three or more by the time you are in production. By contrast, much software is a winner-take-all hits-driven business, like Hollywood movies. There's generally no way that adding an extra offramp to a highway or an extra excavator to a mine will increase revenue by two orders of magnitude, while that kind of thing is commonplace in software. That means that you win at building highways and mining largely by controlling costs, which is a matter of decreasing variance, while you win at software by "hitting the high notes", which is a matter of increasing variance.

So trying to run a software project like a coal mine or a highway construction project is a recipe for failure.

And as a side note, this is why LLMs are such a huge sugar rush for large companies. The performance of LLMs is directly correlated to capital investment (in building the model and having millions of GPUs to process requests).

Software rarely has a system that someone cannot under cut in their bedroom. LLMs is one such (where as computer vision was all about clever edge finding algorithms, LLMs are brute force (for the moment))

Imagine being able to turn to your investors and say “the laws of physics mean I can take your money and some open source need cannot absolutely cannot ruin us all next month”

That's an interesting thought, yeah. But it also limits the possible return on that capital, I think.