This is something that's doubly frustrating for dedicated autodidacts. For some areas, once you start diving into them (and giving up all of your free time in the process to make any progress), you realize just how much more time is necessary to get to the "baseline" / "start of the art / be able to make a useful contribution (and how useful it is to be in an immersive learning environment surrounded by people you can collaborate with.)

You can easily hit a wall even with all the books and MOOCs at your disposal, due to having to do all your learning after work when you're tired or on weekends.

To have to go into debt in order to work on interesting things is unfortunate; just how many people who could have made interesting contributions in a variety of areas never do so because of that?

Studying yourself into web dev and such is eminently possible, but if you want to do, for example, deep learning work at even one-half the Richard Socher-level, you almost surely need to go for an advanced degree. It might be different if you're a superhuman autodidact rather than just a "standard" one, but there are a lot more of the latter (a set that I include myself in) than the former.

You're right that more technical areas (ML, algorithms, maybe even systems) expect more advanced degrees.

But there is far more demand for standard development/architect roles and even those roles command a generous salary.