No, they're not "all ML". ML is the whole package, not one part of the algorithm.

Obviously if you apply any tech badly it isn't magic. ML does what it's intended to, which is find the best model to approximate a specific phenomena. But when it's applied correctly to an appropriately scoped problem, it does a good job.

LLMs do not do a good job at anything but telling you what language looks like, and all the investment is people trying to apply them to things they fundamentally cannot do. They are not capable of anything that resembles reasoning in any way, and that's how the scam companies are pretending to use them.