Think half an abstraction layer higher. You're on the right track with multiple PHP virtual runtimes on a single VM - that could conceptually be viewed as a sort of precursor to function runtimes.

The serverless function has higher-order features included as part of the package: you get an automatic runtime (just as with PHP but in this case it can be golang or dotnet), the function gets a unique endpoint URL, it can be triggered by events in other cloud services, you get execution logging (and basic alerting), multiple functions can be chained together (either with events or as a state machine), the function's compute can be automatically scaled up depending on the traffic, etc.

Think of it as: What do I have to do, in order to scale up the conpute of this URL? For hardware it's a call to DELL to order parts, for VMs or containers it's a matter of scaling up that runtime, or adding more instances - neither of those processes are simple to automate. One key characteristic of the function is that it will scale horizontally basically however much you want (not fully true, aws has a limit of 1500 instances/second iirc, but that's pretty massive), and it will do it automatically and without the request sources ever noticing.

Functions are also dirt cheap for low/burst traffic, and deployment is almost as easy as in the PHP FTP example. Personally I also think they are easier to test than traditional apps, due to their stateless nature and limited logical size (one endpoint). The main downsides are cost for sustained load, and latency for cold starts.

With that said, they are not "endgame". Just a tool - a great one for the right job.