Having used both Erlang and Elixir "in anger" there are a whole bunch of things that bug me about Erlang. There are a whole bunch of things that bug me about Elixir too. I'm not sure what any of that has to do with what I said. Elixir is a more complicated and larger language than Erlang.
It has structs (which are weirdly implemented as maps instead of records). It has type-classes/traits by way of "protocols" (which are a whole different layer of metaprogramming that could be replaced with behaviours and data-structure definition convention or dialyzer type-spec'ing). It has the pipe operator (which strangely elides the first argument in a function such that map/2 becomes map/1 when written... why no placeholder sigil). It has Agents (which are something which is confusing to have in the standard library as its a specific implementation of a gen_server which acts as a KV-store which will be really, really prone to data races). It has a fair bit of metaprogramming to be aware of in terms of macros and hooks that trigger at code loading/import time (which forces one to need to be a lot more aware of the inner details of 3rd party libraries being used). It eventually devolves into requiring knowledge of Erlang when attempting to do anything regarding debugging, tracing, or distribution related.
Elixir is a great language, and it has a lot of neat features, and I love teaching it to people and being increasingly critical of Erlang based on advancements in usability and community engagement/feedback that I see happening around Elixir. But I don't really need to be proselytized to about it.
There are just so many things that make dealing with OTP nicer in Elixir. This is to say nothing of the meta-patterns that Elixir has driven home: Agents, Tasks, the actual supervisor/worker interfaces themselves. You could go on here.
I haven't even gotten to macros -- Have you ever had to make your own behavior, say for an acceptor/worker pool of some variety (presumably non-tcp, otherwise just ranch that out). Making your own behavior (what armstrong himself called "really advanced") is the easy part here. You've got to create a system where you pass around the TRUE module that holds the correct `handle_call/cast/info` implementation back and forth through the true module and the OTP state parameter.
This whole pattern (which comes with very real runtime costs) just doesn't even exist in Elixir. You just create a macro. A macro that sets all of this up at build time. No complex supervisor hierarchies where you need to constantly inspect each and every message, no custom behaviors, etc. You just have a mostly sane way of sharing software logic to begin with.
You can work up really clever solutions with parse_transforms, even standard "-define", but today Elixir has a totally brilliant solution to this problem and many more classes of problems.
This isn't to say there an't blindspots (beyond "wrap proc_lib/inets/other hated library"). There's minor issues here and there. Process registry naming can get confusing, especially operating between Erlang+Elixir supervision trees. Low-level details aren't nearly as well-publicized in Elixir either. There are even big issues, like the Elixir community leaning more "I'm playing around with Elixir because Phoenix is Webscale" than the grizzled realtime adtech/gambling/finance gurus and so if you want help with OTP platform/VM details, you going to have a harder time.
I first tried Elixir in v0.08. & function capture syntax had just been standardized. I suspect that like me, you tried it then, when it wasn't a fully-featured alternative yet. Things have changed and Jose Valim has a vision for the future of Erlang that I think you'd be foolish not to pay attention to.