The conception of language virtual machines, as we know it, has given us many things: high performance for even highly dynamic languages or intermediate representations (IRs); convenient and fast garbage collection; and reliable high-level tooling. All are enabled by the internal uniformity of a virtual machine. But virtual machines also take from us: through removing access to certain abstractions offered by the host system, supporting only one or a subset of the languages of interest, and providing only prescriptive tool interfaces, they create barriers and external needs which must be worked around in ways that are different for each VM (e.g. different FFIs) and often reinvented by each end programmer unfortunate enough to need them (e.g. the relative familiarity of custom tooling via bytecode manipulation).