The adoption of atom-typing allowed for the creation of incredibly widely used force-fields (FF) such as
GROMOS/AMBER/OPLS and many more. Because of the assumptions required in Molecular Mechanical (MM) models, in order to be accurate, at least part of the force-field must be fit to experimental data. This is very tedious, and there are many years, decades even, between updates to a FF.
Recent FF development has utilized chemical perception rather than atom-typing and seems to have alot of potential. For instance, the openFF consortium seems to be making large strides in creating an open source, easy to use and tinker with, FF.
Also, Parallel and GPU computing have come along way in the last decade. I would hope/assume that GPUs should make parameter optimization considerably faster than in previous decades.
Are there FF groups/projects in development that have the aim of being open source and highly "tinkerable", and are we in sight of being able to generate custom FFs somewhat on the fly, as we need them rather than relying on databases of atom-types fit nearly exclusively to solutes in water?