It’s not early days for it, I feel it’s been done and done and failed every time so far. I think the fundamental problem is most useful work, which we can define as work that can be used in another context rather than just validating difficulty — eg protein folding, is very difficult to reason about the energy requirements of a “valid solution”, and linearly scaling the difficult to adjust to work magnitude. Useful work tends to be expensive to validate also, whereas verify for SHA256^2 POW is trivial. Ideally you want easy to reason, linearly scaling and highly asymmetric create/validate.