I favor betting on proper orders of magnitude more rounds of reliable and proven constructions optimizing for a minimum mem/cpu/gpu/fpga- & asic-hard cost factors for a particular use-case than entirely new and unproven constructions (recall SIDH) and from untrustworthy intermediaries like NIST (recall P curves, Dual_EC_DRBG).

PS: We'll sure be in relative trouble if we find faster &| cheaper ways to do prime factorization or invert matrixes.