The long term target is JEPA-like primitives so models become irrelevant. You only have to accelerate critical paths to see massive gains.