AI can only be as good as its available training. The only improvements have been clever ways to approach that asymptote, but no true work arounds. To get them to truly learn, would require algorithms better than diffusion and backpropagation, and those currently do not exist.