AI models still far from AGI-level reasoning: Apple researchers