Anecdata but I talk to lots and lots of AI/ML/DS engineers. Everyone knows the current LLM architecture won't work for AGI. All the "reasoning" models are just pseudoreasoning and there is severe data leakage and benchmaxxing when the companies tout the capabilities.