With the current tech, AGI is literally impossible. They believed scaling, clever architecture, and context window tricks would solve the issues of the current LLMs, but fundamentally diffusion and backpropagation cannot iteratively learn as they have no way to train during inference.