They're banking on diffusion eliminating the hallucination problem, but it's too onerous to run very, very large models with it yet. Auto regressive LLMs are a dead end. One that is far away. LLMs are not and we will continue to learn a lot about them as we continue to implement them. Anyone who thought we were at a dead end should use the new Gemini. It's like a GPT 3.5 to GPT 4 level of improvement.