Actually we do, it's widely believed by top researched not sponsored such as Yann LeCun that we are delaying actual progress on AGI by aiming for LLMs.
There's a buuuunch of reasons why AGI is still very far off, like that our current Von Neumann architeture is still very far away from resembling sophisticated neurons.
Trying to achieve AGI with current Transformer models is like trying to get to Mars by putting more blades and fuel into a helicopter.
I simply don't agree. Denoising with temporal accumulation will always be inherently blurry. Even with ray reconstruction it can get very ugly at times. Worst offender is a moving shadow in a low light situation. Blurry mess. You need to brute force that with raw performance to avoid that, it's the only way. I predict we will never get to that point ever with GPUs working as they currently do.