I don't really agree.

Recent AI inovations are pretty modest and use the innovation of raw fucking power to achieve goals.

Gpt4 uses 230B parameters, whereas to run a 7B LLM you need 16gb of vram already, and llms are o(n²) in complexity in terms of parameters, I'll let you do the maths

Stable diffusion (latent diffusion to be more precise) is about the same, the initial training required billions of teraflop, while it was relatively cheap (100k$), it still rides on modern GPU technology .