diffusive AI context imagination

I suddenly hit upon a new way to compress AI context. Since our human understanding of a project is always holistic and abstract, why can't AI do the same? A recent paper from ByteDance suggests using a diffusion model to replace traditional next-token prediction in LLMs. Essentially, it denoises within a latent space to sketch out a "broad, abstract meaning" before using a VAE decoder to turn it into individual tokens. ​Maybe AI context could also pivot to this kind of abstract, latent approach. It would massively boost context capabilities by breaking free from the limitations of tokens. Plus, it would drastically drive down costs for developers, since storing a latent tensor requires far less compute than keeping every single token previously generated. Reflections after spending an entire afternoon battling Cursor's atrocious context management.