Day page

1 Jul 2026

Paper summary

Paper summary: Orca: The World is in Your Mind

Orca learns one latent "world state" space from 125K hours of video and 160M event annotations via next-state prediction, and frozen, it supports text, image and robot-action readouts through small decoders.

The objective is next-state prediction: not next token, next frame or next action, but how the world's state changes.
"Unconscious" learning from dense video transitions is paired with "conscious" learning from language-described events.

Read the post →

Study

Generative AI · Stable Diffusion

Latent diffusion runs the diffusion process in a compressed latent space from a VAE; encoding and decoding a real photo with the Stable Diffusion VAE to see the 48× compression the whole model relies on.

Stable Diffusion doesn't denoise pixels: it denoises a 4×64×64 latent and decodes it with a VAE at the end.
That latent has 48 times fewer numbers than the 512×512×3 image, which is what makes training and sampling affordable.

Read the post →