Elorian’s Glimpse model aims to reason about images by mentally editing and examining them, rather than relying on visual shortcuts. On the DecoyCount test, it scored higher than the other models tested, though its result came from just one run. Elorian says this kind of visual reasoning could help with real-world tasks and may be essential for artificial general intelligence.
ORIGINAL LINK: https://latent.space/p/elorian/
