Seonglae ChoBoth models are kept frozen; only the bridge is trained.
It closes 50% of the accuracy gap between the small and large models. The paper claims a 25% relative improvement in accuracy over the small model, and 2.5x less compute than a mid-sized model of comparable performance.
mostik.ai — latent communication between AI models
The channel that lets AI models exchange computation instead of text.
https://mostik.ai/read-more
