This paper teaches AI to pay attention better by training its focus, not just its words.
The paper studies how to teach a smaller language model using a bigger one by only focusing on the most useful bits instead of everything.
Typhoon-S is a simple, open recipe that turns a basic language model into a helpful assistant and then teaches it important local skills, all on small budgets.
LiveTalk turns slow, many-step video diffusion into a fast, 4-step, real-time system for talking avatars that listen, think, and respond with synchronized video.