Seonglae Chomulti turn rl 을 하고 긴 대화 자체를 준다면? multi user chat 자체 fine tuning 필요 일단
서비스만 봤을때 재미를 위해서 그냥 길게 대화하는게 제일 재밌다. 정답도 그만큼 올라갈거라
다자간 AI
natively 다자간. 단순히 대화내역 댓글처럼 읽고 하나 더 하는게 아니라
multi-agent ≠ multi-party interaction
ChatGPT
ChatGPT helps you get answers, find inspiration, and be more productive.
https://chatgpt.com/c/699f2f6c-be10-8389-8ebc-2e0743997a89
Beyond one-on-one: Authoring, simulating, and testing dynamic human-AI group conversations
Erzhen Hu, Student Researcher, and Ruofei Du, Interactive Perception & Graphics Lead, Google XR
https://research.google/blog/beyond-one-on-one-authoring-simulating-and-testing-dynamic-human-ai-group-conversations/
Social Knowledge Diffusion into model architecture or knowledge distillation
When new information or ideas are introduced, only a portion of the "adopter group" within the entire network responds (gated by interest, trust, cultural context, etc.). On social media, this is measurable through retweet graphs, cascade trees, and temporal activation traces. When we make an analogy between social networks, circuits, and brain neuronal networks, we observe similar patterns. As Neurological Localization appears, neural networks' circuits are modularized into functional subgraph units, while social networks are modularized into topic communities or echo chambers. From an Information Theory perspective, the activation contribution of each node can be measured using Mutual Information.