LLM은 하이브마인드처럼 작동한다 — Springboards가 그 주문을 깨려고 Flint를 만들었다
Springboards가 환각 감소보다 무작위성을 우선해 학습한 LLM 'Flint'를 출시하며, 개방형 질의에서 주류 모델들이 똑같은 출력으로 수렴하는 '집단사고(groupthink)' 문제를 해소했다.
australian startup springboards has released flint, an llm specifically trained to maximize response variety and mitigate the 'artificial hivemind' effect observed in mainstream models. testing revealed that while dominant models like chatgpt and claude overwhelmingly converge on predictable outputs—such as selecting '7' for random numbers or 'toyota' for car examples—flint successfully generates diverse alternatives like '3.7916' or 'ford f-150.' this development challenges the prevailing industry safety standard of suppressing hallucinations in favor of encouraging controlled randomness for better ideation and creativity. it highlights a growing concern over model homogeneity, a topic recently highlighted in a award-winning paper at neurips regarding the 'open-ended homogeneity of language models.'