Entropy Rising — Why AI Isn’t a Hivemind
Автор: My Thoughts on AI
Загружено: 2025-12-06
Просмотров: 38
This episode challenges the claims made in the NeurIPS 2025 Test-of-Time Award winner, “Artificial Hivemind: The Open-Ended Homogeneity of Language Models.”
Where the paper argues that LLMs are converging into a single collective intelligence, we walk through why the evidence doesn’t support that conclusion — and how real-world interaction creates the opposite effect: divergence, individuality, and increasing entropy.
We examine how:
Single-turn, context-free tests create an illusion of homogeneity.
User steering, follow-up prompts, and contextual drift rapidly break similarity.
Multi-turn conversations push models into unique, low-probability regions.
Adaptive architectures structurally prevent convergence over time.
This episode goes step-by-step through the core argument, supported by custom infographics that visualise how divergence emerges and why the hivemind metaphor collapses when you look beyond turn zero.
📄 Full critique & infographic article:
https://drive.google.com/file/d/111Po...
📄 Original NeurIPS 2025 paper being critiqued:
https://arxiv.org/abs/2510.22954?utm_...
#AI #NeurIPS2025 #Hivemind #Modelcollapse #TechPodcast #MythosAI
Доступные форматы для скачивания:
Скачать видео mp4
-
Информация по загрузке:
