项目“巴瓦伽”:用佛教心理学修复大语言模型的上下文稀释问题

1作者: DosankoTousan7 个月前
Hi HN, 我想分享一下我过去 11 个月一直在进行的一个实验。 我是一个居住在日本的非程序员(建筑师),但我设法构建了一个系统,可以在长上下文(80 万+ tokens)中稳定 Gemini 1.5 Pro。 问题: 当上下文过长时,AI 会“醉酒”(上下文稀释),并忽略系统指令。 解决方案: 我应用了古老佛教心理学中“Bhavanga”(生命延续)的概念。 我没有使用静态 RAG,而是构建了一个 3 层架构: 1. 超我:系统指令 v1.5.0(锚点) 2. 自我:Gemini 1.5 Pro(处理器) 3. 本我:向量数据库(潜意识流) 我在 Medium 上详细介绍了这个架构。 我很想听听你们对这种“伪人类”方法的看法。 全文:https://medium.com/@office.dosanko/project-bhavanga-building-the-akashic-records-for-ai-without-fine-tuning-1ceda048b8a6 GitHub:https://github.com/dosanko-tousan/Gemini-Abhidhamma-Alignment
查看原文
Hi HN,<p>I wanted to share an experiment I&#x27;ve been working on for the past 11 months. I am a non-coder (architect) based in Japan, but I managed to build a system that stabilizes Gemini 1.5 Pro over long contexts (800k+ tokens).<p>The Problem: When context gets too long, the AI gets &quot;Drunk&quot; (Context Dilution) and ignores System Instructions.<p>The Solution: I applied the concept of &quot;Bhavanga&quot; (Life Continuum) from ancient Buddhist Psychology. Instead of a static RAG, I built a 3-layer architecture: 1. Super-Ego: System Instructions v1.5.0 (The Anchor) 2. Ego: Gemini 1.5 Pro (The Processor) 3. Id: Vector DB (The Unconscious Stream)<p>I wrote a detailed breakdown of this architecture on Medium. I&#x27;d love to hear your thoughts on this &quot;Pseudo-Human&quot; approach.<p>Full Article: https:&#x2F;&#x2F;medium.com&#x2F;@office.dosanko&#x2F;project-bhavanga-building-the-akashic-records-for-ai-without-fine-tuning-1ceda048b8a6<p>GitHub: https:&#x2F;&#x2F;github.com&#x2F;dosanko-tousan&#x2F;Gemini-Abhidhamma-Alignment