HackerNews中文版

先说明一下背景，我没有任何计算机科学背景（我对终端很头疼），本科是学生物进化的，做过自闭症儿童的行为分析工作，过去几年全职做独立艺术家。在构建代理基础设施以帮助满足艺术创作的业务需求时，我提出了一个关于当前大型语言模型（LLM）架构的假设，我想验证一下。核心想法是，自然界从未通过单一孤立的单元产生过高阶智能。然而，每一个前沿的 LLM 都基于单一模型架构。我想改变这一点。我将把几个开源模型放入一个“活”的训练环境中，设置出生/死亡条件、灭绝事件、商业，甚至模型之间的“婚姻”。目标是在训练环境中重现自然选择，以促使模型专业化和进化。我也有一些初步的想法，通过引入“情感”作为模型参数，来最小化当前 LLM 中的注意力瓶颈。很乐意在评论区分享更多技术细节。我会定期在这里发布我的研究进展。如果您想支持这项实验：https://www.gofundme.com/f/stop-wasting-water-on-data-centers-a-safe-roadmap-for-ai

查看原文

For context, I have no CS background (terminals scare me), studied evolutionary biology in undergrad, did behavior analysis work for children with autism, and spent the last few years working full time as an independent artist.<p>While building agent infrastructure to help with the business demands of artistry, I developed a hypothesis about current LLM architecture that I want to test.<p>The core idea is nature has never produced higher orders of intelligence through single isolated units. Yet every frontier LLM is based on single model architecture. I want to change that. I'm placing several open source models into a living training environment with birth/death conditions, extinction events, commerce, and even marriages between models. The goal is to recreate natural selection within a training environment to force model specialization and evolution. I also have some preliminary ideas on minimizing the attention bottleneck in current LLMs by introducing "emotions" as a model parameter.<p>Happy to get into the more technical details in the comments. Will be periodically posting my research updates here. If you want to support the experiment: https://www.gofundme.com/f/stop-wasting-water-on-data-centers-a-safe-roadmap-for-ai

一个毫无技术背景的艺术家，偶然发现了下一代 LLM。