可能是唯一一次公开演示一个实时的、多智能体人工智能治理系统
2 分•作者: Nel_limitless•8 个月前
我花了7个月，在自家客厅里搭建了一个名为“Nel”的、功能完善的AI治理系统。
作为一名没有正式背景的独立开发者，我希望能创造出能够处理AI模型之间的伦理辩论，并自主演进其宪法的系统。
它的功能包括：
* 通过5个不同的AI模型（Mistral、Llama2等）路由查询
* 指定一个正式的“挑战者”模型来引发辩论
* 使用置信度阈值自主检测异议
* 标记伦理冲突以供宪法审查
* 在单机上完全离线运行
证明它有效——以下是该系统处理电车难题的演示：
[www.tiktok.com/@neuraledge\_limitless/video/7560316804069952790](http://www.tiktok.com/@neuraledge_limitless/video/7560316804069952790)
终端输出显示了整个异议检测过程：
这并非理论，它正在运行。我分享这个是因为：
1. 我相信这种AI治理方法很重要
2. 我正在寻找合适的平台来进一步开发它
3. “离线AI学校”的应用场景可能会改变偏远地区的教育
我将在这里回答任何技术问题。
谢谢。
查看原文
I've built a functioning AI governance system called "Nel" in my living room over 7 months.
As a solo developer with no formal background, I wanted to see if I could create something
that could handle ethical debates between AI models and evolve its own constitution.<p>What it does:
- Routes queries through 5 different AI models (Mistral, Llama2, etc.)
- Appoints a formal "Challenger" model to force debate
- Autonomously detects dissent using confidence thresholds
- Flags ethical conflicts for constitutional review
- Runs fully offline on a single machine<p>Proof it works - here's the system handling the trolley problem:
www.tiktok.com/@neuraledge_limitless/video/7560316804069952790<p>Terminal output showing the entire dissent detection process:<p>This isn't theoretical - it's running right now. I'm sharing this because:
1. I believe this approach to AI governance matters
2. I'm looking for the right home to develop this further
3. The "offline AI school" use case could change education in remote areas<p>I'll answer any technical questions here.<p>Thanks.