展示HN：自主复制哈佛关于人工智能对就业影响的研究
我们使用NeuGBI在相同的Revelio Lab数据集上复制了《生成性人工智能作为资历偏见技术变革》（哈佛商学院，2025）的研究，该数据集包含3亿条美国就业记录。<p>论文的发现是：人工智能对初级职位的影响显著高于高级职位（初级职位减少29.4%，高级职位减少5.8%）。NeuGBI自主得出了相同的结论。<p>NeuGBI发现了一点论文中没有提到的：在软件开发领域，受到影响的主要是初级（L2）职位，几乎减半，而不是入门级（L1）职位。<p>NeuGBI使用NeuG（一个支持多跳关系的图数据库）作为查询引擎，采用超图重构进行分析，并打包了探索性技能，供大型语言模型（LLM）调用，以逐步分解问题并深入探讨。<p>NeuGBI的关键能力是端到端的无偏采样——在3亿条记录上，复杂的多跳查询可以在几秒钟内返回，而不是几个小时。<p>博客文章：<a href="https://graphscope.io/blog/tech/2026/06/16/NEUGBI-BLOG.html" rel="nofollow">https://graphscope.io/blog/tech/2026/06/16/NEUGBI-BLOG.html</a>
原始论文：<a href="https://arxiv.org/abs/2603.10625" rel="nofollow">https://arxiv.org/abs/2603.10625</a>
查看原文
We used NeuGBI to replicate "Generative AI as Seniority-Biased Technological Change" (HBS, 2025) on the same Revelio Lab dataset — 300M U.S. employment records.<p>The paper's finding: AI disproportionately affects junior positions (−29.4%) vs. senior (−5.8%). NeuGBI arrived at the same conclusion autonomously.<p>One thing NeuGBI found that the paper didn't: within software development, it's specifically junior-level (L2) positions that nearly halved, not entry-level (L1).<p>NeuGBI uses NeuG (a graph database with multi-hop relationship support) as its query engine, Hypergraph reconstruction for analysis, and packaged exploratory Skills that an LLM can invoke to decompose questions and drill down step by step.<p>The key capability of NeuGBI is end-to-end unbiased sampling — on 300M records, complex multi-hop queries return in seconds rather than hours.<p>Blog post: <a href="https://graphscope.io/blog/tech/2026/06/16/NEUGBI-BLOG.html" rel="nofollow">https://graphscope.io/blog/tech/2026/06/16/NEUGBI-BLOG.html</a>
Original paper: <a href="https://arxiv.org/abs/2603.10625" rel="nofollow">https://arxiv.org/abs/2603.10625</a>