初创公司 Leeroo 宣布将其 AI 引擎 Kapso 开源,该项目由曾就职于 LinkedIn 和 Meta 的开发者 Alireza 与 Majid 打造 1。Kapso 的核心机制是“教训库”(bank of lessons),用于存储从实验中归纳的规律、置信度分数及适用条件,旨在通过迭代实验和从过往运行中学习来自动化企业数据处理任务 1。该系统能根据环境反馈优化解决方案并部署到目标技术栈中 1。
官方称 Kapso 在多项基准测试中表现优异 1。它在 RelBench 预测建模基准测试中排名第一,超越了 Kumo AI (Nvidia) 和斯坦福的最佳表格基础模型 1。此外,该引擎在 OpenAI 的 MLE-bench 提交时排名亦为第一 1。在 2026 年国际 AI 奥林匹克竞赛(IOAI 2026)的 AI 模型赛道中,Kapso 自主运行排名前 3,高于所有 471 名人类参赛者 1。
在实际应用方面,Leeroo 展示了一个遗留数据系统迁移案例 1。一家拥有 30 年历史的企业,其数据系统向现代湖仓架构的迁移项目停滞一年后,仅用数周便完成了这一任务 1。作为生态支撑的一部分,公开知识库 Leeroopedia 拥有约 10 万页内容,月请求量约 600 万次 1。Kapso 的代码已托管在 GitHub 上,地址为 https://github.com/Leeroo-AI/kapso 1。
Leeroo has announced the open-source release of Kapso, an artificial intelligence engine designed to automate enterprise data processing tasks through iterative experimentation and learning from past operations 1. The system is available on GitHub at https://github.com/Leeroo-AI/kapso and was developed by Alireza and Majid, engineers with prior experience at LinkedIn and Meta 1. Kapso’s core mechanism relies on a "bank of lessons," which stores patterns derived from experiments alongside confidence scores and applicability conditions, allowing the system to optimize solutions based on environmental feedback before deploying them to target technology stacks 1.
The developers claim significant performance achievements for Kapso across several benchmarks and real-world scenarios 1. In the RelBench predictive modeling benchmark, the engine ranked first, outperforming Kumo AI (Nvidia) and Stanford’s best tabular foundation models, while also securing the top position upon submission to OpenAI’s MLE-bench 1. Additionally, during the 2026 International Olympiad in Artificial Intelligence (IOAI), Kapso operated autonomously in the AI model track, placing in the top three and scoring higher than all 471 human competitors 1. A cited case study illustrates its practical utility: Leeroo used Kapso to migrate a legacy data system from a company with 30 years of history to a modern lakehouse architecture, completing in weeks a project that had previously stalled for over a year 1. Supporting this ecosystem, Leeroo also maintains Leeroopedia, a public knowledge base containing approximately 100,000 pages that receives around 6 million monthly requests 1.
评论
还没有评论,欢迎留下第一条。