初创公司 Plurai 宣布完成由 Team8 领投的 1300 万美元种子轮融资,旨在通过借鉴自动驾驶技术的仿真方法来解决 AI 代理可靠性不足的问题 1。该公司联合创始人 Ilan Kadar 和 Elad Levi 曾从事自动驾驶技术开发,他们观察到 AI 代理领域正重现自动驾驶早期的挑战模式,因此决定将成熟的仿真验证思路引入这一新兴领域 1。Kadar 表示:“当我们看到 AI 代理浪潮时,我们看到了同样的模式……我们要构建这个仿真层。”1
当前主流 AI 模型(如 Claude Fable 5.2, ChatGPT-6 Astra)在复杂任务基准测试中准确率约为 75%-83%,显示出不稳定性 1。为应对这一瓶颈,Plurai 的技术核心是创建企业电子系统的“数字孪生”,并生成合成数据用于模拟训练,以优化 AI 模型处理边缘案例的能力 1。
在与 Nvidia 和 Nebius 的合作测试中,该方法被证实可训练出比中等前沿模型更具成本效益的小型开源权重模型 1。这些小型模型在保持性能的同时,大幅降低了推理成本 1。
Startup Plurai has secured a $13 million seed funding round led by Team8, aiming to tackle the reliability issues of current AI agents by applying simulation techniques similar to those used in autonomous vehicle development 1. Co-founders Ilan Kadar and Elad Levi previously worked on self-driving technology and recognized a parallel between the challenges faced by early autonomous cars and modern AI agents 1. "The moment that we saw the wave of AI agents, we saw the same pattern... we want to build this simulation layer," said Ilan Kadar 1.
Plurai’s core solution involves creating "digital twins" of enterprise electronic systems and generating synthetic data to simulate training scenarios 1. This method is designed to help AI models handle edge cases more effectively, addressing instability observed in mainstream models like Claude Fable 5.2 and ChatGPT-6 Astra, which currently achieve accuracy rates of approximately 75% to 83% on complex task benchmarks 1.
In partnership tests with Nvidia and Nebius, Plurai demonstrated that its approach can train smaller open-weight models that offer greater cost-efficiency compared to mid-tier frontier models while maintaining performance 1.
评论
还没有评论,欢迎留下第一条。