AI研究人员贾扬清离开英伟达约一个月后创办了Intent Lab，并启动了名为Fleet的自主AI团队项目[1]。该项目旨在将用户意图转化为生产级软件系统，通过多项技术突破展示了AI在完整软件工程流程中的能力[1]。
Fleet在推理优化方面取得显著成效[1]。通过优化TensorRT-LLM，该团队使GLM-5.2的推理速度从102token/s提升至647token/s，实现了534%的性能提升[1]。
在数据库开发方面，Fleet从零开始构建了SQLite兼容数据库，产生8720万个输出token，通过了全部600万项验收测试[1]。该项目使用Opus 4.8的成本约2000美元，若采用开源模型成本可降至约350美元[1]。同时，Fleet还创建了分布式文件系统AgentFS，目录状态查询速度达到EFS的626倍，文件状态查询速度达到EFS的625倍[1]。在测试中，Fleet检查了约190万个可达状态，完成了约300项集成测试[1]。
Jia Yangqing, an AI researcher who recently left NVIDIA, has founded Intent Lab and unveiled Fleet, an autonomous AI team designed to transform user intent into production-grade software systems [1]. The project has demonstrated significant technical achievements across multiple domains, showcasing AI's capacity to execute complete software engineering workflows.
Fleet has delivered three major technical accomplishments. The team optimized TensorRT-LLM to accelerate GLM-5.2 inference speed from 102 tokens per second to 647 tokens per second, representing a 534% improvement [1]. Additionally, Fleet built a SQLite-compatible database from scratch that passed all 6 million acceptance tests, consuming 87.2 million output tokens with estimated costs of approximately 2,000 US dollars using Opus 4.8 or around 350 US dollars with open-source models [1]. The team also created AgentFS, a distributed file system that vastly outperforms comparable solutions, achieving directory status query speeds 626 times faster than EFS and file status query speeds 625 times faster than EFS [1].
The project involved extensive testing and validation, with Fleet examining approximately 1.9 million reachable states and completing around 300 integration tests [1]. Jia Yangqing launched Intent Lab approximately one month after departing from NVIDIA [1].
