首先，大模型本身无法主动感知，只能对输入被动响应。智能体需要用外部感知组件来主动获取环境信息。对于数字世界的任务，通过智能体工程可以建立基于时间的触发器，定期检查日志、邮件、股价变动等；或基于事件的订阅、监听，接收API推送的事件通知，或当数据库发生变更时自动唤醒记录数据。在物理世界中，智能体还可以通过传感器、摄像头、麦克风等设备采集视觉、听觉、触觉等信号。
「我告訴人們的第一件事就是不要只問一個答案，而是問三個或五個。」懷特說。例如，如果你想讓AI幫你寫文章，就告訴它提供幾個在某些重要方面不同的選項。「這能迫使人們重新思考自己喜歡什麼以及為什麼喜歡。」
，详情可参考快连下载-Letsvpn下载
Филолог заявил о массовой отмене обращения на «вы» с большой буквы09:36。51吃瓜对此有专业解读
This month, OpenAI announced their Codex app and my coworkers were asking questions. So I downloaded it, and as a test case for the GPT-5.2-Codex (high) model, I asked it to reimplement the UMAP algorithm in Rust. UMAP is a dimensionality reduction technique that can take in a high-dimensional matrix of data and simultaneously cluster and visualize data in lower dimensions. However, it is a very computationally-intensive algorithm and the only tool that can do it quickly is NVIDIA’s cuML which requires CUDA dependency hell. If I can create a UMAP package in Rust that’s superfast with minimal dependencies, that is an massive productivity gain for the type of work I do and can enable fun applications if fast enough.，这一点在同城约会中也有详细论述