这方面，国内的光帆科技（Lightwear）其实已经给出了先行版本——他们在耳机上装了一对摄像头，连续不间断地捕捉现实世界，但这些信息在 AI 使用之后便「阅后即焚」，既保证可用性，又具备私密性——这套方案，是不是也有点果味儿？
圖像來源，Getty Images
，更多细节参见下载安装 谷歌浏览器 开启极速安全的 上网之旅。
"We have to let them participate in that. But I do think that access to children should be conditional, and it should be conditional on respecting their rights, on keeping them safe, on actually not addicting them and not extracting economic value out of their behaviour."
Anthropic’s prompt suggestions are simple, but you can’t give an LLM an open-ended question like that and expect the results you want! You, the user, are likely subconsciously picky, and there are always functional requirements that the agent won’t magically apply because it cannot read minds and behaves as a literal genie. My approach to prompting is to write the potentially-very-large individual prompt in its own Markdown file (which can be tracked in git), then tag the agent with that prompt and tell it to implement that Markdown file. Once the work is completed and manually reviewed, I manually commit the work to git, with the message referencing the specific prompt file so I have good internal tracking.