注册并分享邀请链接,可获得视频播放与邀请奖励。

Jon Shulkin
@jon
加入 November 2011
435 正在关注    6.9K 粉丝
Responding to the questions I'm getting on "how do I use AI to ask questions that securely access all my private files" problem. Your files are the company. The model has never read them and they don't fit in the context window. Connectors are not good enough when precision matters. Grok 4.7 enabled the process and workflow described below. The Brain is the memory. Not a trained model with your documents baked into the weights. Every file is cut into windows, and the page stays attached. Search is the words on the page first. Vectors only re-rank that shortlist. The wiki is compiled from the record. The model does not write it. Intelligence is the job. A question comes in. An always-on machine, on the premises, claims one job at a time. Empty workspace. No source code. No prior answers. No one else’s credentials. That machine is not optional. It is the security boundary. Grok 4.7 does the work. Think, search, read, compute, write, critic. If the number does not re-add, it goes back. If the date or the folder is missing, it asks. Silence is not a yes. Ask does not post the books. A person still posts. Model context protocol "MCP" is the only door from that machine back to the record. The credential is a person. Same grants as that person. The model does not get a database password. Money never moves. If the audit fails, the call fails. 10,000+ private files. One record. The model can be swapped. The files, the grants, and the audit cannot.
显示更多
0
10
25
2
转发到社区