Docs

SeeMaps

On this page

How AI works

Your AI is layer 1. jade.os is layer 2, on your computer. Here is where each part runs, and how one request moves through them. Tap a box, or press Play to follow a request.

The desk is everything your AI is given to read in one turn: its instructions, your memory, the chat so far and your message. AI makers call it the context, and its size limit is the context window.

The map

Your computer
Internet
Layer 1 · your AI, in its data center
jade.os · layer 2
Rows of servers, each full of chips encrypted reply
  1. You ask for something, on any page.
  2. jade.os sets your AI's desk: your memory, this page's helper, the chat so far, and your message.
  3. The whole desk goes to your AI company, over an encrypted connection.
  4. In the data center, the model reads everything and writes the reply one token at a time, on rows of chips. All of this is layer 1, your AI.
  5. The reply comes back to your computer. When it asks to use a tool, jade.os runs it here.
  6. The result goes back on the desk and out again. Look, act, check, repeat, until the job is done.
  7. A second pass reads the finished work again before you see it.
  8. It waits on Today. Nothing is sent, posted or paid until you give your OK.
  9. What matters is saved to your memory, so every session after starts from it.
The mapTap a step, or press Play.

Play follows one request through, a step at a time.

The two layers, on this map

Layer 1 is your AI, in your AI company's data center: the chips, the model, and the work of running it. Layer 2 is jade.os, on your computer: the desk, tools, memory, the second check and your OK.

What it means for you

jade.os decides what your AI reads, runs its tools on your computer, checks its work, and waits for your OK. The model does not change while you chat. Whether your chats help train future models is a setting with your AI company. What makes it good at your business is what goes on its desk, and that is memory.

How AI memory works

Under the hood

  • The desk. Everything your AI is given to read in one turn: its instructions, your memory, the chat so far and your message. AI makers call it the context; its size limit is the context window.
  • Tokens. Models read and write in tokens, small pieces of text. Usage and plan limits are counted in tokens.
  • The math. Most of a model's work is matrix multiplication: large grids of numbers, multiplied and added. GPUs and similar chips are built to run it in parallel.
  • Attention. At each step, every token weighs the other tokens it can see. That is how a model follows context.
  • Picking a token. Every possible next token gets a probability and one is picked, so the same question can get slightly different answers.
  • Fixed weights. The weights do not change while you chat. What changes an answer is what is in front of the model.
  • Tools. A model never runs a tool itself. It writes a request, and the app runs it and sends the result back as more context.