The three levers.
A generic model becomes yours in three places: its weights, what it looks up at question time, and what it is told at the start.
new words
- Weights
- The numbers inside the model, billions of them. Training sets them, and nothing else changes them.
- Training
- Changing the weights with examples. Pretraining builds the model from nothing; fine-tuning adjusts a finished one with your examples.
- Retrieval
- Finding the right paragraph at question time and putting it into the prompt. Lessons 3 to 7 take it apart.
- Context
- Everything the model is given before and in your message: instructions, pages, the results of a lookup. Claude basics, lesson 2.
the lesson
Every company uses the same models. What makes one AI yours is where you put what it must know and how it must act. There are three places: the weights, set by training; the paragraphs fetched into the prompt; and the pages and instructions it reads first.
| Lever | Facts | Behaviour | Fresh |
|---|---|---|---|
| Weights | Poor, with no source | Yes | Stale from the day training ends |
| Retrieval | Cited: the answer names its chunk | No | After the next reindex |
| Context | Whole, in the window | By instruction | On save |
Facts go in context or retrieval. A way of answering goes in an instruction file, which costs nothing to change, or in a fine-tune. Judgment is the lab's job: pick a better model.
Fine-tuning fails three tests: fresh, cited and private. Context passes all three when the store is yours. instibrain is the third lever: your pages are read at the start of every session, and the folder you start in decides which ones.
try it
Type this and press Enter:
What do you know about my company? For each fact, say where it came from: your training, a page you read at the start, or a lookup you made just now.you should see
- Every fact names a page or a lookup. Nothing about your company came from training.
- If Claude read nothing yet, it says so and offers to read your spaces.