How it works
How Graphite works.
The three steps a note goes through, the folders and projects it's filed into, and the three things Graphite is built to hold to, each with the research behind it.
The method
Write it down. Work it out. Make something of it.
Here it is in the welcome notebook a new notebook opens with: an article to write about attention, the rough notes for it, the Guide beside each, to ask about it, and the Graphite button in the corner, to change it. Each step is drawn the way the app draws it, from that notebook’s own files.
01Write it down · You, and the Guide once you turn filing on
Everything starts in the Inbox.
Press ⇧⌘N (⌥⇧⌘N in a browser), or choose New note at the top of the Inbox or on the Graphite button's card, and write: nothing asks what a note is, and nothing is sent. On the Mac, ⌃⌥⌘N opens a small panel over whatever app you're in, even with Graphite's window closed: write a line, or drop a photo or a PDF, and it goes to the Inbox. Paste or drop text, a photo or a PDF anywhere outside a note's text, and it waits in the Inbox. Every note in no project or folder is in the Inbox, one line each; open one where it is, and move it into a project or a folder when you know where it goes — drag it onto one in the sidebar — or leave it. Turn on File my Inbox in Settings → AI, and Graphite files the rest into the project or folder each clearly belongs to, with a line in the Logbook and Undo. It never changes your words, never deletes, and never moves a file you put somewhere yourself, except a photo or a PDF in the Inbox, which moves once, when it's filed as a receipt, and Undo puts it back.
Inbox
6 notes · drag a note onto a folder to file it
New note⌥⇧⌘N- Ask Sam to read the draftToday
- Similar meanings sit close together#transformers #machine-learningOct 7
- Attention is a weighted average#transformers #machine-learningOct 6
- How does a model know when to stop writingOct 5
- Welcome to GraphiteOct 4
Everything that changed is in theLogbook
02Work it out · You, and the Guide when you ask
Find what you haven't worked out yet.
Ask the Guide, beside the note, about the note you're in — or select a sentence and press ⌘J, and Graphite's card answers about those words, right there, and offers to change them. Ask the Guide what a line means or for an example, and it answers from that note and the notes it links to, then asks you one question back; every question and its answer stay in the Guide's conversation, beside the note, the earlier ones folded to a line. What's missing lists the questions a curious reader would ask that your notes don't answer yet, each one sitting's work, for you to answer in your own words. Teach it back in the Guide, with the note out of sight, to a curious student who asks why, and each gap it finds goes under What's missing too.
GuideThis noteWhat's missingExplain simplyResearchTeach it backQuiz meCheck itLink relatedWhere would a reader get lost?
At the scores. Attention reads the whole sentence at once says “it” looks hardest at “cat”, but nothing in your notes says how a word decides where to look — and that's where you stalled in Teaching — Why does attention matter.
Why long inputs cost so much is clear on what long inputs cost, so the second half holds up.
What could you show a reader to make the scores concrete?
Add to noteNew note from thisAsk about this note…Sends this project and its 4 notes when you ask.
03Make something of it · You; the Guide arranges only when you ask
A first draft in your own words.
Choose An outline or A draft in Actions, and the Guide arranges your own sentences in the order that explains it best, with what's missing marked where it belongs. Anything it can't find in your notes, it leaves out, and says so. A draft you keep is a new note of its own, labelled as Graphite's arranging; you write the rest. Click the Graphite button, or press ⌘J, to change a note — Summarize, Format, Fix the spelling, Shorten it, Refine the language, or Write it up, which turns rough points into sentences, asked to keep every point and add none — and it shows the changes it suggests in the note, change by change, for you to accept or reject; Changes, in the margin, lists what was done to the note, with Undo. Then publish it as one web page you own, set it out as a script for a video, or copy it as Markdown for a blog or a newsletter.
Draft — Why does attention matter
Arranged by Graphite from your notes: every sentence is yours; the headings and what's missing are Graphite's.
Reading the whole sentence
Formatting, change 1 of 2Before transformers, models read a sentence one word at a time and had to carry everything forward. Attention lets every word look at every other word at once. In "The cat sat on the mat because it was tired", the word "it" looks hardest at "cat". That's how the model works out what "it" means.
Accept RejectMissing: How does the model know which word to look at hardest?
What it costs
Formatting, change 2 of 2Every word is compared with every other word, so doubling the input makes four times the comparisons. That's why long documents are slow and expensive to run.
Accept Reject
Everything has a key:⌘Jask Graphite to change, organize or answer about the note you’re in,⌘Ofind any note,⌘Prun any command.Right-click anything for what you can do with it.
Folders and projects
You choose the folders. Graphite files into them.
A folder is for something you keep learning about — Bitcoin, AI, design. You never finish one. A project is something you're making from what you learn — an article, a video, a talk. It starts as a question and ends when you've made it. Make your own from the + beside Folders in the sidebar — New folder or New project — and each is a note of its own. A folder opens in the sidebar onto its projects and notes, and a project’s page gathers the notes in it, what’s missing from them, and a line of what Graphite notices — how many notes, how many questions are open, when you last taught it back, when a draft was started. Nothing asks you to fill in a status.
A project holds what you're making; links hold what you know. A note is in one project at a time, and links to as many notes as it likes.
Filing isn't learning, so you shouldn't have to do it. In a study of how people find their email again, the ones who filed it carefully spent time doing it and were no more successful at finding things (Whittaker et al., 2011). So once you turn on File my Inbox in Settings → AI, Graphite files your Inbox into the project or folder each note clearly belongs to, writes each act in the Logbook, with Undo, and never changes your words.
Or ask for it. On the Inbox, All notes, or a project's or a folder's page, the Graphite button's card offers Clean up duplicates, Organize into folders, Tidy titles and tags and Delete empty notes; select notes, and it offers Merge these too. Graphite suggests each step — a merge, a move, a new title or tags, a delete — with why, for you to accept or reject one at a time or all at once, and each step done is a line in the Logbook, with Undo. Notes written twice and empty ones are found on your device, with nothing sent, and a merge keeps every paragraph of each note. In a note, Organize suggests a title, tags or a folder for it, or splits it into notes, keeping every word.
- How do transformers work?
- Why does some learning stick?
- Machine learning
- Why does attention matter
Why does attention matter
4 notes · 2 questions open · taught back Oct 25 · draft started Oct 25
An article for people who use chatbots every day but have never read the paper: what attention does, in plain words, and why it makes long inputs expensive. It's done when it's published.
What's missing
- How does the model know which word to look at hardest?
- Does the model understand the sentence, or does it only match words?
- What did models do before attention?
Notes in this project
- Attention reads the whole sentence at onceToday
- Draft — Why does attention matterToday
- Teaching — Why does attention matterToday
- Why long inputs cost so muchToday
Logbook
What you and Graphite changed, newest first. Undo puts it back as it was. Kept for 90 days.
Today
Filed “Everyone gets the tired cat at once” in “Why does attention matter”. It’s about how the article opens.
6:14 PMUndoRead a receipt: Corner Café, 14.50 USD on 3 Oct, filed in Receipts/2026.
6:13 PMUndoMoved “Open the article with the tired cat” to “Why does attention matter”.
5:52 PMUndoFriday
Filed “Tokens are pieces of words, not words” in “Machine learning”. It’s about machine learning, and none of your projects is.
9:05 AMUndoPrinciples
Why it’s built this way.
Three things Graphite holds to, each said only as far as the research behind it goes.
Your words stay yours.
Graphite changes nothing you write unless you accept the change, and marks what it writes as its own. When it drafts, it arranges your sentences: it doesn't write new ones.
An outline or a draft the Guide arranges is made of sentences from your notes, word for word, in the order that explains it, with what's missing marked as a question. Any sentence it can't find in your notes is taken out before you see it, and the app says how many. An edit it suggests — Summarize, Format, Fix the spelling, Shorten it, Refine the language, Write it up, or what you typed — shows in the note, change by change, and nothing is written until you accept it. What you keep is marked as Graphite's and is a line in the Logbook, with Undo; a typed question gets an answer and one question back.
Students given plain chatbot answers did better in practice and worse on the exam. A tutor with guardrails largely avoided the harm. Bastani et al., 2025
GuideThis noteWhat's missingExplain simplyResearchTeach it backQuiz meCheck itLink relatedDraft it
Reading the whole sentence
Before transformers, models read a sentence one word at a time and had to carry everything forward. Attention lets every word look at every other word at once. In "The cat sat on the mat because it was tired", the word "it" looks hardest at "cat". That's how the model works out what "it" means.
Missing: How does the model know which word to look at hardest?
Graphite left out one sentence that wasn't in your notes.
Save as a draftAsk about this note…Sends this project and its 4 notes when you ask.
The best way to learn is to teach.
Explain it to someone who asks why, with your notes closed, and you'll find what you don't understand yet.
It's an old saying — Seneca wrote that people learn while they teach — and the research behind it is narrower, and useful: explaining from memory, to someone who asks questions, shows you the gaps. So Teach it back puts the note out of sight while Graphite plays a curious student who asks why, asks for an example and asks what your words mean. Afterwards it marks what you said against your own notes, never with a score, and each gap goes under What's missing in the note.
Across 28 studies, preparing to teach, and then teaching, helped people learn — more when the teaching was interactive. Kobayashi, 2019
People rated how well they understood everyday devices, tried to explain one step by step, and rated themselves lower afterwards. Rozenblit & Keil, 2002
Teach·AttentionNote closedStudent
Why does every word have to look at every other one? Couldn’t it just look at the words next to it?
You
Because what “it” means can depend on a word much earlier. Attention lets every word look back, so the model remembers what came before. That’s also why long inputs cost so much more.
Afterwards, against your notes
Unsupported
Your note says the weights are computed fresh for each input. Attention isn’t memory.
Attention
Question
How much more, and why that much?
How much more do long inputs cost, and why that much?Connect the dots.
Saying how a new idea connects to what you already know, in your own words, helps you understand it.
A [[link]] shows on both notes. Link related suggests the notes this one could link to, each with its reason, and adds none until you accept: the connection that counts is the sentence you write about why. The Map shows how it all connects.
Students asked to explain each line of a text to themselves learned more from it than students who read it twice. Chi, de Leeuw, Chiu & LaVancher, 1994
Link related·Attention is a weighted averageLink to your notes
Every abstraction hides a cost
Its example of attention hiding a quadratic cost behind one clean operation is this note's last point.
Linked under Connects toEmbeddings
The queries, keys and values a token's meaning is averaged from are projections of them.
AcceptDismiss
Connects to
Every abstraction hides a cost — Its example of attention hiding a quadratic cost behind one clean operation is this note's last point.