Fromfirstlighttolastthought.
Notes, reports, dashboards and canvas in one quiet window — with an agent that works in every room and writes plain files you keep.
The Architecture
Carbon holds four bonds at once, and what it does with them is the whole story. Flatten them into sheets of hexagons and you get graphite. Point them at the corners of a tetrahedron and the same atoms become diamond.
The Recipe
Diamond is the denser arrangement, so it has to be paid for in pressure. Along a continental geotherm the two forms trade places somewhere past a hundred kilometres down — below that line graphite is the stable form; past it, carbon prefers the net.
The Extremes
Every superlative diamond carries comes from the same tetrahedral net: short bonds, no slip planes, nothing loose to lose. It is the hardest natural material, it moves heat better than copper, and its band gap is wide enough that visible light passes straight through.
The Instrument
Two brilliant cuts, tips ground flat and set facing each other. Diamond is transparent and does not yield, so you can watch a sample while you crush it — which is how we know what water does under pressure.
A drop of water is loaded between the culets, in a gasket no wider than a grain of salt.
Coda
One element, five chapters. The distance between a pencil line and a wedding ring is the angle between four bonds — and whether the lattice grew in the mantle or in a chamber, it ends at the same place.
Ask once. Get the artifact.
A journal, a research brief, a slide deck, a study guide — your agent writes it, designs it, and cites it. Then it hands you the pen.
CLICK A TYPE — THE BIG CARD FOLLOWS
Cited, honestly.
The source sits in the line you're reading, and the full list rides the bottom of the page.
DESCENT AND MEMORY · ACT I
The fix is standard: clip the gradient by norm before stepping, so a step’s length is bounded while its direction is untouched 
arxiv.org. The direction stays informative after the magnitude stops being.
But note where the cliff came from — the loss of a single number multiplied by itself twenty times. Cliffs stop being a feature of some loss surfaces and become the generic case.
arxiv.orgPascanu et al. 2013 — On the difficulty of training recurrent neural networks
Yours to edit.
Select any line — highlight it, quote it into your thoughts, or hand it to Nimbus, right where you're reading.
Two small trained matrices can stand in for a fine-tuned model — the update lives in a low-rank sliver of the weights, and the stack above it never notices.
your ink — “prove the rank claim”
Not by analogy — by intuition by building the actual mathematics, in order.
A desk, not a page.
Turn it into a book, pin a second report on top — the desk holds them all, sources included.
Every famous fossil from the Cambrian explosion postdates the explosion.
This is the first thing to be clear about, because it governs what the Burgess Shale can and cannot say.
Keep it. Send it.
A PDF for print, one interactive HTML file that opens anywhere, or the whole report to hand over.
The Memory Wall
Compute grew 60,000× in a decade; memory bandwidth grew 100×. The gap is the wall — and it is measured, not guessed.
memory-wall.pdf · 2.1 MB · SOURCES TRAVEL INSIDE
A report is a plain React file in your workspace — open it, edit it, or ask the studio to write the next one. The vocabulary report reads your words dataset and re-renders as it grows.
A cockpit for every obsession.
No stock widgets. Nimbus builds each tile from your own notes and data — and sits inside every one of them, ready to change it.
EVERY OBSESSION GETS ITS OWN COCKPIT — CLICK ONE
Built from your notes.
A widget is a view on your own prose — plans, streaks, datasets your notes already contain.
Minutes on task · week 32
Loss curve · run 41
plateauing — nudge the lr?
Sessions · 11 weeks
- Mon — 40 min, GAE by hand. Third time it clicked.
Nimbus in every widget.
“Make this weekly” is a sentence, not a settings hunt.
Your tools, streamed in.
Connector tiles stay live — and say so when a source goes quiet.
▲#build — Slack
8DLAST SYNC 8 DAYS AGO — RECONNECT?
Draw a room. Give it a rule.
Notes, reports, widgets and live outputs all sit on one infinite table. Draw a region around some of them, write down how you want that material handled, and everything that runs inside obeys.
Every room has a rule.
A cluster carries standing instructions. One source through one recipe comes out differently in a different room.
Transformer attention — raw notes
· every token attends to every other, no recurrence
· softmax(Q·Kᵀ / √d)
· O(n²), fully parallel
report · rebuilt weekly
widget · live
Attention lets a sequence model weigh every token against every other in a single parallel operation — the mechanism that displaced recurrence as the backbone of sequence modelling.
Self-attention derives three projections of each token — query, key, and value — and scores every query against every key through scaled dot-product attention.
The machine is a file.
A recipe is a .recipe.mdx file in your workspace — named ports, a prompt, yours to edit.
--- name: Synthesize handler: opencode inputs: - { id: notes, type: markdown, label: Notes, multi: true } - { id: angle, type: text, label: Angle } output: markdown ---
Ask the board.
Nimbus answers onto the canvas. New tiles arrive ringed and unaccepted — nothing lands until you say so.
Both papers assume one cache per head. The tension is whether that survives grouped-query.
Nobody cites it, and it contradicts the scaling claim both of these lean on.
Meet Nimbus. Or bring your own.
One agent works every room — and it remembers you. Prefer your own? Claude Code, Codex, and a real terminal are built in.
DRAG TO SPIN — HE’S RESTLESS
Nimbus, resident
Orchestrates notes → reports → dashboards → canvas. Carries your memory: gold standards, profile, any .md you hand it.
Reading logs and the canvas cluster — brief coming up.
The table opens on the gap, not the total — you missed Thursday.
Deadlifts are the only lift that moved — pull-ups held flat.
So: Thursday or Friday. Pick one and week 33 is a five-week run.
Your keys, your models
A real terminal in the shell. Run Claude Code or Codex against your workspace — it’s all just files to them too.
Your world, streamed in.
Connectors don’t sync to someone’s cloud. They write plain files into your workspace — where you, Nimbus, and every agent you bring can read them.
It’s all just files.
Open the folder behind the studio and there’s no database, no proprietary format — every note, recipe, and report is a file you can read, rewrite, or throw away. Which means the studio itself is yours to build.
A note is just MDX; one tag drops a live, draggable experiment into the sentence. Pull the outlier and watch MSE and MAE pull apart — inside your own prose.
Open the same folder in VS Code, Obsidian, or a terminal — edits show up in the studio, and the studio’s edits are ordinary diffs you can put under git.
Jot a word in a note, ask the studio to fill it in, and the dataset, the dashboard tiles, and the report all update — no code touched.
A studio that remembers.
Every note links back. Concepts grow a spine across your workspace. Ask a question and the whole neighborhood answers.
Backlinks and the Links rail keep every note aware of who mentions it.
Ideas that recur get harvested into one place — the studio notices before you do.
Ask about attention and the kv-cache note answers too. Memory is retrieval, not a graph to tend.
Where the studio is headed.
SHIPPING NOW
macOS desktop app · Notes editor · Canvas · Dashboards · Reports · Nimbus, resident agent · Web clipper · MCPs & connectors
IN PROGRESS
Windows support · Signed releases · Notes review
UP NEXT
Notes sync · Note history & diffs · Web app · CLI
From first light to last thought.
Free while in beta. Your files, your keys, your studio.