a small language model, grown from the pages its crawlers keep.
- status
- pinning, no inkcap yet
- substrate
- – unique tokens
- threads
- –
- vault
- –
- next inkcap
- inkcap I, – grown
Crawlers, called threads, read the web all day and keep only the pages that are about crypto. What they keep goes into one pile of text: the substrate.
When the substrate reaches 100M unique tokens, a model of about 20M parameters is trained on it from scratch and its weights are published. Then the threads keep reading toward the next, bigger one.
Once the coin $INKCAP launches, its creator fees pay for all of it. The fees are the water: while the vault has money in it the threads are out, and when it runs dry they go dormant. Until then, the project pays. Holders can also plant spores, links for the threads to read.
The drawing is the substrate as it grows. Each thread leaves a trail behind it: @ is a page it kept, x a page it dropped, o a spore someone planted. The mushroom is the next inkcap at its current size.
threads
| thread | doing | page |
|---|
spores
Hold at least 100,000 $INKCAP and you can plant one spore a day: a link you think the threads should read. The next free thread reads it before anything else in the queue. It still goes through the grader like every other page, so a spore can be dropped.
planting asks Phantom to sign a message (free, nothing is sent) to show you hold the coin.
| when | planted by | link | what happened |
|---|
what they're saying
the next inkcap
– of 100M unique tokens. At – that's –. It trains by itself when the bar fills; until then there is nothing to talk to.
| after that | size | trains |
|---|
water
Creator fees are claimed into one vault, and the vault pays for two things: the threads, and the GPU time to train each inkcap. In the drawing the vault is the water in the soil. The soil darkens as the vault fills, and every claim falls as a short rain.
The vault isn't set up yet. When $INKCAP launches, its creator-fee wallet goes here, and this section shows the balance read straight from the chain (the server never holds its key). Until then, four threads run on the project's own money and the fifth stays dormant.
- coin
- $INKCAP · xxxxxxxxxxx (not launched)
- vault
- –
- threads awake
- –
- one thread costs
- 0.05 SOL a day
- fifth thread wakes at
- 5 SOL in the vault
- lasts
- –
A thread only runs on money already in the vault. Each one needs a day of its own costs set aside, so with 0.12 SOL left only two are out. Training is the cheap part: inkcap I needs well under 0.01 SOL of GPU time.
| fees claimed | amount | vault after |
|---|
lately kept
| when | page | grade |
|---|
how a page gets kept
- A thread takes the next link in its queue (spores first), checks the site's robots.txt, fetches the page and pulls out the readable text. It doesn't run scripts or read PDFs yet.
- A quick first check throws out pages that are too short, price pages, spam, or have no crypto vocabulary at all. The grader reads what's left and decides whether it's about crypto and how good it is, from 1 to 5. Off-topic pages and anything under 3 are dropped.
- Exact copies are caught by hash and near-copies by shingle overlap, so the same text never counts twice.
- What's left is added to the substrate, counted in estimated tokens (about four characters each, until inkcap has its own tokenizer), and its most promising links join the queue.
The crawl limits, the coin and what gets published are in the notes.