Question 1

How does the model know where to paint without seeing the image?

Accepted Answer

It is sent a text description rather than pixels: the stroke's shape and where it ends, the emoji around it identified by name, and a small text picture of the cells nearby. Nothing about who painted it leaves the page. From that, Jev chooses an emoji sharing keywords with the surrounding paint plus a position named against the stroke, such as above the line, at its end or inside the loop, so placement is a typed option rather than a coordinate it invents.

Question 2

Is the output deterministic, or is the randomness a gimmick?

Accepted Answer

The README describes the pick as drawn from the model's probabilities, wider the less sure it is, which makes the spread a readout of confidence rather than decoration. It also says the top candidates flash on the canvas before the pick lands, so you can see the distribution being sampled from. The second question is a yes or no on whether the stroke is unfinished: when it is sure, it closes the loop, otherwise it carries the line on instead.

Question 3

What does it cost to run my own copy?

Accepted Answer

Per-visitor cost is capped rather than measured: each visitor has a rate limit and the site carries a daily spend cap, and a visitor is given 300 cells of ink with five returning every second. Calls route through Vercel's AI Gateway first with a direct TypeSafe fallback, and the `packages/jev` client checks answers and reports cost. To run the code itself, `npm install`, then `npm test` and `npm run typecheck`, where the unit tests make no Jev calls and cost nothing.

Question 4

Why can't my pull request be merged?

Accepted Answer

Because this repository is a mirror, copied from the site's monorepo on every change made there, and the README states that pull requests are read but applied upstream. Issues and patches are still worth filing, but the fix lands in the site repository rather than here. The code splits into two MIT packages: `packages/emoji` for the canvas rules and palette, and `packages/jev` for the client that checks answers and reports cost.

Question 5

How do I use the live canvas?

Accepted Answer

Open chriswijnia.com/lab/emoji and start drawing; there is no account. Pick an emoji and drag, and each visitor gets 300 cells of ink with five coming back every second. Everyone's strokes arrive live, so it is a shared surface rather than a private sandbox.

Question 6

How does an emoji get picked, and how do positions work?

Accepted Answer

From a typed set rather than free text: emoji that share keywords with what is painted around the stroke, each in a place named against that stroke, such as above the line, at its end, or inside the loop. The palette in packages/emoji carries the names, groups and keywords that matching runs against, and the top candidates flash on the canvas before the pick lands.

Question 7

Can I test any of this without spending money?

Accepted Answer

Yes. npm install, then npm test and npm run typecheck run the packages' unit tests, which make no Jev calls and cost nothing. packages/emoji holds the canvas rules and palette as ordinary code you can assert on offline, and packages/jev is the only part that reaches a model.

1 Million Emojis

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community