Watch them decide

Two decision models, the same game, the same situations, no training. Both play at once, side by side, and every option each one weighed is on the page.

Playback

Change lanes and speed through dense traffic without crashing. Scenario #0 gives both models the same traffic; each runs its own episode, so they never meet. These are recordings: run the arena yourself to watch live.

Jev

TypeSafe, closed API

Press Play. Every decision this model makes appears here.

Speed
-
Time
-

Laya

Open source, runs on this machine

Press Play. Every decision this model makes appears here.

Speed
-
Time
-

Metres travelled, both models on one scale

Jev · not started
Laya · not started

Jev

Press Play to see the wire.

Laya

Press Play to see the wire.

Earlier decisions

newest first · pick one to pin the whole page to it

#JevLaya

Decisions appear here as they are played.

Watch them play live

The games above are recordings: this site is a static page, so it cannot run a model. The same arena runs on your machine in four steps, and there the models play live, on any game and any seed you like.

  1. 1. Get the code

    Clone the repository. You need Docker, or Python with uv and Node.js 22.

    git clone https://github.com/nadeem4/jev-demo.git
    cd jev-demo
  2. 2. Add a Jev key

    Put an AI Gateway key in .env at the repo root. Laya needs no key: it downloads once and runs on your machine. Without a Jev key you can still watch Laya and the baselines.

    AI_GATEWAY_API_KEY=your_key
  3. 3. Start it

    One command builds and runs the arena server and the site. The first start downloads Laya's model, about 2.3 GB.

    docker compose up --build
  4. 4. Play

    Open localhost:3000, pick a game and two models, choose Live, and press Play. Now the models are deciding on your machine, in real time.

Full setup, the benchmark commands and the Colab notebook are in the README on GitHub.