What it does
Providing possible next boards helped most, but the best Jev setup still landed near a simple fixed policy in this small author-run comparison. The replays show recorded games, not wall-clock inference timing.
Benchmarks & research
Kicking the tires on Jev with 2048
Four ways of asking Jev to play 2048, compared with random moves and fixed rules, with replays of the resulting games.
Recording unavailable. Open the experiment ↗
How you can use it
Sketch out the four directions a player can slide tiles in 2048. You can ask a developer to build an app that creates the future board for each valid slide. Your developer can then ask Jev to pick the best move from those previewed boards and point values.
Running this setup requires an account access key that connects your app to Jev. Keep in mind that in the author's tests, the model scored slightly lower than a simple rule that always tries the same directions in order. The complete test software was also not shared.