← All work
Project / AI / Games

Jev plays Tetris

A hybrid Tetris experiment: code enumerates placements, Jev selects moves, and recorded replays show the results.

Open the original ↗Original Tetris replay showing the full board, recorded results, and playback controls

Code computes, the model chooses

The engine enumerates legal Tetris placements and applies the selected move. Jev receives either the board or richer information about candidate placements; rich mode uses a shortlist ranked by a code heuristic.

The recorded runs let you inspect moves alongside a code baseline and see the model’s recorded decision latency. They demonstrate a hybrid decision loop, not unaided spatial reasoning or a general game-playing benchmark.

Watch the Tetris recordings · Explore the shared source repository

Playback uses saved runs and makes no model requests. Offline baselines and optional local API experiments are included in the repository.

Related work.