Make papers playable.
Upload an ML/AI research paper — it becomes a playable, faithful game you learn by playing, not reading. Real computation under the hood, not a narrated animation.
▸ Featured
Race the Paper
Tune precision, sparsity, and cache against the paper's ghost car. You can't beat it on speed and quality at once — that wall is the Pareto frontier, and their trick is what reaches it. Speed from real FLOPs, quality from a measured grid.
Pipeline Defense
Route queries through a retrieval pipeline where bad retrieval becomes a visible breach. Naive knobs alone let wrong context through; the paper's corrective module (CRAG) catches the specific failure. Replayed from a precomputed grid — zero model calls at play time.
▸ Or bring your own
Drop in your own paper and watch it auto-detect the right game — summary, concept map, and a challenge built from the paper's core idea.
▸ More games
The Leash
Fine-tuning · RLHF / DPO / alignment
Tune how far a fine-tuned policy may drift — too little and nothing changes, too much and it reward-hacks into gibberish.
Play →AI Conference Scheduler
Reasoning · chain-of-thought / test-time compute
A single sample fails the 8-constraint puzzle; best-of-N + verifier + voting under a compute budget finds the solution.
Play →Vision Detective
Multimodal · VQA / OCR / grounding
Interrogate a vision-language model, then ground its answers in real visual evidence before it hallucinates.
Play →Maze Runner
Pathfinding · search / planning
Naive BFS burns its step budget; the paper's learned A* heuristic reaches the goal with a fraction of the expansions.
Play →Spotlight
Attention · transformers / NLP
Find the token the model should focus on — uniform attention smears across everything, learned attention spotlights it.
Play →