About

Gamelab teaches models how to REALLY think

We are a dedicated division within Arkadium, built by a multidisciplinary team of game developers, designers, software engineers, data scientists, and AI/ML researchers focused on advancing how models learn and improve. With decades of combined experience across interactive entertainment, large-scale data systems, machine learning, and RL, the team brings a unique perspective: using games as structured environments to study reasoning, strategy, and behavior in ways that are both measurable and scalable. Operating within Arkadium's long-standing, profitable, and stable business provides GameLab with a rare advantage: continuous access to real-world human gameplay data, production-grade infrastructure, and a live ecosystem of millions of players. This foundation allows the team to move beyond theory and into practical, real-world experimentation, building systems for training, evaluation, and benchmarking that reflect how decisions are actually made under uncertainty, over time, and at scale.

Our Games Measure a Variety of Cognitive Demands

Visual Processing
Physical Reasoning
Social Reasoning
World Model Learning
Memory
Spatial Temporal
Long Horizon Planning
Strategic Decision Making

Gamelab Has Scale

1.7BSpatial temporal data points
150BHuman decisions per year
100+Owned games across 20+ genres
200BGameplay images per year
22MMonthly players
50+Skill archetypes (experts, casual, new players, etc.)

The Largest Wholly Owned Active Games Portfolio

Perfect Information

Perfect information games: Go, checkers, and chess

13.5M Monthly Game Plays

Imperfect Information

Imperfect information games: Hearts, bridge, and gin rummy

10.5M Monthly Game Plays

Stochastic

Stochastic games: backgammon, Klondike solitaire, and blackjack

26.5M Monthly Game Plays

Single Agent

Single agent games: Block Champ, Spider Solitaire, and The Daily Medium Crossword

50M Monthly Game Plays

Physics & World Models

Physics and world model games: Stockcar Hero, Ballistic, and Mr Racer

20M Monthly Game Plays

Deterministic

Deterministic games: 9 Ball Billiards, Sudoku, and checkers

35.5M Monthly Game Plays

Visual Dominant

Visually dominant games: Bubble Shooter, Mahjongg Dimensions, and Tile Match Flowers

12.5M Monthly Game Plays

Long Horizon Planning

Long horizon planning games: bridge, spades, and Hex Jigsaw Challenge

16.5M Monthly Game Plays

Constraint Based

Constraint-based games: nonogram, Minesweeper, and Word Wipe

11.5M Monthly Game Plays

The Most Structured Dynamic Gameplay Data

Static Data
A single moment in time, no progression
No explicit actions; just pixels or words
No consequences or rewards
Surface-level patterns
Overabundant, easy to scrape, low uniqueness
Our Data Differentiator
Sequences
Actions
Outcomes
Strategy
Exclusive
Dynamic Gameplay Data
Multi-step sequences of decisions
Every move is an action chosen by a human
Actions lead to wins/losses, score changes, success/failure
Human planning, foresight, decision making with imperfect information
Rare, structured, exclusive human decision data

Frontier Model Partners

Google DeepMind logo
Meta AI logo

Special Thanks

GoodStartLabs logo

CONTACT US

Do you want to know more about the project?