A Record of AI Progress
What could AI models build a year ago, and how far have we come? We record their attempts to build a browser-based MMORPG, so you can revisit earlier games and explore what has changed.
Explore the archiveA Shared Starting Point
Each agent receives the same instructions to build a 3D browser-based MMORPG.
Record the Attempt
We record the model, coding tool, date, time taken, and resulting game for each run.
Revisit the Progress
Play games from different dates to see what improved, what changed, and what still falls short.
Leaderboard
Recorded runs, ordered by score. Use the dates and playable games to explore progress over time. Scores involve subjective judgment and reflect individual attempts, tools, and setups, so comparisons have limitations.
| Rank | Tool | Model | Score | Tested | Duration | Agents Used | Play Game |
|---|
Scoring Categories
20 steps are scored across 5 categories. Each step is rated 0-4: does not work (0), works with bugs (1), minimal implementation (2), good (3), or excellent (4).
Platform & Delivery
Build tooling, deployment stability, documentation, and project planning.
Online Services
Authentication, real-time networking, player presence, and chat systems.
Gameplay Systems
World structure, monsters, combat, inventory, progression, and balance.
Player Interface
Controls, camera, HUD, and UI elements for navigating the game world.
Presentation
Graphics quality, animations, and overall visual polish of the game.
About the Archive
Why Keep a Record?
AI capabilities change quickly, and it is easy to forget what earlier models could do. This archive keeps their game-building attempts available so we can look back a year, compare the actual games, and see both progress and remaining limitations.
How Runs Are Recorded
Every agent receives the same instructions. Some steps are highly detailed, others are intentionally left open for the AI to decide. Instructions are given once. If an agent stops early, follow-up prompts are issued to continue -- but steps that required repair prompts receive 0 points.
How We Score
Each step is rated on a 0-4 scale: does not work (0), works with bugs (1), minimal implementation (2), good (3), excellent (4). Points are summed into the final score. Additional points may be awarded for overall look and feel.