The quest to create realistic flocking behavior in games has long challenged developers, but AI language models are now demonstrating impressive capabilities in game development. A recent experiment challenged various AI models to create a sheep herding game called Shepherd's Dog in a single attempt, with Claude 3.7 emerging as the clear winner.
The concept behind Shepherd's Dog is simple yet engaging: players control a dog to herd sheep into a pen before nightfall. What makes the game interesting is the realistic flocking behavior of the sheep, which must demonstrate authentic group dynamics while reacting to the dog's presence and navigating around obstacles.
AI Models Show Varying Degrees of Success
Claude 3.7 dominated the competition with a score of 24/28, creating what many users described as a legitimately fun game with natural sheep movement. The Claude version even animated the nightfall transition without being specifically prompted to do so. One player noted that the game felt reminiscent of early iPhone days mobile games in its simplicity and appeal.
Other models showed mixed results. Anthropic's o3-mini scored 16/28, implementing the classic boids algorithm for flocking behavior. While technically impressive, some users noted that this approach didn't feel entirely natural for sheep since the boids algorithm maintains constant movement, whereas real sheep can remain stationary.
At the bottom of the leaderboard, models like Deepseek failed to produce valid JavaScript, while GPT-4o and Gemini Pro scored just 4/28 points each, with limited features and functionality issues.
AI Model Leaderboard for Shepherd's Dog Game
| Model | Score | Notes |
|---|---|---|
| Claude 3.7 | 24/28 | Really impressive demo, missing some obstacle dynamics |
| o3-mini | 16/28 | Misses features, but has good flocking & gameplay |
| o1 Pro | 12/28 | Similar to o3-mini |
| Mistral | 12/28 | Herding not properly implemented |
| GPT-4o | 4/28 | Limited features, model wouldn't continue |
| Gemini Pro | 4/28 | Difficult as sheep don't stay in pen |
| Deepseek | 0/28 | JavaScript is not valid |
| Cursor | 0/28 | Impressive but may not qualify as "one-shot" |
Flocking Algorithms: The Heart of the Challenge
The core challenge in creating this game lies in implementing realistic flocking behavior. Many AI models defaulted to using the boids algorithm, a well-known approach for simulating flocking behavior that was developed in the 1980s. However, as one commenter pointed out, boids may not be ideal for sheep simulation:
Seems o3-mini implements the 'boids' algorithm for flocking (likely due to its prevalence online), but I find that here it doesn't really fit. Indeed in boids each element has a constant (or minimum) velocity, s.t. the sheep never stop 'running'. I find the Claude flocking behaviour looks more natural, for sheep.
This insight highlights how Claude 3.7's implementation managed to create more realistic mammalian behavior by allowing sheep to have zero minimum speed—a subtle but important distinction that made its simulation feel more authentic.
Community Engagement and Improvements
The community response to these AI-generated games has been enthusiastic, with several users sharing their own versions or suggesting improvements. One user created an enhanced version using multiple AI models in combination, demonstrating how human guidance can help refine AI-generated code.
Some users noted playability issues that would benefit from human refinement, such as maps being too large on desktop or sheep getting stuck in corners. These observations highlight that while AI can create functional games in one shot, human play-testing and iteration remain essential for polishing the experience.
Interestingly, several commenters mentioned that they had previously attempted to create similar games themselves, showing that the sheep herding concept has broad appeal among game developers. The challenge of getting flocking behavior just right appears to be a common hurdle that many have faced.
The experiment demonstrates the rapid advancement of AI coding capabilities while also revealing their current limitations. As one-shot creations without feedback or iteration, these games represent impressive technical achievements, but they still benefit from human refinement to reach their full potential.
For those interested in trying these AI-created games, they're available through GitHub HTML previews, though some users reported security warnings from certain browsers—an unexpected hurdle for what are essentially simple HTML/JavaScript games hosted on GitHub Pages.