Building LLM Bots for Gaming

Hey builders!!! I’ve had such a fun month with building Large Language Model (LLM) bots to play, compete and create experiences in video games. In this post, I will share my key insights and takeaways from these experiments, focusing on the importance of model choice, how to build a model persona, and dealing with hallucinations.

Choosing the Right Model

There is a school of thought that you only need one model to accomplish every task. My experiments on the other hand prove them wrong. Every traditional benchmark will have Claude Haiku performing worse than leading frontier models such as GPT-4 or Claude Opus, but as my Street Fighter Experiment shows, the faster leaner model able to return intelligent results fast was the key difference for the task.

Conversely, in games where complex data processing is needed, larger models have their place. Models such as Mistral Large and Claude Opus were much better in Pokémon battles. It came at the cost of speed, with Opus taking seven times longer to select a move than Haiku.

This trade-off between speed and data processing capabilities illustrates the importance of model selection based on the task at hand. Being able to leverage many different models through Amazon Bedrock was a great boon for this experience.

Defining How Models Interact and Think

Creating a persona for your LLM and setting clear constraints and guidelines is vital for directing the model's behavior and output. This not only helps in making the interactions more predictable, but also enhances how the AI integrates with the game's mechanics.

In my work with a Super Mario level maker, increasing the number of examples provided to the model from one to three dramatically improved the quality and playability of the levels generated.

In the Pokémon battles, updating the system prompt and instructing the model to adopt a more aggressive strategy boosted its win rate from 5% to 50% against the heuristic bot. This adjustment not only made the game more competitive, but also led to the model generating entertaining and innovative responses, adding an element of surprise and enjoyment to the gameplay. The ability to tweak the model's approach shows how flexible and adaptable LLMs can be in navigating the “Jagged Frontier”.

Unpredictable Creativity

Despite clever prompting and hacks, there is no way to prevent LLMs from hallucinating. In building my Slay the Spire bot, this led to the program crashing due to unexpected moves. I was able to build guardrails thanks to the help of Amazon Q Developer, but it highlights the need of adopting an "error handling" mindset when building with LLMs as you would with traditional software. Here are some of the hallucinations I experienced when building the bots for each game:

• Super Mario: Models occasionally misplaced game elements, like pipes.
• Slay the Spire: Errors included suboptimal play decisions and incorrect attack uses.
• Pokémon: Hallucinations ranged from misunderstanding type matchups to inaccuracies in damage calculations.
• Street Fighter: Invalid moves, and even refusing to play

These examples highlight as the intelligence required to complete a task, the current generation of LLMs won’t be able to automate everything. This why it's import to act as a human in the loop when leveraging LLMs to assist with tasks.

This is also why I see more mechanisms for enabling LLMs to use specialized tools. For example, a battle calculator in Pokémon, could take inputs and provide results for an LLM to use, ensuring more accurate and strategic gameplay.

Conclusion

My experiments in building generative AI bots for gaming shows us how we can democratize the tools necessary for creators to engage in innovative ways of learning and playing with LLMs.

You don’t need a full research team or extensive programming knowledge. With a well-crafted prompt and a dash of creativity, you can build exciting and enriching experiences.

If you have any other ideas for games to test, or unique experiences to build, let me know in the comments. Until then, keep building!