Because it’s hard…
We’ve dedicated the last four years to this endeavor.
Making a simple proof of concept (PoC) is easy (we’ve seen many), but polishing it to production level requires 1,000 times more effort where deeper complexities emerge:
1) Cost: Using LLM via API means someone has to pay for those thousands of tokens per hour per player. Yes, you can run a smaller LLM on a player’s GPU (e.g., Mistral, Llama3 8b), but the quality and capabilities may not be sufficient for an engaging narrative and gameplay.
2) It’s too early: GPT-3.5 is only 1.5 years old, and GPT-4 is only one year old.
3) Current LLMs: They may not be capable of running multi-agent narratives for long game sessions. Context sizes are not large enough to hold scene history consistently. LLMs only attend perfectly to some instructions in the context and often hallucinate.
4) Multi-modal LLMs (with Vision): You may need these to unlock spatial reasoning and better grounding to the game environment.
For LLM-driven NPCs to fundamentally transform gameplay—making them indispensable, like a painkiller rather than a mere vitamin—they must do more than enhance; they must be crucial, realizing games that, without LLM-driven AI NPCs, would be unachievable (Beat Saber wouldn’t be what it is without VR).
- NPCs need to interact with one another and their environment.
- NPCs must consistently understand their surroundings, avoiding any contradictions related to their past or actions that the game mechanics do not support.
- NPCs’ behavior and the plot must be consistent over long periods.
- NPCs must have intriguing personalities and narratives that unfold with the drama and tension of a well-crafted movie, rather than resembling dull everyday interactions.
- These narratives should continuously evolve, navigating from problem-solving to new challenges repeatedly.
The benefits are clear: pioneers will create a new market niche and a defendable, tough-to-duplicate technology.
