The smartest NPC in the world

We can add a new immersion to VR training by using LLMs to power characters and avatars that interact with the user in real-time.

Traditionally, NPCs in games and VR training applications have followed pre-scripted dialogues and behaviours. This approach has several limitations, including a lack of variability in responses and a limited capacity for understanding user inputs. These factors can make interactions with NPCs feel repetitive and artificial, detracting from the realism and immersion that VR training aims to provide.

Creating Dynamic NPCs with GPT-3.5 and NLP APIs

By combining the power of a LLM with voice processing, developers can overcome the limitations of traditional NPC development. User voice inputs can be used as part of a prompt for enabling NPCs to understand and respond user’s voice. The model can then generate contextually appropriate responses although latency is still a bit of an issue.

The ability to create dynamic NPCs with unique behaviours has several benefits for training applications:

  1. Enhanced realism: As NPCs become more human-like in their interactions, trainees are more likely to feel immersed in the virtual environment, leading to a better overall training experience.
  2. Improved replay value: Trainees must now adapt to dynamic characters.
  3. Increased engagement: Dynamic NPCs can encourage trainees to explore and interact with the virtual environment more deeply, leading to a greater level of engagement and improved learning outcomes.

Applications in VR Training

The potential applications of realistic NPCs powered by LLMs like GPT-3.5 are vast, spanning various industries and use cases. For example:

  • Healthcare: Medical professionals can practice their communication skills with virtual patients, who can present a range of symptoms, medical histories, and emotional states, allowing for more comprehensive training.
  • Law enforcement / Security Officers: Police officers can engage in realistic simulations of high-stress situations, such as hostage negotiations, where the outcome depends on effective communication and decision-making.
  • Customer service: Trainees can interact with virtual customers to practice handling inquiries, complaints, and other scenarios, honing their skills in a low-stakes environment before applying them in real-world situations.


1. Computational resources: Large language models  require significant computational power for real-time processing. Currently the only way to do this practically is to use cloud compute. This can introduce latency and requires a fast internet connection (obviously).

2. Context awareness: . The NPC should be able to remember past interactions, understand the user's goals, and provide relevant responses to ensure a believable experience.

3. Integration with game mechanics: Ensuring that the NPC's responses align with the rules and goals of the VR experience can be difficult. The language model needs to be well-integrated with the game mechanics, environment and also their avatar.

Want to start a project together? Get in touch

background small cog
slightly bigger background cog