The area I think is most exciting (and in need of more innovation) is using natural language to create (and modify!) the actual simulation / rules / behaviors. Our approach was to map language outputs to actions that could be chained together using Goal Oriented Action Planning plus an Entity Component System. The user's verbs / prepositions / etc. would add layers of goals, each of which would enable or disable certain behavior components when triggered.
This is very cool! And I think it's a perfectly good approach. My understanding is that the usual way of doing things with LLMs is to train them on a series of specialized tokens that represent actions in your environment. E.g., this sequence of words results in this completion of action tokens.
The code approach is intriguing and I'd like to explore further but controllability is a real problem and bullet-proofing it would require a lot of effort, if it's even possible at all. I do think that a hybrid environment where one speaks what they want and then sees the code and can interact with it in a friendly way would be very intriguing for a sandbox experience.
We got pretty far with this a few years ago using more basic ML/NLP. The app was called Moatboat: https://twitter.com/moatboat/status/1082425681210859520
The area I think is most exciting (and in need of more innovation) is using natural language to create (and modify!) the actual simulation / rules / behaviors. Our approach was to map language outputs to actions that could be chained together using Goal Oriented Action Planning plus an Entity Component System. The user's verbs / prepositions / etc. would add layers of goals, each of which would enable or disable certain behavior components when triggered.
More details here for anyone interested: https://medium.com/@mikejohnstn/whatever-you-say-happens-2fa...
Directly generating source code from natural language would be a fun alternative approach to try today.