Where should I start creating a model?

I wish to use Sentis and keep thinking of use cases but there is so much to take in … I simply don‘t know where to start and thus far, my expectations for AI seem to be at a mismatch with where tech currently is at (quality, capability, applicability).

Most importantly the key to using AI and Sentis lies entirely with the model you use. Sentis being „only“ the execution layer.

If so I wish to learn how I can train my own simple model.
Something that has the AI pick existing (categorized) assets for me using very simple prompts. Like, let the AI choose ten assets and you can then use those to design a room, or possibly let the AI do that for you eventually.

What would I need to look at to achieve something simple as this? Or is there even something like this available?

Surely I could write this using classic scripting parameters and Random() but hey, it also seems like something I could train an AI to do. Unless I understand the creation and training of custom models I feel I cannot grasp the true power of AI, nor its limits for that matter.

So, I hope to get some pointers for up to date and Sentis compatible resources on creating simple models and that you don‘t mind me asking this here. :slight_smile:

2 Likes

To answer my own question: I’m going to start experimenting with ml-agents.
Sentis unfortunately doesn’t support it yet but it has been said by Unity staff that support is coming. (can’t find the citation again)

After looking into different options I realized that:

  • most AI models are pretty academic … and often totally unapplicable to games or game tech IMO
  • creating models using terms and technology I am not familiar with means I’m looking at an extreme learning curve and lots of potential for frustration
  • ml-agents works within an environment that I’m comfortable with (Unity, with just a bit of command line), provides a really simple API to get started and potentially allows experimentation with all game engine features as well as gameplay elements, and easily have them visualized in realtime 3D

That makes using ml-agents for training models a no-brainer! At least it provides a relatively smooth entry into the AI model training field.

1 Like

Hi you can check out our samples or our models on Hugging Face. It is a good place to start with tinkering with the code.