DeepSeek-R1 is here!

[agrimes]

Hey, the new reflexive/thinking LMs are here! I'm using a DeepSeek-R1+ Qwen model.

The problem is that it needs to do "out of band" thinking. This creates a whole new set of problems for this game because it needs to sort and store the AI's internal state of mind regarding a thing along with the explicit text of that thing. When prompting the AI to update a person/place/thing/quest then it will now need to be fed back its own internal state of mind regarding the relevant things. This was released just in the past day or two so everything is very new, maybe there's a guide as to how to program this new system. Here's some examples from how the system deals with a basic Bkerkely scenario:

The Ai seems to be using XML style <think> </think> tags to mark its internal thoughts, It is not clear whether the system benefits from receiving these internal prompts back for subsequent processing. A minimal immediate solution is to filter out <think> blocks so that the game can function, and then later improve performance by actively maintaining the AI's internal monologue.

[agrimes]

Here's the precise link: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B

[agrimes]

Here is a video explaining the model I mentioned in the previous comment:
https://www.youtube.com/watch?v=LYxQbgAUzsQ