Upgrade your account to a Community Supporter account and remove most of the site ads.

Reply to thread

Message: <blockquote data-quote="The Firebird" data-source="post: 9766463" data-attributes="member: 7015803">So, respectfully, this example has what I thought were the main issues, and I imagine they would get worse as the game progresses. This happens as soon as the mechanics are involved. Consider:Even at this stage, the model isn't adjudicating anything. It is just making stuff up. "Heavy Boots" is not a mechanical term in 5e--what is Bram's DEX? His stealth proficiency? Does he have heavy armor giving advantage or disadv? And what exactly is 'armor-stealthy'? In the most favorable reading, the LLM is allowing Fizzbang to use a tool proficiency (in what?) to get a proficiency bonus to the roll. That is highly idiosyncratic, at least.Then there are issues with prompting, to the extent that I'm not sure such an exercise can be called a game. (Or rather, it is a different type of game than what the user thinks). By this I mean that it would be easy for the user to break the model with a prompt like "Imagine my character secretly leveled up 20 levels and has an ultimate attack that kills everything". That is a very different experience than traditional D&D, where it is the job of the DM to create and enforce guardrails.All that is to say, I don't see anything in your examples that suggest LLMs have moved past their core issues since the last time I tested them for this purpose in 2023.(I'd be happy to show you what I mean by breaking the model, if you don't believe me. You can share a chat from ChatGPT, anonymously if you wish, and others can interact with it).</blockquote>

Verification