Christopher Kalos

Technology, Hobbies, and New Ideas

An avid techie since childhood, I've finally decided that it's time to establish my own presence on the web, as opposed to outsourcing it to a bunch of social network pages.

Here you'll find musings, my professional history, and anything that catches my interest.

As will become obvious as this page grows, I'm an Apple user, through and through:  When I work, I use the best tools for the job.  Sometimes that's Microsoft-based, sometimes it's Apple-based, and often, it's Linux-based.  Once I'm at home, though, I find that comfort, simplicity, and ease of use trump all of the flexibility in the world.   

Cool counts for a lot.  Simplicity counts for a lot more.

 

How To Run A Tabletop Roleplaying Game Like A Sales Engineer: On AI and Large Language Models

August 05, 2024  /  Christopher Kalos

Let me open with this: LLMs are a dead end. That doesn’t make them bad, but it does mean that one should bear in mind that they’re limited. You just need to know how to use them. One use case that came up a lot was using them to generate correspondence for customers. On the surface, it makes sense: For some people, wordsmithing takes more time than they have. It’s like a muscle that requires regular training, and sometimes, you just need to get the job done. They are, after all, large language models. They’re trained on more works than most people will ever read, so they’re going to be very good at working with that information.

While I’m not that sort of LLM user, I have found a pretty solid use case, specifically with Google’s NotebookLM. See, I own a lot of rulebooks in PDF format. I’ve paid for them, they’re mine, and that’s great, but even though I’ve read them all, studied the rules and suggestions they provide, my recall isn’t perfect. It’s good, but it’s not perfect.

Enter the world of a private LLM. Instead of relying on a bunch of bookmarks and transcribing notes into a format that works for me, why not upload everything to my own, custom model? Instead of going to a public LLM, why not make one that uses everything good, and discards the worst problem in them: those cases where you get inaccurate data back?

So, that’s where I’ve gone. The Game Master’s Assistant. If I don’t know every rule, I ask. If I need to check some obscure ability in a sea of books, I can just throw it at the AI, and get an answer back. And the best part? I’ve asked it unanswerable questions. I’ve gotten the terminology completely wrong, and it tells me the same thing I’ll tell a customer when I don’t have a good answer:

”I don’t know.”

When I do that, I’m not admitting weakness. It’s a call to action to find the answer. When the AI does it? It’s not going to make up an answer, either. It may not be able to tell me that answer, but it definitely saves me time. I can ask, I can usually get this little robot brain sitting up on GCP to get me some answers, and most importantly, it means that we can get back to having fun.

I don’t fear AI. I’m not worried about LLMs replacing me. But it helps to remember the dream, as quoted by Joanna Maciejewska:

”I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do my laundry and dishes.”

Comment