Mobile Musings: Video Game NPCs and Data Science Proxies

I’m trying something a little different today - I’m writing from my mobile because I’m not in front of my computer. My texting prowess is better described as prow-less, so expect some creative grammar and spelling.

After writing about Inworld.AI last week, I spent a little (translation: a lot) time thinking about NPCs and what it would actually be like to have them be more multidimensional. I started asking myself do I even want that as a player? Imagine a nightmare scenario where I say something that pisses off the NPC and then he would refuse to give me the thing I need to progress. That nonsense happens all the time in real life, why do I want that in a game?

The truth is, I don’t. NPCs are one dimensional character for a reason - they are a proxy for a real person, without all the messiness that comes with being a person. It’s beautifully simple. It’s great, the NPC serves its purpose and then I move on. As a stand-in for human interaction, it’s kinda great because people are a lot.

Naturally my internal monologue decided to up the ante and drew a comparison to proxies in data science. In data science, a proxy is a stand in for something the data scientist can’t directly measure. For example, I can’t determine someone’s income directly, but I can look at where they shop to guess about their household income. It’s a simplified, one dimensional representation of something that’s complex.

But here's where it gets interesting: data scientists rarely use just one data point to build a proxy. They're looking at multiple factors to triangulate what they want to know. It's not just where someone shops - it's where they live, what car they drive, whether they have kids, the whole nine yards.

Pretty useful, right? We can infer all kinds of things using proxies. But what happens when you use proxies to create more proxies? Maybe my “do you have kids” data point is actually a proxy made up of a whole bunch of other data points - and so on and so on.

Eventually, the quality goes off the rails because your proxy becomes so far removed from what it's supposed to represent. It's like playing a game of telephone with data - by the end, you're left with something that barely resembles the original message.

Let’s go back to our friendly NPC. We know they are proxies of proxies of people. But they are so far removed from the “source” (aka us, real flesh-and-blood people), that while we can recognize them as the thing they represent, we’d never mistake it for the real deal. The NPC is a proxy of a person, that is a proxy of a type of person that fits into this virtual world, that is a proxy of a designer’s understanding of a real world group of people that they’ve only ever read about, which is yet another proxy … it’s an endless web.

So our NPCs end up being these fascinating patchwork creatures - recognizable as human-ish, but fundamentally incapable of replacing what they represent.

So, then I have to muse: since GenAI is built on the same principle - an amalgamation of proxies cobbled together to simulate consciousness - does that mean it's also inherently limited? Has it become so far removed from what it's trying to represent that we could never mistake it for the real thing?

Maybe that's not such a bad thing after all.

Did you know that for only a few dollars per month, you can help support my work here. If you have the means, please consider subscribing— and as always share with you friends!