Why No AI Games?

This has been discussed ad nauseum in other places here over the years, but basically I don't think the conclusions in the post are correct.

It does not make a whole lot of sense to let an AI determine story, whether it be dialogue, changing the game state, or whatever. The reason basically being a few things - one, the fundamental aspect of a game is that there are rules and boundaries these rules need to remain consistent (and testable). It is entirely jarring if an AI NPC that says something that's not consistent with the game state, or changes the game state in a way that violates the constraint of the understood rules/boundaries of the game - this isn't fun to the user at all, even if it sounds cool. If you do not believe me, you can test this out trying to play DND with these things and see how annoying it can become. If you decide ok, I'm going to try to tightly bind what the AI can do or not with branching and rules, you are basically engineering the same way that games have engineered already, so why even use AI? It's a solution looking for a problem here.

The second big issue is determinism - LLM's are fundamentally non-deterministic. In most games, you would expect an action to have the same or at least very similar reaction most of the time - RNG comes into play naturally already in a lot of games in a predictable way (dice rolls, chance to hit, etc.) LLM's bring nothing new to the table here.

For world gen, we already have games like No Man's Sky deterministically generate quadrillions of worlds. What help is an LLM here? We already have the technology.

One area that would be interesting, is agentic bot players, but that sort of leads down the same path as the above arguments - there already are extremely sophisticated bots that play a huge variety of games. What do LLM's bring here?

It's not only just signalling. I've never said anything about AI art in games before this moment, but if I become aware that a game has used it, I'm not getting that game.

Signalling means you are doing the thing only to identify yourself to others as being in the tribe. That's kind of the opposite of my personal take.

I tried to write a different "convince an AI" game about a year ago, but it was hard to work with, hard to figure out a business-model for, and more importantly-- it just wasn't very fun to play. Maybe there's a different scenario than the one I chose.

I've spent way too many years chipping away at genai gaming, here's my 2c.

First, the intersection between someone who understands and cares about genai (and keeping up with the SOTA), with someone who wants to put in the work and compromise to make great games, is slim.

Think about the intersection between a procedural generation/shader programmer, someone focused on making a really fun and compelling game tech be damned. The skills don't really overlap. One requires deep knowledge, technical mastery, geometric thinking, algorithmic optimization, and time in a dark room, and the other requires empathy for humans, brutal iteration, artistic taste, and a knack for storytelling (even if just to sell the player on your game). You basically need both for your "ai game" to even qualify as an ai game. And in an industry that's so anti-ai it's hard to find teams where these two personalities meet and vibe enough to work on a game together.

Secondly, is model performance: quality, speed, cost. I don't think the industry has yet crossed the threshold where generation is high enough quality, fast enough for players, and cheap enough that it amortizes compared to traditional gamedev. You see tons of AI used in development but deploying this stuff to the front lines doesn't really napkin math. That will change though.

Third, game engines are NOT designed for genai. They are decades of optimization for a completely different pipeline, both technical and human: pitch -> publisher -> game engine -> art team + coders -> trailer -> marketing -> release -> support. Adding even simple very, very basic genai requirements to this (say, dynamic texture generation and loading over the network, streaming character dialogue) completely breaks decades of assumptions of how games are built and shipped (e.g. everything mega-super-duper-compressed and loaded to statically allocated gpu buffers). So doing anything new means throwing away large parts of the game engine. And if you pitch not using 70% of the game engine, or making your own, to a 20 year game veteran/studio boss calling the shots in this industry: well, your game will not be getting funded.

This will all get fixed though. The technology jsut came faster than the industry could adapt.

Looking forward to GDC this year!