AI Won't Take Gamedev
Game development is already commoditized. November 17, 2022 Writing AI Art
A lot of VC thinkfluencer types have been salivating at the possibility of how AI is going to take over gamedev. This won’t happen. Additionally, a lot of the “arts” bandied around as being prime for being taken over by AI (movies, animation, etc.) also won’t be taken over by AI. Here’s why.
Video games, film, and whatever Fortnite “is” are something like “second order arts”. Or as Robert Yang, quoting Wagner put it, a Gesamtkunstwerk. They are by-products of multiple creative disciplines, and are more than the sum of their parts. AI diffusion models (for now) aren’t looking at this space. Instead they are focused around “first order” art, where the work and the product it produces are nearly one in the same. Aka, a writer writes a story, an illustrator illustrates a drawing, etc.
How much AI takes over “first order” art, though it is happening now, I think is also overstated, but not necessarily the purview of this article (I addressed this a bit in a previous post). But for second order art, we’re still incredibly far away.
What tweetfluencers like to do is look at how first order art is being commoditized by AI, and then say: “Well if all these things that go into making games and done by AI, it follows that AI can make games”. Here’s the thing though, that any working professional in games understands: the things that AI is commoditizing are already commoditized. Or if not completely comooditized, are in a race to the bottom.
Midjourney stuff looks scary, but at the same time everything i see is sort of showelare outsourcer style art that is nightmare to work on. Probably already destroying the market that was rotten to the core.— Computer can do art, now lets teach it some pain (@BakalVitaliy) November 9, 2022
There were and are plenty of channels to produce both low bar and high bar “content” for games before AI “happened”. Most AAA outsource the bulk of game content creation to the lowest bidder — AI here is just replacing already marginalized and overworked people here who are competing to not be displaced by the next company with insane work conditions and lower budgets.
And that’s the “hard” content to make — ask any games writer or audio designer how much a game director “cares” about their disciplines (monetarily) and you’ll likely hear that it’s near 0 (compared to investment in “trailer features”). Basically, if you want to pay little to nothing for content for a game, AI isn’t offering anything new here. It’s just a different interface to a comparable output.
But what about code? Surely stuff like Github Copilot is making games more commoditized? Well, yes and know. Copilot, in my experience, works really well with problems shaped like koans. Easily chunkable pieces of functionality that are mostly self-contained. Additionally, it specifically operates on code, which isn’t how most modern games are made (engines like Unity and Unreal often work with meta file content in addition to code in order to have features and serialized values inside an editor context — complex data types instead of simple HTTP responses). Said differently, Copilot is great at helping you write idomatic code, but isn’t great at systems-level thinking for complex games (and obviously has no idea about rendering).
Let’s imagine it does though. Copilot can “write” games. Does this mean tweetfluencers win? Again, probably not. Games suck to make, because everything in the game is a knob that can be tweaked and tuned. Many games would look very similar from Copilot’s point of view (once again thinking of Ralph Koster’s idea of “Game Topologies”), but would vary fierecly in quality. For AIs to takeover gamedev, this would have to not be true, but it seems so far from possible I’m going to go ahead and say that it won’t happen.
The reason for this is that “good” games are incredibly hard to quantify in a way that can be passed to an AI for training. You can’t just feed an AI screenshots of prestige videogames for it to be able to generate “good” games. They parts of a game that manifest in ways that are legibile to AIs are not the parts of games that make games good or not. Not only this, but even a stream of screenshots from something like Grand Theft Auto wouldn’t likely even be able to determine what a player is doing/why. Or what it would do is take an open world game and quantify it as a predetermined route, like a projection of the possibility space into a “baked” route.
What’s hard about making games isn’t the stuff an AI can gobble up. It’s not even code. It’s tweaking all the million parameters of the game, making the game legible to players (both in UI, systems design, UX, etc.), and getting people to find out about the games themselves. An AI could probably generate the stuff that someone could use to set off on this journey, but that work of stiching things together into a cogent experience, whether in games, film, etc. is still a far way off.
DateNovember 17, 2022
Up nextHeisenfunctions, Incremental Determinism, and The Future of Programming Exploring what it could mean to program with GPT from the very start
PreviouslyTalking To Machines, Or "Serendipitous Prompting for AI Image Generators" Talking about how Midjourney has changed how I approach creative work.
subscribe to my newsletter to get posts delivered directly to your inbox