Game programmer, designer, director; retired quadball player; antimeme; radical descriptivist; antilabel; Moose;

Working at Muse Games. Directed Embr, worked on Wildmender and Guns of Icarus, Making new secret stuffs

Opinions are everyone else's


gaw
@gaw

I just finished Firmament, the new game from Cyan, the creators of Myst. It's... just fine! It scratches that Myst-like itch; the realms are gorgeous, it has a lot of puzzles that are grounded in physicality and mostly involve powering or fueling things and turning them on. I was disappointed that there's not really a big turn to the mechanics; you do a thing in each of three realms, collecting three upgrades for your tool along the way, then you do a harder thing in each of the three realms, then you finish the game. Many of Cyan's other games add a fun complication in the second half; I'm thinking of the lovely revelation of how Obduction's worlds are connected. The story in Firmament is fine, but mostly boils down to introducing a mystery in the first hour, letting you solve puzzles for five hours, and then explaining the (not hugely novel) mystery in the last hour. Not Cyan's best work, but not their worst.

The VR was gorgeous but the controls were almost unusable and my rig chugged way too much in complex scenes for it to be playable. That's okay; I'm happy to play in pancake mode.

There's one big issue I have, though: they seem to have made extensive use of AI content generation.


There's a bit in the credits that goes by so quickly that I couldn't read it all, and I read quickly. It's almost like they're a bit embarrassed about it:

A.I. Assisted Content

Journals, logs, checklists, newspapers, stories, songs, poems, letters, loosely scattered papers; all backer portraits; all founders portraits; the "sunset" paintings; the art-nouveau wallpaper in the Swan dormitory hallways; propaganda banners; coastal spill decal kit; all voiced mentor, announcer, founder, and other speeches; backer-exclusive content

This isn't okay. I hope I don't have to go into how virtually all AI-generated art is plagiarism. In limited cases, I think it's fine to use AI-generated stuff; for example, the StyleGAN system used by This Person Does Not Exist uses a dataset of permissively-licensed images from Flickr. The photographers and models gave permission for their work to be used, and generating a bunch of images of people for something like the portraits mentioned in the credits seems fine to me. Far cheaper than getting model releases, and for that sort of generic work I'm not too hung up on the fact that humans were deprived of their sliver of royalties from Getty Images.

Edit: Kotaku wrote an article and Cyan released a statement in response. In short: Cyan says that while they used AI tools, it didn't count as AI generation, and human creators at Cyan were involved in everything. I'm not happy about this, but it's not as bad as I worried it was based on the vagueness of the credit.

Click to see my original speculation that was repudiated by Cyan: The voicing is a much bigger issue. I remember thinking throughout the game that the "mentor" voice, who guides you and delivers most of the plot, was a bit flat, but I chalked it up to a cheap voice actor (perhaps a developer) for whom English was not a first language. That would be fine! I've played plenty a game, including many from Cyan, where the acting was just passable, and this was a Kickstarter project. But using AI to generate the voice? Shameful. Just pay someone to record it! Trust me; we did it at @FPG, and we have an extremely shoestring budget. It's not that expensive, and it'll be as good or better than what they ended up with.

It's also shameful that they AI-generated a lot of the writing. To be clear, I'm pretty sure that the good-enough general story was human-authored. But it now makes sense to me how much I found myself skimming documents rather than reading them in depth like I have journals in Myst or Uru. None of these documents matter; they can't, because Cyan couldn't depend on them to make sense or say what they wanted them to! And writing is cheap. I know! I'm a writer! I could have written better flavor text in a few days. They have writers on staff! Good ones! Rand Miller is their CEO, and he wrote Myst!

Based on the development timeline, I'd guess that Cyan grew nervous about delays and their increasingly-insistent backers, some of whom were already a bit grumpy at some of the design changes from the initial proposal. At one point, Cyan projected the game to be released in late 2022; but in November, they switched to Q1 2023. AI art generation became super trendy in mid-2022. It sure seems that, low on time and budget, Cyan turned to AI to avoid paying the time and money for real artists, writers, and voice actors.

I hope Cyan stays in business. I like their stuff. But I'm very, very disappointed in their decisions regarding the use of AI in Firmament.


You must log in to comment.

in reply to @gaw's post:

The Miller brother's have been posting various AI-generated stuff of twitter for a while now. So to some extent they appear to be true believers and I doubt they'd be particularly ashamed of it.

Firmament has always seemed like a weird non-priority for Cyan. It was half-positioned as a VR-first game, but then walked back when that raised expectations for VR users. And the game was seldom mentioned across the years of development in preference to plugging more Myst remakes. I'm fine with them prioritising whatever they want, but it does also suggest it was never meant to be "The next big game by Cyan!"

The original projected fulfilment date for the Kickstarter was mid-2020. Those are usually complete fiction but the Kickstarter itself was run in 2019, so you'd hope they'd have had at least some idea of what it would be and what was required by then.

Maybe to some degree there was some pre-AI fad experimentation with Neural Nets/PCG that inadvertently got caught up in later culture wars. But to me it suggests the game was always meant to be a smaller scope/"cheaper" side project, and they saw an opportunity to paper over the cracks with various AI tools.

I mostly backed the Kickstarter mostly because Obduction turned out fantastically. But haven't actually played more than 10 minutes after discovering I was affected by a critical error (narrator audio not playing after the first monologue) in the launch version.

I was perfectly happy to chalk up the release delays to COVID. The wild thing is that the 3D environments are gorgeous and amazing and clearly handmade, but they seem to have used AI for some of the easier parts (and also for 2D art, which is not easy but which they still shouldn't use AI for).

Yeah this is a part I don't quite feel happy about. The group seemed deep into the AI prior to it taking off in the public's eye. It's unclear to me what was used, how, etc. I find AI at best a child's trinket in how it is being used and there are serious legal problems to using it.

To some extent, a lot of the worlds feel much different than it's closest comparator in a 3-dimensional space, Obduction. That is not to say the realms of Firmament are entirely bland, but the brush strokes of say, the various ice fields, seem devoid at times of human-intervention and feel lifeless. It would not surprise me if much larger portions of the game than we could guess were AI-generated.

Given the constraints/changes of the release date, and what one can assume having played through, that a portion of such things as the posters and newspaper clipping were AI-generated. I didn't see a credit to any acting or voice-artists but I could be wrong.