Microsoft recently showed off the latest version of its generative AI game technology, publishing a version of Quake II playable in a web browser. It quickly attracted a lot of attention, particularly from the game development community, who questioned the benefits of such a technology.
Id Software’s Quake II originally released in 1997, and the influence of its fast-paced first-person shooter action remains strong today. It’s part of why the reaction to Microsoft’s AI experiment has drawn such strong reactions. One critic described the demo as “an insult to life itself”, evoking the famous quote from Studio Ghibli director Hayao Miyazaki oft attributed to his stance on AI animation.
Microsoft first demoed its generative AI gaming technology in February, sharing multiple clips based on Bleeding Edge, a 2020 game made by the Microsoft-owned Ninja Theory studio. The research focused on a technology called “World and Human Action Model (WHAM)”, known by its nickname of “Muse”.
Showcasing generative AI technology that aims to replicate gameplay, Microsoft then shared an updated version of Muse designed to run an interactive version of Quake II. However, since its launch, the generative AI-based shooter demo has faced strong backlash.
“Underwhelmed” was the reaction from Unpacking and Tempopo developer Tim Dawson, one of the Queensland-based founders of Witch Beam.
“From what I could tell [Microsoft] combined a team of researchers, the most cutting edge genAI models, a huge amount of resources and the fully functional videogame Quake II, and the result is Quake II that can’t remember which room you’re in,” he said.
Chantal Ryan, the Adelaide-based director of We Have Always Lived In The Forest, whose upcoming game darkwebSTREAMER uses a combination of complex AI technologies, labelled the Quake II demo a “barely-glorified image simulator”.
“I can’t imagine releasing something in that state and calling it an example of next-gen technology,” she said.
“It has no congruence — it struggles to remember where you are and what you’re looking at. There’s no interaction, no meaningful mechanic.”
It’s even been suggested that the way Microsoft presented the demo left its researchers “hung out to dry.” Presenting in-progress work without adequate context has its drawbacks, like the impact the 2022 Grand Theft Auto VI leaks had on employees.
But Muse’s issues appear to run deeper than simply highlighting unfinished work.
Artificial intelligence and game development
Artificial intelligence and video games are inseparable. Any form of interactivity – like pathfinding, non-playable character (NPC) behaviour, or which levels will appear next – rely on some form of AI.
“AI has existed in video games since the dawn of them – the earliest video game enemies had rudimentary forms of AI,” Ryan said.
“When a game responds to us in a way that makes us feel like they’re assessing and predicting our movements or actions, that’s AI.”
Differences between AI technologies often gets lost in the generative AI discussion. Procedural generation forms the basis of darkwebSTREAMER, a psychological horror game that includes NPCs and an in-game internet that reacts to player choices. It’s a technique employed by many games so that it feels as if every experience is unique, encouraging replayability.
“In layman’s terms, procedural generation is a bit like having a large deck of cards and saying, ‘The experience you will encounter will be determined based on the random sequence of cards I will draw’,” Ryan said.
“The underlying data being distributed remains the same – the cards in the deck do not change, they’re always four suits of thirteen cards a piece – but the order in which you experience them, if you experience them at all, will make each encounter seem different.”
Ryan, who has an academic background in digital humanities and machine learning, added that generative AI is “far more complex”, including the well-documented concerns surrounding copyright violations. She also clarified that the AI used by darkwebSTREAMER’s relies on “overwhelmingly hand-written” data by her team and obtained consent for all other data.
So far, Microsoft’s generative AI game demoes and research have used data it owns and has the rights to use. Both Bleeding Edge and Quake II are from teams under the Xbox Game Studios banner that has grown substantially as part of the company’s acquisition-driven strategy.
A concern among developers is that other companies may not limit generative AI use to datasets they own. Legal battles continue between copyright holders and AI companies about consent, compensation, and fair use.
“I started working in the academic generative AI space in 2019 and it was well known you could not charge for data trained on private works without permission and licenses,” Ryan said. “Unfortunately capitalism often sides with protecting wealth generation even at the expense of justness and legality.”
Generative AI and game preservation
Microsoft made several big claims when it started talking about Muse. Fatima Kardar, Microsoft’s Corporate Vice President of Gaming AI, claimed that, among other things, the generative AI technology has the potential to assist with game preservation.
“We believe [Muse] could radically change how we preserve and experience classic games in the future and make them accessible to more players,” Kardar said in a company blog post.
Microsoft’s discussion around Muse so far indicates that the company is trying to see if generative AI can get older games running on modern hardware. The intention is to attempt automating what is typically labour-intensive process.
Based on the public demoes of Bleeding Edge and Quake II, scepticism remains over generative AI’s capacity to properly handle the task of game preservation. Seb Chan, ACMI Director and CEO, explained that the act of preserving a game goes far beyond what appears on the surface.
“There is a key difference between preserving a game and remaking it, and it appears that Muse GenAI tools as they stand are able to make increasingly plausible replicas of some well-known old games based on training data,” Chan said.
“But this is very different from preserving the game: the original code, the original hardware and/or emulating the original hardware to run a game long into the future.”
ACMI is one of multiple major Australian institutions taking an active role in preserving games for future generations. But the process is far more involved than just curating source code and old hardware; it’s also about preserving a game’s history. ACMI’s upcoming Game Worlds exhibition, opening in September, embodies this, aiming to present the relationships between developers and players across games like World of Warcraft and The Sims.
“…game preservation also involves understanding and documenting the context in which the original game creators were working – what were they trying to do, what was possible with the hardware they had available to them, what was the economic context they were operating in, all that,” Chan said.
“And also, it’s vital to understand and document the experiences of players: how did they enjoy or find meaning in the game, did players make mods or cheat codes to get adapt the game to their needs, all that too.”
Chan also elaborated on the complex work behind retaining a game’s files. Often, it involves obtaining the original physical media the software came on, including ROM chips, floppy disks, and cartridges.
However, modern industry trends have made this task increasingly difficult. The popularity of “games-as-a-service”, online-based games that evolve over time through regular updates, are nearly impossible to preserve.
“Games-as-a-service titles that are run from cloud servers are only currently able to be documented through developer and player interviews and screen capture recordings,” Chan said. “Even if one version of a client and server version was able to be saved, the ability to run it would require vast infrastructure and also pose many new legal challenges.”
Aside from aiding with repetitious tasks like documentation and cataloguing, Chan is “doubtful” about generative AI’s usefulness in preserving games’ code.
“Preservation is full of complex human challenges, not least of all the legal ones – and AI is not useful in helping humans find agreement with each other that game preservation matters.”
Remastering a classic
Microsoft’s decision to use Quake II as the poster child for generative AI gameplay has proven to be a curious one. In 2023, Bethesda released a remastered version of the 1997 shooter, adding an extensive list of enhancements designed for modern platforms.
It went beyond the standard spit-and-polish of a standard remaster, adding an entirely new campaign expansion and restoring cut content. “This is how to remaster a game” was the headline on John Linneman’s technical analysis on Digital Foundry, praising Quake II’s re-release for bringing it to modern audiences while retaining the heart of the source material.
A collaboration between Nightdive Studios, id Software, and Machine Games, the 2023 version of Quake II stands as a testament to the value of a finely crafted remaster. Describing Quake II as “a great example of game preservation”, Dawson believes that the act of remastering is “a poor fit for automation.”
“I think the hope is that maybe a sufficiently advanced AI could one day be shown footage of a game and reconstruct it, but to me that sounds like wishing for magic,” Dawson said.
“Today, videogame historians, emulator developers and communities of enthusiasts are doing the boots-on-the-ground work of actually preserving videogames, and to not acknowledge that in favour of a technology that may never actually be viable, feels disrespectful.”
Does generative AI game development have a future?
Ryan believes that “any tech can feasibly be used as a tool in game development”, but “just because we can, doesn’t mean we should.”
“I can see the potential for a radically advanced version of Muse that communicates with a backend [large language model] to communicate code to visuals, and visual simulation back to code,” she said.
“But it would require enormously powerful computers and likely preposterous amounts of energy to do so.”
Reflecting on generative AI’s potential, Ryan doesn’t believe it currently is capable of improving game development, adding that it “seems like a highly inefficient, unreliable way to do something we already do well.”
In addition to the ethical concerns and energy demands of generative AI, Dawson also struggles to see a future for the technology. He argued that any benefits of automating repetitive tasks are outweighed by the limitations place on creativity.
“When I see a system that instead promises to automate creative decisions, my first question is ‘what’s making the decisions?'” Dawson said.
“If the answer is ‘training data’ then what’s really being offered is a way to recycle other people’s ideas, which might seem like a time saver if you struggle with that part, but feels ultimately short sighted in a field that relies on new and novel experiences to push the medium forward.”
Prevalent throughout Dawson’s work is intentionality. From the carefully sculpted levels of Assault Android Cactus to the rhythmic clockwork nature of Tempopo’s musical puzzles, Witch Beam’s games – like many games made by teams around the world – are the result of deliberate decisions made by people.
“Ultimately it’s the human-to-human connection that drives successful art – commercial or otherwise – and the big secret at the heart of any creative endeavour, including video games, is that when we make art, the art makes us,” he said.
“It’s a two-way process, the decisions we make every step of the way teach us and plant the seeds for future ideas, and I believe tech companies that fail to understand this do so at their peril.”
Crucially, Microsoft and Xbox leadership isn’t mandating the use of generative AI. Kardar’s blog post mentioned that Xbox’s creative leaders get to make that decision.
“There isn’t going to be a single solution for every game or project, and the approach will be based on the creative vision and goals of each team,” Kardar wrote.
Dawson added that while he appreciated Microsoft’s ambitions in trying new things, he hopes that the company sticks to what it does best.
“The ID@Xbox program in particular has been amazing at reaching out and empowering small teams from around the world to bring their titles to a major platform,” Dawson said.
“My sincere hope is Microsoft never forget about the human beings that love making video games that they already have on their side.”
The post “Barely-glorified image simulator”: Developers react to Microsoft’s AI Quake II demo appeared first on GadgetGuy.
0 comments:
Post a Comment