07-what-shipped 4,422 words about 19 min

Chapter 7. What shipped

You pick a fantasy setting and type that you draw your sword and step into the tower. A paragraph comes back describing the stairwell, the smell of smoke, and someone moving on the landing above. You type that you call out, and another paragraph follows. An hour later you mention the silver key you picked up near the start, and the story has no idea what you mean.

There was never a key, in any sense the program could check. AI Dungeon has no world state, no inventory, and no rules engine. Each turn it assembles a prompt from the scenario premise, whatever you pinned to its memory box, and the last stretch of the story, and a language model writes the next passage: GPT-2 with 1.5 billion parameters when AI Dungeon 2 launched in December 2019, GPT-3 for the premium model from July 2020 [1]. A fact survives as long as it stays inside the model's context window. Your tools for keeping the story straight are undo, retry, edit, and the memory box, where you write down the key yourself.

Four years later Suck Up! put a model behind a front door. You play a vampire who has to be invited in. You choose a disguise, knock, and talk into a real microphone; the resident answers out loud, in character, and decides whether to let you in. Each resident's personality is written by the developers. Each reply comes from a hosted language model given that personality and your words; the store page names ChatGPT, and the 2023 terms name GPT 3.5 and Llama 2 and meter use in tokens [2, 3]. A disguise works because it changes what the model is told about you. Players write new residents and whole neighbourhoods and share them on Steam Workshop.

The dataset classes both games as boundary cases. In AI Dungeon nothing is simulated outside the model's text, and in Suck Up! the model decides whether your pitch worked, with no social system underneath keeping score. They show what a game looks like when the model is the whole game.

How the dataset counts maturity

A technique enters the dataset only as a dated, citable result that changes at least one mechanic. There are 45 of them, from tree search and wave-function collapse to world models that generate a playable scene from a sentence. Each carries a maturity judged for deployment in games, which is a stricter standard than maturity in AI generally. Constrained decoding, which forces a model's output into a fixed format, is routine in software engineering and rare in shipped game runtimes, so it is scored as research here.

The scale has four steps. Research means papers and lab demos. Prototype means public playable demos or tech previews with no sustained commercial title. Shipped (niche) means at least one commercial game uses it at runtime or in production. Shipped (mainstream) means it is common across studios.

Figure: Techniques by maturity in games. Every technique in the dataset under how far it has got in games, each with the year of the result the dataset cites, oldest first; the one shipped technique that is a norm across studios is marked. The mark shows what the technique does to reproducibility when it runs. Interactive version: https://sim.yohei.me/#fig-technique-landscape

1 technique is a genre norm, 12 have shipped in at least one commercial game, 10 exist as public prototypes, and 22 are research only. The shipped ones fall into four groups: rendering and assets, characters that talk, opponents and tools, and a single case of a model writing game rules.

The one norm: frames

The one learned technique that has become a genre norm is neural rendering. Upscaling reconstructs a sharp frame from a lower-resolution render and the motion of each pixel; frame generation predicts whole frames between the ones the game rendered; small networks inside shaders decode compressed textures and denoise ray-traced light. By January 2026 more than 250 titles supported multi-frame generation and more than 400 games and apps supported upscaling [4]. Simulation games are among them: Microsoft Flight Simulator (2020) supports both upscaling and frame generation, and Cities: Skylines II supports upscaling [5, 6].

A generated frame is drawn between two frames the game computed, and the simulation never learns it exists. Nothing about the rules, the saves, or the multiplayer code has to change, and the dataset records it as preserving determinism. The same separation shows in the one place it has gone badly. DLSS 5 launched in September 2026 with a final network pass that adds lighting and material detail to the finished frame, and its reveal drew a largely negative reaction from players and developers over altered faces and art direction [7, 8].

Assets, places, and motion

Microsoft Flight Simulator (2020) rebuilt the planet from map data. Machine learning running on Azure classified buildings, trees, and terrain in Bing Maps imagery, and procedural rules turned them into about 1.5 billion buildings and 2 trillion trees, streamed to the player from more than two petabytes of data as they fly [9]. The reconstruction is done offline and the result is fixed data, so two players flying over the same town see the same houses. Microsoft Flight Simulator 2024 carries the same scenery system over [10]. Its record lists two limits: where imagery is blurred the game falls back to procedural fill, and outside the photogrammetry cities each building is generic, rebuilt from an inferred class and height.

Text-to-3D models such as TRELLIS turn a prompt or an image into a textured mesh in seconds to a minute on a large GPU [11]. inZOI exposes the idea to players: a 2D image becomes a 3D object in the game, and an uploaded video becomes a character's motion [12]. A generated asset has to be stored to be reproduced, since asking the model again gives a different chair, so the dataset marks the technique as mitigable rather than safe.

TRELLIS.2, at 4 billion parameters, outputs a textured asset in about 3 seconds at its lowest resolution and about a minute at its highest on an H100 [13]. Tencent's Hunyuan3D 2.1 added physically based texturing and needs about 29 GB of video memory for shape and texture together [14]. Roblox's Cube models generate shapes and, since May 2026, objects with separately controllable parts, aimed at assets that work with the platform's animation, physics, and scripts [15, 16]. Hardware of that size puts player-side generation in the cloud or on high-end PCs. A generated chair also arrives as a single object with no function, footprint, or price, which the game has to assign, and inZOI's store page describes its AI tools as auxiliary and says the team reviews generated content [12].

Learned motion has shipped in sports and life games. Ubisoft's learned motion matching replaced a 590 MB database of motion-capture frames with about 17 MB of network weights and nearly identical animation [17]. FIFA 22's HyperMotion generated player animation in real time from a model trained on more than 8.7 million frames of full-team match capture, on the new consoles and Stadia only [18]. No physics is simulated in either case; the networks choose the next pose. Kinematic output ignores physics, so a character can clip or float when the environment differs from the training motion, and learned motion matching saves memory without saving the per-character CPU time that limits crowd sizes. Unreal Engine ships the non-learned version, database motion matching in its Pose Search plugin, which is what the learned version compresses [19].

Generated assets are also where most of the store's AI activity sits. About a thousand Steam games had disclosed generative AI by April 2024 [20]. By July 2025 the count was 7,818 titles, about 7 percent of the store, with just under a fifth of that year's releases disclosing and visual assets named in about 60 percent of disclosures [21]. Valve's content survey separates pre-generated content from content generated live during play and asks developers of the second kind to describe their guardrails [22].

Characters that talk

In inZOI, released into early access in March 2025, a small language model running on the player's graphics card lets typed prompts steer a Zoi's thoughts and actions [12]. NVIDIA describes these Smart Zois as agents that plan, act, and reflect on their decisions [23]. The needs, the animations, and the action set are inZOI's; the model reads a Zoi's needs, surroundings, and thoughts to help choose among those actions, where a life sim would otherwise rely on scored need curves alone. Where Winds Meet, a martial-arts RPG, discloses that its NPCs answer free text and voice through a live model [24], while its quests and combat stay authored and rule-based. Whispers from the Star is built around one character, the stranded astronaut Stella, whose spoken replies and choices are generated in a voice developed with an actor, behind safety filters; whether she survives depends on what you tell her [25]. Outside the catalog, KRAFTON ran a two-week public beta in June 2026 of PUBG Ally, a teammate that takes spoken orders through an on-device model [26].

Voice got cheaper and usable models got smaller in the same two years. One leading realtime speech API cut its audio prices from 100 and 200 dollars per million input and output tokens at launch in October 2024 to 32 and 64 dollars in August 2025 [27, 28]. On the device side, pruning and distillation produced instruction models small enough that a 2-billion-parameter one fits in about 1.5 GB of video memory beside the game's own rendering [29, 30], although NVIDIA's in-game inference kit still asks for at least an 8 GB card from the RTX 30 series [31].

Moshi, a full-duplex speech model, reported 160 ms of theoretical and 200 ms of practical latency [32], and Inworld's TTS-1.5 lists time to first audio under 130 ms for its small model at 15 dollars per million characters, about a cent and a half per minute of speech [33]. A turn through a text model is slower: in a study of 28 players talking to GPT-4 characters, lag of about 15 seconds frustrated 40 percent of them [34]. Delivery still trails people. In Sesame's listening tests, its generated speech matched human recordings when each was heard alone, and listeners preferred the human recordings once they had heard 90 seconds of the conversation before it [35]. The record's estimate from list prices puts an hour split evenly between player and character speech at about 9 dollars at the 2024 realtime rates and about 3 dollars at the 2025 ones, before the model rereads its context. Most games keep a text layer between speech and game state, because a model that works in speech directly is hard to constrain to the actions a game allows.

A local model can be seeded and run the same way twice, so on-device characters are mitigable. Characters built as full generative agents, language-model game masters, and realtime voice are recorded as breaking reproducibility outright: the same input can produce a different reply, and a replay has to store what was said.

A grounded variant of the language-model storyteller was in early access by August 2025. Hidden Door runs each player intent through planning, a check against the setting's canon and rules, and an update to a typed world state before any text is shown; its example blocks Dracula from entering a house he was not invited into. The company reports that a 1.7-billion-parameter model tuned on real play beat a 120-billion-parameter general model on one of its core narrative tasks [36, 37]. Hidden Door is outside the catalog, but its order of operations, state first and prose second, is the architecture chapter 11 describes.

Societies in the lab

Two catalog entries are research demonstrations rather than products. Generative Agents, the 25-character town of chapter 6, showed memory, reflection, and planning spreading news through a community [38]. Project Sid reported Minecraft runs of up to 500 agents; attempts at 1,000 overloaded the server. In a 29-agent run with a constitution and a 20 percent tax, influencers argued for and against the tax, amendments passed, and payments followed the new rate; in a 500-agent run, a religion seeded by 20 priests spread to other towns [39].

Both are research runs: they need a server, they cost real money per hour, and neither can be replayed exactly. The dataset records the generative-agent technique as shipped on the strength of inZOI, whose Smart Zois plan, act, and reflect with a small local model.

A third research line puts a game master inside the experiment. In Concordia, a Google DeepMind library for generative social simulation, a Game Master turns each agent's stated action into an outcome and checks it for physical plausibility, and a 2025 design pattern sorts its uses into simulationist, dramatist, and evaluationist [40, 41, 42]. The dataset files this grounded Game Master under the same technique as AI Dungeon, where the model continues the story from the player's typed intent.

Opponents and tools

The oldest shipped technique in the set plays opponents. Monte Carlo tree search, introduced in 2006, plans by playing many random games forward from the current position and backing up which moves won [43]. Total War: Rome II used it for campaign decisions and Fable Legends for tactical planning [44]. A 2012 survey catalogued its variants across board, card, and video games after its first six years [45].

In a simulation game, search needs a forward model it can copy and step quickly, and a full simulation with thousands of agents is expensive to clone, so practical use relies on an abstracted model or macro-actions. Random playouts judge positions badly in systems where consequences arrive late, such as economies and colony survival, unless a designer supplies a rollout policy or an evaluation function. And when some of the state is hidden, search that averages over sampled worlds can choose moves that are good only for a player who could see what is hidden. Chapter 11 takes up search on a time budget.

The best-known learned opponent in the catalog is Gran Turismo Sophy. Sony AI trained it with deep reinforcement learning on many parallel copies of the Gran Turismo simulator, with a reward that values course progress and encodes racing's sportsmanship rules, and in 2022 it beat four of the world's best Gran Turismo drivers head to head [46]. It reached players in Gran Turismo 7 as a time-limited event in February 2023 [47] and became a permanent Quick Race rival that November on nine tracks and 340 cars [48], then moved into custom races in 2025. Sophy drives the same tire, suspension, and weight-transfer model the player does, so the dataset records it as deepening vehicle dynamics and embodied skill rather than replacing either.

Studios ask more of an opponent than strength: whether it is fair, whether it can be tuned to a player's level, and whether it fits inside a content update, and Sophy arrived one update at a time with Beginner, Intermediate, and Expert settings [47].

Two tools shipped on the developer's side. Agent-based playtesting predicts difficulty and finds exploits before release: a network trained on logged moves in Candy Crush Saga predicted level difficulty better than tree search at a fraction of the compute [49], and EA used search agents to balance careers in The Sims Mobile [50]. Wave-function collapse, a constraint solver that fills a grid tile by tile from an example, generates levels in Bad North and Caves of Qud and runs live in Townscaper, where every block you click reruns the solver over the cells around it [51, 52]. Both are deterministic given their seeds.

EA found that deep reinforcement learning needed heavy tuning for each game, so the effort grew roughly in step with the number of titles [50]. Its research division's test agents found a wall with no collision mesh that scripted navigation agents had missed and located all five stuck areas on a test map, and rewards for curiosity raised how much of a complex 3D map the agents covered [53, 54]. When it added an experimental reinforcement-learning system to the scripted test bots of titles including Battlefield 2042 and Dead Space, integration took more time than the algorithms [55]. In mobile games, agents combined with a model of player skill, persistence, and boredom predicted pass and churn rates for each level of Angry Birds Dream Blast [56]. The record's limit for simulation games is the goal: an open-ended sim has no single objective to optimize, so coverage or curiosity rewards stand in for design intent and can miss balance problems that appear only over long play.

Wave-function collapse is constraint solving with familiar heuristics, and Karth and Smith showed what that implies: with one added global constraint, its restarts failed to finish within a minute while a backtracking solver resolved the same case quickly [57]. It enforces only which tiles may sit next to each other, so reachability, room counts, and path lengths need constraints of their own. The shipped games handle contradictions differently. Bad North, with more than 300 tiles, restarts generation, and Townscaper accepts a silent local failure where cells cannot fit [52]. Caves of Qud, one of the early users, reached version 1.0 in December 2024 after an early access that began in 2015 [58].

What writes into the simulation

Sort the shipped techniques by whether they change authoritative game state, the state that saves, replays, and multiplayer depend on, and the list is short. A cached model output can become a rule, as in Infinite Craft, where the first answer to a new combination is stored and served to every later player; chapter 8 opens there. Language-model characters help choose actions in inZOI and a few AI-native titles. Search and learned policies drive opponents. Everything else in the shipped group changes what the player sees and hears, or does its work before release.

The 13 shipped techniques affect 41 mechanics, and those mechanics account for 23 percent of the uses in the catalog. Of the 33 edges in which a technique would replace how a mechanic works, 9 come from techniques that have shipped.

No learned replacement for physics, fields, or the world itself runs in a catalog game. The models that would replace the engine outright are prototypes. GameNGen simulated DOOM at 20 frames per second on a single TPU, and human raters picked the real game over the simulation only 58 to 60 percent of the time on short clips; its authors state it cannot be used to create new games [59]. Genie 3 renders 720p at 24 frames per second from a text prompt and keeps about a minute of visual memory [60]. When Google opened it to subscribers as Project Genie in January 2026, each world allowed sixty seconds of play, a limit DeepMind attributed to compute, and a reviewer watched characters walk through walls [61]. Microsoft's WHAMM ran a section of one Quake II level at more than ten frames per second with blurry enemies, damage that could come out wrong, and health counters that could not be trusted [62]. Learned physics has one prototype, subspace neural physics, described below, and graph-network simulators remain research.

A world model holds the game's state as pixels. Position, health, and the layout of the map exist only in the frames it remembers, so anything that has scrolled out of that memory can come back different. Simulation games depend on the opposite property: a cow in a field on day one is the same cow on day three hundred, with a history the player can inspect.

Prototypes beside the engine

4 of the prototypes are world models, WHAM among them. None of the others has an edge that replaces how a mechanic works, and those described below, other than WHAM, work beside a game's simulation and leave it alone.

Gaussian splatting represents a scene as millions of small 3D Gaussians fitted to photographs and drawn without a neural network at runtime. The 2023 method rendered new views at 1080p and 30 frames per second or more [63], and a scene trains in about 35 to 45 minutes against about 48 hours for an earlier neural method, with training memory peaking above 20 GB [64]. PhysGaussian runs a material solver on the Gaussians themselves, so a scanned object can bend, dent, or flow without a mesh [65], and Khronos released a glTF extension for splats as a release candidate in February 2026 [66]. The colours are captured light, so a splat scene has its lighting baked in, and it carries no collision geometry or semantics: a game needs separate meshes for physics and pathfinding, and a day-night cycle needs another method.

Generated 3D worlds build on splats. World Labs' Marble, generally available since November 2025, turns text, images, video, or a coarse layout into an explorable scene and exports splats, meshes, and coarse collider meshes; its Chisel mode takes a blockout of boxes and planes as structure and a prompt as style [67]. Tencent's HunyuanWorld 1.0 lifts a generated panorama into a layered mesh for standard graphics pipelines [68]. The result persists and can be saved, unlike the frames of a world model, but it is static: no agents, rules, or moving objects come with it.

Real-time restyling repaints a running game. Decart's MirageLSD returns each frame in under 40 ms at 24 frames per second from the current frame, its own recent output, and a text prompt [69], and Oasis 2.0 packages it as a Minecraft mod whose prompt changes by chat command or by command blocks placed in the world [70]. Collision, inventory, and world state stay in Minecraft. The record's warning is about legibility: a block repainted to look like a door is still solid.

Microsoft's WHAM, which preceded the WHAMM demonstration above, was aimed at designers. It was trained on more than a billion images and controller actions, seven years of continuous play, from Bleeding Edge matches recorded under the game's licence agreement, and it predicts both the next frames and what a player is likely to press [71]. A designer can insert a character into a frame and the model carries the edit forward into plausible play, which its Nature paper presents as support for gameplay ideation; the 1.6-billion-parameter version produces about one frame per second at 300 by 180 pixels [72].

Subspace neural physics is the only prototype the dataset judges usable in a production title started today. A small network trained on offline cloth or flesh simulation predicts deformation in a compressed space, and Ubisoft La Forge reported it running hundreds to thousands of times faster than the simulation it learned from [73]. Unreal Engine's ML Deformer trains such models on deformation caches from Maya or Houdini and applies them at runtime as GPU morph targets [74]. The record found no commercial title confirmed to use these models at runtime, and their output is usually cosmetic: it does not feed back into gameplay physics unless the game is built to read it.

Reproducibility

Among the shipped techniques, 3 break it: generative agents, language-model game masters, and realtime voice.

The fix is the move Infinite Craft makes with its cache: record each model output, and a replay reads the record instead of asking the model again.

How it was received

Developers' opinion moved against generative AI while their use of it held steady. In GDC's surveys, 36 percent of respondents reported using it in both 2025 and 2026, and the share who believe it harms the industry rose from 30 to 52 percent [75, 76]. inZOI drew negative user reviews after Krafton announced its generative features [77]. ARC Raiders was criticized for synthetic voice lines, some of which were later re-recorded by actors [78], and the SAG-AFTRA video game strike that ran from July 2024 to July 2025 ended with contract terms on consent for AI voice replicas [79]. AI Dungeon had met the moderation side of the problem earlier, with an audience of more than 100,000 players in AI Dungeon 2's first week and 1.5 million by June 2020. Its 2021 content filter, added to block sexual content involving minors, flagged innocent text, and human moderators read the flagged private stories, which led to a wave of negative reviews [1]. Games with disclosures still sold: an estimate in November 2025 put them at roughly 8 percent of the store with about 660 million dollars in combined gross [80].

Exercise: classify an AI feature

Pick a game with an AI feature, from the catalog or from a recent announcement.

1. Name the mechanic it changes, using the catalog's names. If it changes none, it belongs to presentation.
2. Decide whether it writes authoritative state (something a save file or a replay must contain) or only changes what the player sees and hears.
3. Note when it runs: before release, while the player builds, or during play.
4. Ask what happens to a replay, a multiplayer session, or a bug report if the model answers differently tomorrow. If the answer is "it breaks", find the log or cache that would fix it.

The features that pass step 2 are rare, and almost all of them sit in agents, narrative, and opponents. Chapter 8 counts every technique in the set, shipped or not, against every mechanic, and finds that the families most games are built from are the ones the field has worked on least.

Sources

[1] Wikipedia contributors (2026). AI Dungeon. Wikipedia https://en.wikipedia.org/wiki/AI_Dungeon
[2] (2026). Suck Up! on Steam. Steam https://store.steampowered.com/app/2726370/
[3] (2023-12-05). Suck Up! Terms of Service. playsuckup.com https://www.playsuckup.com/terms
[4] NVIDIA (2026-01-06). NVIDIA DLSS 4.5 Delivers Major Upgrade With 2nd Gen Transformer Model For Super Resolution and 6X Dynamic Multi Frame Generation. GeForce News https://www.nvidia.com/en-us/geforce/news/dlss-4-5-dynamic-multi-frame-gen-6x-2nd-gen-transformer-super-res/
[5] PCGamingWiki contributors (2026). Microsoft Flight Simulator (2020). PCGamingWiki https://www.pcgamingwiki.com/wiki/Microsoft_Flight_Simulator_(2020)
[6] PCGamingWiki contributors (2026). Cities: Skylines II. PCGamingWiki https://www.pcgamingwiki.com/wiki/Cities:Skylines_II
[7] NVIDIA (2026-09-01). NVIDIA DLSS 5 Available September 3rd: 3D-Guided Neural Rendering Debuts In NBA 2K27. GeForce News https://www.nvidia.com/en-us/geforce/news/dlss-5-3d-guided-neural-rendering/
[8] Wikipedia contributors (2026). Deep Learning Super Sampling. Wikipedia https://en.wikipedia.org/wiki/Deep_Learning_Super_Sampling
[9] Wikipedia contributors (2026). Microsoft Flight Simulator (2020 video game). Wikipedia https://en.wikipedia.org/wiki/Microsoft_Flight_Simulator(2020_video_game)
[10] Wikipedia contributors (2026). Microsoft Flight Simulator 2024. Wikipedia https://en.wikipedia.org/wiki/Microsoft_Flight_Simulator_2024
[11] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang et al. (2024-12). Structured 3D Latents for Scalable and Versatile 3D Generation. CVPR 2025 https://github.com/microsoft/TRELLIS
[12] inZOI Studio (2025-03). inZOI on Steam. Steam https://store.steampowered.com/app/2456740/inZOI/
[13] Microsoft (2025-12). TRELLIS.2 repository. GitHub https://github.com/microsoft/TRELLIS.2
[14] Tencent Hunyuan (2025-06-13). Hunyuan3D-2.1 repository. GitHub https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1
[15] Roblox Foundation AI Team (2025-03). Cube: A Roblox View of 3D Intelligence. https://arxiv.org/abs/2503.15475
[16] Roblox (2026). Roblox Cube repository. GitHub https://github.com/Roblox/cube
[17] Ubisoft La Forge (2020-07-30). Introducing Learned Motion Matching. https://www.ubisoft.com/en-us/studio/laforge/news/6xXL85Q3bF2vEj76xmnmIu/introducing-learned-motion-matching
[18] Wikipedia contributors (2026). FIFA 22. Wikipedia https://en.wikipedia.org/wiki/FIFA_22
[19] Epic Games (2026). Motion Matching in Unreal Engine. Unreal Engine documentation https://dev.epicgames.com/documentation/en-us/unreal-engine/motion-matching-in-unreal-engine
[20] Totally Human Media (2024-04-24). The Surprising Number of Steam Games that Use GenAI. Totally Human blog https://www.totallyhuman.io/blog/the-surprising-number-of-steam-games-that-use-genai
[21] Totally Human Media (2025-07-13). The NEW Surprising Number of Steam Games that Use GenAI. Totally Human blog https://www.totallyhuman.io/blog/the-surprising-new-number-of-genai-games-on-steam
[22] Valve (2026). Steamworks Documentation: Content Survey (generative AI disclosure). Steamworks documentation https://partner.steamgames.com/doc/gettingstarted/contentsurvey
[23] NVIDIA (2026). NVIDIA ACE for Games. NVIDIA Developer https://developer.nvidia.com/ace
[24] Everstone Studio, NetEase Games (2025-11-14). Where Winds Meet on Steam. Steam https://store.steampowered.com/app/3564740/Where_Winds_Meet/
[25] (2026). Whispers from the Star on Steam. Steam https://store.steampowered.com/app/3730100/Whispers_from_the_Star/
[26] KRAFTON (2026-06-16). KRAFTON Introduces PUBG Ally Beta Test. KRAFTON press release https://www.krafton.com/en/news/press/krafton-introduces-pubg-ally-beta-test/
[27] OpenAI (2024-10-01). Introducing the Realtime API. OpenAI blog https://openai.com/index/introducing-the-realtime-api/
[28] OpenAI (2025-08-28). Introducing gpt-realtime. OpenAI blog https://openai.com/index/introducing-gpt-realtime/
[29] Sharath Turuvekere Sreenivas, Saurav Muralidharan, et al., Bryan Catanzaro (2024-08-21). LLM Pruning and Distillation in Practice: The Minitron Approach. arXiv https://arxiv.org/abs/2408.11796
[30] NVIDIA (2025). NVIDIA ACE autonomous game characters: PUBG Ally, inZOI Smart Zoi, NARAKA AI teammate. GeForce News https://www.nvidia.com/en-us/geforce/news/nvidia-ace-autonomous-ai-companions-pubg-naraka-bladepoint/
[31] NVIDIA (2026). NVIGI-Core: NVIDIA In-Game Inferencing SDK core. GitHub https://github.com/NVIDIA-RTX/NVIGI-Core
[32] Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Hervé Jégou et al. (2024-09-17). Moshi: a speech-text foundation model for real-time dialogue. arXiv https://arxiv.org/abs/2410.00037
[33] Inworld AI (2026-01-21). Inworld TTS-1.5: Upgrading the #1 Ranked TTS Model with Production-Grade Latency, Expression and Stability. Inworld blog https://inworld.ai/blog/introducing-inworld-tts-1-5
[34] Xiangyu Peng, Jessica Quaye, Sudha Rao, Weijia Xu, et al., Bill Dolan (2024-04-25). Player-Driven Emergence in LLM-Driven Game Narrative. arXiv https://arxiv.org/abs/2404.17027
[35] Sesame (2025-02-27). Crossing the uncanny valley of conversational voice. Sesame research https://www.sesame.com/research/crossing_the_uncanny_valley_of_voice
[36] Hidden Door (2026). Hidden Door blog. hiddendoor.co https://www.hiddendoor.co/blog
[37] Hidden Door (2026). Hidden Door. hiddendoor.co https://www.hiddendoor.co/
[38] Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, Michael S. Bernstein (2023-04-07). Generative Agents: Interactive Simulacra of Human Behavior. UIST '23 (ACM Symposium on User Interface Software and Technology) https://arxiv.org/abs/2304.03442
[39] Altera.AL, Andrew Ahn, Nic Becker, Stephanie Carroll, Nico Christie, Manuel Cortes et al. (2024-10-31). Project Sid: Many-agent simulations toward AI civilization. arXiv https://arxiv.org/abs/2411.00114
[40] Alexander Sasha Vezhnevets, John P. Agapiou, Avia Aharon, Ron Ziv, Jayd Matyas, Edgar A. Duéñez-Guzmán et al. (2023-12-06). Generative agent-based modeling with actions grounded in physical, social, or digital space using Concordia. arXiv https://arxiv.org/abs/2312.03664
[41] Alexander Sasha Vezhnevets, Jayd Matyas, Logan Cross, Davide Paglieri, Minsuk Chang, William A. Cunningham et al. (2025-07-10). Multi-Actor Generative Artificial Intelligence as a Game Engine. arXiv https://arxiv.org/abs/2507.08892
[42] Google DeepMind (2026). Concordia: a library for generative social simulation. GitHub https://github.com/google-deepmind/concordia
[43] Levente Kocsis, Csaba Szepesvári (2006). Bandit Based Monte-Carlo Planning. Machine Learning: ECML 2006 (LNCS), pp. 282-293 https://doi.org/10.1007/11871842_29
[44] Wikipedia contributors (2026). Monte Carlo tree search. Wikipedia https://en.wikipedia.org/wiki/Monte_Carlo_tree_search
[45] Cameron B. Browne, Edward Powley, Daniel Whitehouse, Simon M. Lucas, Peter I. Cowling, Philipp Rohlfshagen et al. (2012-03). A Survey of Monte Carlo Tree Search Methods. IEEE Transactions on Computational Intelligence and AI in Games 4(1):1-43 https://doi.org/10.1109/TCIAIG.2012.2186810
[46] Peter R. Wurman, et al. (2022-02-09). Outracing champion Gran Turismo drivers with deep reinforcement learning. Nature 602:223-228 https://doi.org/10.1038/s41586-021-04357-7
[47] Polyphony Digital (2023-02-20). Gran Turismo Sophy Race Together event in Gran Turismo 7. gran-turismo.com https://www.gran-turismo.com/us/gt7/news/00_5962264.html
[48] Polyphony Digital (2023-11-02). Gran Turismo Sophy 2.0 permanently added to Gran Turismo 7. gran-turismo.com https://www.gran-turismo.com/us/news/00_4176595.html
[49] S. Gudmundsson, Philipp Eisen, E. Poromaa, Alex Nodet, Sami Purmonen, B. Kozakowski et al. (2018). Human-Like Playtesting with Deep Learning. IEEE Conference on Computational Intelligence and Games (CIG 2018) https://doi.org/10.1109/CIG.2018.8490442
[50] Yunqi Zhao, Igor Borovikov, Fernando de Mesentier Silva, Ahmad Beirami, Jason Rupert, Caedmon Somers et al. (2019-03-25). Winning Isn't Everything: Enhancing Game Development with Intelligent Agents. arXiv https://arxiv.org/abs/1903.10545
[51] Maxim Gumin (2016-09-30). WaveFunctionCollapse. GitHub (mxgmn) https://github.com/mxgmn/WaveFunctionCollapse
[52] Tommy Thompson (2022-04-06). How Townscaper Works: A Story Four Games in the Making. Game Developer https://www.gamedeveloper.com/game-platforms/how-townscaper-works-a-story-four-games-in-the-making
[53] Joakim Bergdahl, Camilo Gordillo, Konrad Tollmar, Linus Gisslén (2020-08). Augmenting Automated Game Testing with Deep Reinforcement Learning. IEEE Conference on Games 2020 https://arxiv.org/abs/2103.15819
[54] Camilo Gordillo, Joakim Bergdahl, Konrad Tollmar, Linus Gisslén (2021-08). Improving Playtesting Coverage via Curiosity Driven Reinforcement Learning Agents. IEEE Conference on Games 2021 https://arxiv.org/abs/2103.13798
[55] Jonas Gillberg, Joakim Bergdahl, Alessandro Sestini, Andrew Eakins, Linus Gisslén (2023-08). Technical Challenges of Deploying Reinforcement Learning Agents for Game Testing in AAA Games. IEEE Conference on Games 2023 https://arxiv.org/abs/2307.11105
[56] Shaghayegh Roohi, Asko Relas, Jari Takatalo, Henri Heiskanen, Perttu Hämäläinen (2020-08-29). Predicting Game Difficulty and Churn Without Players. ACM CHI PLAY 2020 https://doi.org/10.1145/3410404.3414235
[57] Isaac Karth, Adam M. Smith (2017-08). WaveFunctionCollapse is Constraint Solving in the Wild. Foundations of Digital Games (FDG 2017) https://adamsmith.as/papers/wfc_is_constraint_solving_in_the_wild.pdf
[58] Freehold Games (2024-12-05). Caves of Qud. Steam https://store.steampowered.com/app/333640/Caves_of_Qud/
[59] Dani Valevski, Yaniv Leviathan, Moab Arar, Shlomi Fruchter (2024-08). Diffusion Models Are Real-Time Game Engines. ICLR 2025 https://arxiv.org/abs/2408.14837
[60] Google DeepMind (2025-08-05). Genie 3: A new frontier for world models. https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/
[61] Rebecca Bellan (2026-01-29). I built marshmallow castles in Google's new AI world generator Project Genie. TechCrunch https://techcrunch.com/2026/01/29/i-built-marshmallow-castles-in-googles-new-ai-world-generator-project-genie
[62] Microsoft Research (2025-04). WHAMM real-time Quake II world model (Microsoft Research blog). https://www.microsoft.com/en-us/research/?p=1135867
[63] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, George Drettakis (2023-08). 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics (SIGGRAPH 2023) https://arxiv.org/abs/2308.04079
[64] Wikipedia contributors (2026). Gaussian splatting. Wikipedia https://en.wikipedia.org/wiki/Gaussian_splatting
[65] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang et al. (2023-11). PhysGaussian: Physics-Integrated 3D Gaussians for Generative Dynamics. https://arxiv.org/abs/2311.12198
[66] The Khronos Group (2026-02-03). Khronos releases KHR_gaussian_splatting glTF extension release candidate. https://www.khronos.org/news/press/gltf-gaussian-splatting-press-release
[67] World Labs (2025-11-12). Marble: A Multimodal World Model. https://www.worldlabs.ai/blog/marble-world-model
[68] HunyuanWorld Team (Tencent) (2025-07). HunyuanWorld 1.0: Generating Immersive, Explorable, and Interactive 3D Worlds from Words or Pixels. https://arxiv.org/abs/2507.21809
[69] Decart (2025). MirageLSD: live-stream diffusion video model. https://decart.ai/publications/mirage
[70] Decart (2025). How to Play Oasis 2.0. https://oasis2.decart.ai/how-to-play
[71] Microsoft Research (2025-02-19). Introducing Muse: Our first generative AI model designed for gameplay ideation. https://www.microsoft.com/en-us/research/blog/introducing-muse-our-first-generative-ai-model-designed-for-gameplay-ideation/
[72] Anssi Kanervisto, Dave Bignell, Linda Yilin Wen, Martin Grayson, Raluca Georgescu, Sergio Valcarcel Macua (2025-02). World and Human Action Models towards gameplay ideation. Nature 638(8051): 656-663 https://www.nature.com/articles/s41586-025-08600-3
[73] Daniel Holden, Bang Chi Duong, Sayantan Datta, Derek Nowrouzezahrai (2019-07). Subspace Neural Physics: Fast Data-Driven Interactive Simulation. ACM SIGGRAPH/Eurographics Symposium on Computer Animation 2019 https://theorangeduck.com/page/subspace-neural-physics-fast-data-driven-interactive-simulation
[74] Epic Games (2026). How to use the Machine Learning Deformer in Unreal Engine. Unreal Engine documentation https://dev.epicgames.com/documentation/en-us/unreal-engine/how-to-use-the-machine-learning-deformer-in-unreal-engine
[75] Game Developers Conference (2025-01). GDC 2025 State of the Game Industry: Devs weigh in on layoffs, AI, and more. gdconf.com https://gdconf.com/article/gdc-2025-state-of-the-game-industry-devs-weigh-in-on-layoffs-ai-and-more/
[76] Game Developers Conference (2026). GDC 2026 State of the Game Industry reveals impact of layoffs, generative AI, and more. gdconf.com https://gdconf.com/article/gdc-2026-state-of-the-game-industry-reveals-impact-of-layoffs-generative-ai-and-more/
[77] Wikipedia contributors (2026). inZOI. Wikipedia https://en.wikipedia.org/wiki/InZOI
[78] Wikipedia contributors (2026). ARC Raiders. Wikipedia https://en.wikipedia.org/wiki/ARC_Raiders
[79] Wikipedia contributors (2026). Artificial intelligence in video games. Wikipedia https://en.wikipedia.org/wiki/Artificial_intelligence_in_video_games
[80] Totally Human Media (2025-11-06). Games With AI Disclosures Have Grossed an Estimated $660M on Steam. Totally Human blog https://www.totallyhuman.io/blog/games-with-ai-disclosures-have-grossed-an-estimated-660m-on-steam

Dataset dfde7895aa6d.

mechanics in this chapter

2 linked or named

Day-night cycleVehicle dynamics