08-where-ai-lands 4,707 words about 20 min

Chapter 8. Where AI lands

You start with four tiles: water, fire, wind, and earth. Drag one onto another and a new tile appears with a name and an emoji. Drag that onto something else. Most pairs you try have been tried before by someone, somewhere, and the answer comes back from a database that every player of Infinite Craft shares, so you get the same result anyone else would.

After a while you combine two things nobody has combined. This time the database has no entry, and the game asks a language model: a Llama model writes a name for the result, at most twenty tokens long, and picks an emoji. The game stores the answer, and your screen marks the new tile as a First Discovery. From then on every player who drops those two tiles together gets your result [1].

In Infinite Craft a model decides a rule of a shipped game, and it works because the decision is made once. The technique record, LLM-adjudicated state transitions, describes the cache as turning a stochastic model into a growing deterministic rule table, with cost bounded by the number of new pairs. It also helps that the mechanic has no right answer. When researchers asked GPT-4 to predict the next state of 31 text games across 76,369 transitions, it got the whole state right at most 59.9 percent of the time, and at most 49.7 percent for changes the environment caused on its own [2]. A combination in Infinite Craft only has to be plausible once and consistent afterwards. A model asked to run a city's water table would have to be right.

Seven ways to change a mechanic

Every technique in the dataset is linked to the mechanics it changes by an affects edge, and each edge records one of seven effects, each with a shipped or prototype example in the records:

• Replaces substrate: keeps the mechanic's role in play and swaps how it works, as when inZOI's small model reads a Zoi's needs, surroundings, and thoughts to pick its next action where scored need curves alone would.
• Deepens: more fidelity, scale, or variety, as when Gran Turismo Sophy drives the same tire and suspension model the player does at champion level.
• Enables new form: a version that was not possible before, as when a PUBG Ally teammate takes spoken orders mid-match.
• Cuts authoring cost: cheaper content, as when Genie 3 could render an injected event without per-event art.
• Explains: makes the mechanic's state or causes legible to the player.
• Tunes: automates balancing and testing, as agent playtesting did for The Sims Mobile's careers.
• Accelerates: the same output for less compute, as frame generation raises displayed frame rates in flight and driving sims.

How techniques change mechanics, by effect type.

Across 218 edges, deepening is the most common effect, at 74 edges, and explaining is the rarest, at 5. Only 3 techniques explain anything. Three of the explaining edges come from one prototype, language-model explanation of state, which turns an agent's thought log into a short account of why it acted, condenses a long chain of modifiers into a ranked list of causes, or annotates a graph with the events that moved it. 1 explaining edge comes from a technique that has shipped: generative agents keep their memories, plans, and reflections as text, so a player or designer can read them or ask the character why it did something.

The imbalance follows research effort. Generating content and replacing substrates attract most of the work, while explaining a running simulation to the person playing it attracts little, although the parts it needs (retrieval over a structured log, constrained generation, summarization) all exist. Chapter 5 described legibility as a cost of depth, and few techniques in the dataset aim at it.

The games show what explanation looks like when it is built by hand. RimWorld lists every thought behind a colonist's mood with its value and expiry. Paradox's grand strategy games break every derived number into nested tooltips of named modifiers (chapter 5). Dwarf Fortress keeps each dwarf's stress as a hidden number between minus and plus a million, and the player sees the tantrum without the account that produced it. A model that reads the simulation's own modifier lists and event log could give every game the Paradox treatment without the years of interface work, if it narrates what the simulation computed instead of guessing. The text-world results above are the reason for that condition: a model asked to predict dynamics gets them wrong about half the time, so the simulation has to compute the cause and the model has to report it. Chapter 10 returns to this as the highest-ranked opportunity in the dataset, a why-query over the simulation's log.

Where the set comes from

The techniques in the dataset are recent. 27 of the 45 are dated 2023 or later, and the largest category is generative world models, mostly the frame-predicting systems of chapter 7.

The set was built by looking for dated results that change a named mechanic, which probably favours fields that publish often and name their applications to games. Language-model agents, world models, and neural rendering publish constantly and point at games directly. Operations research on scheduling, or econometrics on markets, rarely mentions a game, so a method that could balance an economy enters only when someone has applied it to one. The pattern below reflects both the field and the search that built the set.

Counting exposure

To compare families of mechanics, the dataset sums each mechanic's edges into a technique exposure. Each edge counts for more when the technique has shipped than when it is a research result, and for more when it replaces or enables a mechanic than when it only tunes or speeds it up; the weights are written down in the rubrics file. A family's exposure is the sum over its mechanics. The measure says where the set of techniques points, weighted by how close each one is to a game. A mechanic with no edge means only that nobody in the set has published a dated result that names it. General-purpose models can write mod code or romance dialogue today, and the dataset leaves those uses out until someone ships or publishes one for a named mechanic. A single pull request that adds a technique can change it.

Figure: Where the techniques land. Each ribbon follows one technique's edges to one mechanic family, through the kind of change it makes. Darker ribbons come from techniques that have shipped in games. Hover or tap a technique, an effect, or a family to isolate its ribbons; switch the width to weight each edge by maturity and effect, the exposure measure of chapter 8. Families with no ribbon: operation & procedure. Interactive version: https://sim.yohei.me/#fig-technique-flows

In total, exposure concentrates in four families: narrative and generation, agents and behavior, control and automation, and physics and materials. These are the families where content is expensive to write (events, dialogue, characters) or behavior is expensive to compute (bodies, vehicles, crowds).

What the techniques would change

The edges into these families describe specific changes, and most of them are proposals rather than products.

A storyteller director picks incidents from an authored catalog. Triggered event chains, the machinery behind Crusader Kings III's stories, fire authored templates when conditions hold and schedule follow-ups days or months later, so a rival's illness can lead to a succession dispute and then to a war. A language-model game master would write the event for the exact state of the game, check it against the setting's rules, and narrate it. The trade is visible in both failure lists: authored chains repeat once players exhaust the pool, and surveyed language-model game masters accommodate requests a human game master would refuse, while language-model characters invent quests that do not exist in the game [3].

Many characters in the catalog pick their next action with a utility scorer: every candidate action gets a score from the character's needs, distances, personality, and context, and the game takes the best one or draws among the top few. The Sims picks the action whose object advertises the largest gain in happiness. Generative agents replace the scorer with a prompt that holds the current step of the day's plan, what the character perceives, and the memories retrieved for the moment; they replace numeric relationship values with remembered interactions; and they replace authored routines with plans each character writes for itself. The research route that keeps most of an existing engine is constrained decoding, which lets a language model choose only among the same enumerated actions the scorer would have ranked, so the game's actions, animations, and rules stay as they were.

In control, PUBG Ally takes spoken orders about movement, looting, and strategy during a match. Research agents that play commercial 3D games through the screen and a virtual keyboard and mouse follow language instructions and explain what they intend to do [4], and agents in the Voyager line write, test, and keep new routines in a growing library of skills [5].

In physics, the proposals that would replace the engine outright come from learned video prediction; the others, such as graph-network simulators and differentiable simulation, speed up, tune, or deepen existing solvers. Genie, WHAMM, and the open interactive video models render collisions, gravity, and car handling from what they learned watching footage, with no conservation laws and no stored geometry. DIAMOND's Counter-Strike model let the player jump again and again in mid-air, which the real game forbids [6]. The physics family's exposure is high partly because these models touch rigid bodies, vehicles, and destruction all at once, and every edge from them comes from a prototype or a research result.

Graph-network simulators learned fluids, rigid solids, and deformable materials as interacting particles and generalized to thousands of rollout steps and to at least ten times more particles than they were trained on [7]; MeshGraphNets ran cloth and aerodynamics one to two orders of magnitude faster than the simulator it learned from [8]; and the Fourier Neural Operator modelled turbulent flow up to three orders of magnitude faster than traditional solvers [9]. Those speedups are measured against the accurate solvers used for training, not against the cheap approximations games already ship, so the net gain can be much smaller, and nothing in the models guarantees that mass or energy is conserved. Differentiable simulators run from a target back to the parameters that produce it. GradABM made a million-agent epidemic model differentiable, simulated it in a few seconds on ordinary hardware, and calibrated it to COVID-19 and influenza data by gradient descent [10], and the dataset points the method at fitting tire and suspension constants to recorded telemetry and at telling a bridge builder which strut most changes a design's score.

Bodies are where agents meet physics. MaskedMimic trained one physics-based controller that generates motion from keyframes, text, VR tracking, paths, or objects, and sits on chairs, stools, and sofas [11]. In a game built on smart objects, a new piece of furniture could advertise an interaction and the body would perform it without a bespoke animation. Each character then needs a physics step and a policy inference every frame, which the record flags as costly at the hundreds of agents a colony or management sim runs.

Opponents draw on research from games that are not simulations. Poker solvers learned to play with hidden information: Libratus beat top professionals at heads-up no-limit hold'em, and Pluribus beat elite players at six-player hold'em, winning about 5 big blinds per 100 hands over 10,000 hands with a strategy computed in 8 days on a 64-core server for under 150 dollars of cloud compute [12, 13, 14]. DeepNash reached a top-three all-time ranking among expert Stratego players while bluffing with the hidden identities of its pieces [15]. The dataset points these solvers at fog of war, scouting, auctions, negotiation, and rival firms: an opponent that reasons over what it cannot see, scouts, and conceals, instead of cheating with full map knowledge or ignoring hidden units. Their guarantees hold for two-player zero-sum games, a continuous economy with many units has no obvious abstraction to solve, and Pluribus took about 20 seconds a hand on two CPUs, far above a per-frame budget.

Cicero added language. It played Diplomacy by pairing a dialogue model with a planner that predicts other players' moves from the board and the conversation, and over 40 games in an anonymous online league it averaged more than double the human score [16, 17]. Its edges in the dataset turn treaty menus into free-text negotiation with factions that propose joint moves, commit to them, and change plans when persuaded. A studio cannot ship the released model: its weights are licensed for non-commercial use, and the repository was archived in April 2025. Retraining is costly too; the scripts' example configuration uses 256 GPUs [18].

Self-play planners offer a second use: testing rules rather than playing them. AlphaZero, trained from scratch on nine variants of chess, was used to compare how decisive each variant was and what each piece was worth [19], and the dataset points the same comparison at tech trees and policy toggles, where converged self-play under each variant would flag research orders and laws that are always or never worth taking. MuZero learned its own model of the rules and matched AlphaZero at Go, chess, and shogi, with 16 TPUs for training and 1,000 more for self-play in its board-game runs [20, 21]. Each significant rule change needs retraining, and self-play converges on strong play, which can find exploits human players never would.

Weather forecasting supplies 3 of the environment family's 7 research edges. GraphCast produced ten-day global forecasts at 0.25 degrees in under a minute on one TPU machine and beat ECMWF's operational high-resolution forecast on over 90 percent of 1,380 verification targets [22], GenCast's 50-member ensembles beat ECMWF's ensemble on 97.2 percent of 1,320 targets [23], and ECMWF made its own machine-learned model operational in February 2025 at about a thousandth of the energy per forecast [24]. The dataset's edges would make a game's weather a coherent field with fronts and cyclones instead of a regional chain of weather states, give in-game forecasts calibrated uncertainty, and turn storms from dice rolls into threats with lead time. The models are trained on Earth's record, so a fictional planet or an altered coastline lies outside what they know, and a grid of about 28 km is coarse for a game map.

Two lines of artificial-life research grow behaviour from local rules. Neural cellular automata learn an update rule shared by every cell: a network of about 8,000 parameters grew an image from a single cell and regrew it after damage [25], 3D versions grew Minecraft castles, apartment blocks, and trees, some of over 3,000 blocks [26], and a 2025 method recovered Conway's Game of Life exactly as a circuit of 336 logic gates [27]. Lenia, a continuous cellular automaton, has a catalogue of over 400 species in 18 families [28], and a 2024 system used a foundation model to search the rules of Lenia and other artificial-life substrates for lifeforms nobody had seen [29]. The dataset points them at falling-sand materials, structures that heal after damage, and natural selection acting on creatures that emerge from the physics rather than from declared fitness stats. Each trained NCA rule targets one pattern and must be retrained to change it, and a Lenia creature is a field with no discrete identity, needs, or stats, so connecting it to a game's economy or goals needs another layer.

Divided by the number of mechanics in each family, the order changes. Narrative & generation still leads and world generation comes second, while agents and physics drop out of the top three: their totals partly reflected how many mechanics they hold.

The economic core

At the other end sit the families most simulation games are built from. Per mechanic, resources, logistics, economy, and population all fall in the lower half of the families, and resources and logistics sit near the bottom. Operation and procedure, the family of checklists, surface tools, and repair, has no technique attached at all.

Resources and production shows the pattern most plainly. It is the most used family in the catalog, and of its 14 mechanics, 12 have no technique in the set pointed at them. The two that do are hand crafting, which Infinite Craft's cached combinations turn into an open-ended recipe space, and machine boosting, where a research method searches module layouts for the dominant configurations that flatten a player's choices. Production chains, extraction, storage, spoilage, crops, and power have nothing.

Logistics fares only slightly better. Of its 14 mechanics, 9 have no technique, and every edge that reaches the family comes from research: mean-field games that would let drivers anticipate congestion instead of reacting to it, GPU simulation that points at agent-level traffic at much larger counts, and calibration methods that fit a traffic model to observed counts. Economy and finance has 19 mechanics, and 11 of them have no technique. The one shipped technique with an edge into the family is tree search, which can let rival firms choose prices and expansions by rolling the market forward.

Chapter 1 noted that most of the ten most used mechanics have no technique attached. Across the whole vocabulary, 99 mechanics are affected by no technique in the set, and they account for 52 percent of every game-to-mechanic use in the catalog.

The production chain's own record lists the problems a technique could address. "A single optimal ratio per product, once solved, turns play into copying calculator output": a design problem that search over recipes and layouts could expose before release. Upstream starvation that players blame on the wrong building is an explanation problem, which the record says is hard to diagnose without flow statistics; chapter 10 opens with it. Victoria 3 sets each good's price from the balance of buy and sell orders in each market, and its tooltips break a price into modifiers, yet the record for supply-demand pricing lists boom-and-bust cycles "that players read as random" among its failure modes.

Some of the neglect is reasonable. A production chain is cheap to compute and, by the rater's scores, about as easy to read as the average mechanic, so there is less for a learned model to fix. Economies still break in ways techniques could address: a price or loan loophole turns the economy into an exploit, recipe graphs have to be authored and balanced by hand, and scenarios have to be tested against strategies no designer tried. Those are tuning and authoring problems, and tuning methods exist. Agent-based playtesting has shipped in production pipelines, and quality-diversity search, which keeps the best design for each niche of a space, is an established research method.

Research that would tune the economy

Four research methods in the set have edges that tune parts of the economic core. In quality-diversity search, MAP-Elites with sliding boundaries mapped Hearthstone deck space by play style and found strong decks in each region [30], and a search through a level generator's latent space produced playable Super Mario Bros levels across designer-chosen counts of enemies and obstacles [31]. The dataset also aims it at competitor strategies, assembled designs, and rating formulas.

AlphaEvolve evolves programs: a language model proposes changes to code, an automatic evaluator scores them, and the best survive. It found a way to multiply 4 by 4 complex matrices with 48 scalar multiplications, and a scheduling heuristic it produced for Google's Borg cluster manager has recovered about 0.7 percent of the company's fleet compute in production for over a year [32, 33]. Assigning hauling jobs in a colony is a scheduling problem of the Borg kind, and ordering loads and stops on a delivery route resembles the bin-packing heuristics an earlier system improved, so the dataset points it there and at the scoring functions of utility AI, where an evolved function stays readable code. It needs a fast and trustworthy evaluator, so it cannot tune anything judged by taste.

Simulation-based inference runs from data to parameters. Given time series, it returns a posterior over the parameters of an agent-based model that could have produced them, and neural estimators did this for economic agent-based models where older likelihood-free methods needed hundreds of thousands of simulations [34]. Besides traffic, its edges would fit a market's elasticities to real prices, a match engine to a real league's statistics, and threat scaling to the loss rates seen in playtests, each with a measure of which parameters the data cannot pin down.

Models of human behaviour are the fourth. Agents built from two-hour interviews with 1,052 American adults answered the General Social Survey about 85 percent as accurately as the same people answered it again two weeks later [35, 36], and Centaur, a model trained on human behavioural experiments, predicted held-out participants better than 14 cognitive and statistical models [37]. Critics found that Centaur still did well with its task instructions removed or replaced by misleading ones, which they read as reliance on statistical cues [38, 39]. The dataset points these models at customer choice, crowd preferences, electorates, and interest groups, where games now use formulas and distributions a designer sets by hand.

The most exposed mechanics

Mechanics most affected by the technique set.

Generative content and procedural terrain lead the ranking, because many techniques generate content or worlds. Utility action selection, the scoring that picks a character's next action from its needs, comes third: language models, planners, and learned policies all offer to replace it, and inZOI's Smart Zois are the shipped case, a small local model helping choose among the game's own actions. The scorer these techniques would replace runs the characters of 18 catalog games, from Theme Park in 1994 through The Sims and Theme Hospital to RimWorld and Prison Architect. Rigid-body physics and direct avatar control follow, exposed mainly to frame-generating world models such as Genie, which replace both with learned prediction and carry over the consistency limits chapter 7 described.

World generation shows the same split inside one family. Procedural terrain is affected by 10 techniques: satellite reconstruction rebuilt the planet for Microsoft Flight Simulator, learned generators trained on example maps would replace hand-written noise and placement rules, and world models generate a scene from a prompt and forget it a minute later. Simulated history, the mechanic behind Dwarf Fortress's world generation, has none. Before the fortress is founded, Dwarf Fortress builds elevation, rainfall, drainage, temperature, and vegetation, then simulates civilizations, sites, and historical figures for as many years as the player chooses, so the fortress starts inside a world with a past. It is used by 2 catalog games and judged highly open to AI, and no technique in the set touches it.

Generative content is used by 7 catalog games, almost all of them AI-native titles, and is affected by 13 techniques.

Judgment against evidence

Every mechanic also carries a judged score for AI leverage: how much, in the rater's reading, current techniques could change it. AI leverage scores and exposure correlate at ρ = 0.53. That number is a consistency check. The same model rater scored AI leverage, wrote the affects edges, and judged each technique's maturity, so a correlation between them shows the rater was consistent with itself. A second, calibrated rater is the remedy, and chapter 13 describes how one is being added.

The places where the two disagree point at specific mechanics. The table below lists mechanics judged highly open to AI that no technique in the set touches, the largest gaps between judged AI leverage and exposure first.

Mechanics where rater-judged AI leverage outruns the techniques currently linked to them (normalized). High gap = under-researched opportunity or missing technique links. Equal gaps are ordered by prevalence, so more-used mechanics come first.

Mod support tops the list because it is the most widely used of them, at 55 games. Romance and family systems, blueprints, creature care, product design, diagnosis, simulated history, and rule-checking inspection are each judged ready for substantial change and have no cited, game-specific technique. Some are covered implicitly by general methods, such as code generation for mods or language models for romance dialogue, but the dataset admits a technique only as a dated result that changes a named mechanic, and none has been recorded.

Exercise: aim a technique

Pick one mechanic from the table of gaps above.

1. Write down the state it owns and the rule that updates it, from its record.
2. Choose one effect from the seven: replace, deepen, enable, cut authoring cost, explain, tune, or accelerate. Pick the one that would matter most to a player.
3. Name the smallest technique that could do it, and say whether its output would enter authoritative state or stay in presentation.
4. Describe the experiment that would show it worked, with a number a stranger could check.

Step 3 decides most of the risk. Chapter 9 turns from techniques to the mechanics themselves and looks for pairs of common parts that almost never appear in the same game.

Sources

[1] Wikipedia contributors (2026). Infinite Craft. Wikipedia https://en.wikipedia.org/wiki/Infinite_Craft
[2] Ruoyao Wang, Graham Todd, Ziang Xiao, Xingdi Yuan, Marc-Alexandre Côté, Peter Clark et al. (2024-06-10). Can Language Models Serve as Text-Based World Simulators?. arXiv https://arxiv.org/abs/2406.06485
[3] Roberto Gallotta, Graham Todd, Marvin Zammit, Sam Earle, Antonios Liapis, Julian Togelius et al. (2024). Large Language Models and Games: A Survey and Roadmap. IEEE Transactions on Games https://arxiv.org/abs/2402.18659
[4] Google DeepMind SIMA Team (2025-11-13). SIMA 2: An Agent that Plays, Reasons, and Learns With You in Virtual 3D Worlds. Google DeepMind blog https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/
[5] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu et al. (2023-05). Voyager: An Open-Ended Embodied Agent with Large Language Models. Transactions on Machine Learning Research (2024) https://arxiv.org/abs/2305.16291
[6] Eloi Alonso, Adam Jelley, Vincent Micheli, Anssi Kanervisto, Amos Storkey, Tim Pearce et al. (2024-05). Diffusion for World Modeling: Visual Details Matter in Atari. NeurIPS 2024 https://diamond-wm.github.io/
[7] Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, Peter W. Battaglia (2020-02-21). Learning to Simulate Complex Physics with Graph Networks. ICML 2020 https://arxiv.org/abs/2002.09405
[8] Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, Peter W. Battaglia (2020-10-07). Learning Mesh-Based Simulation with Graph Networks. ICLR 2021 https://arxiv.org/abs/2010.03409
[9] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart et al. (2020-10-18). Fourier Neural Operator for Parametric Partial Differential Equations. ICLR 2021 https://arxiv.org/abs/2010.08895
[10] Ayush Chopra, Alexander Rodriguez, Jayakumar Subramanian, Arnau Quera-Bofarull, Balaji Krishnamurthy, B. Aditya Prakash et al. (2022-07-20). Differentiable Agent-based Epidemiology. AAMAS 2023 https://arxiv.org/abs/2207.09714
[11] Chen Tessler, Yunrong Guo, Ofir Nabati, Gal Chechik, Xue Bin Peng (2024-09). MaskedMimic: Unified Physics-Based Character Control Through Masked Motion Inpainting. ACM Transactions on Graphics (SIGGRAPH Asia 2024) https://research.nvidia.com/labs/par/projects/maskedmimic/
[12] Noam Brown, Tuomas Sandholm (2018-01-26). Superhuman AI for heads-up no-limit poker: Libratus beats top professionals. Science 359:418-424 https://doi.org/10.1126/science.aao1733
[13] Noam Brown, Tuomas Sandholm (2019-07-11). Superhuman AI for multiplayer poker. Science 365:885-890 https://doi.org/10.1126/science.aay2400
[14] Facebook AI, Carnegie Mellon University (2019-07-11). Pluribus: First AI to beat pros in 6-player poker. Facebook AI blog https://ai.facebook.com/blog/pluribus-first-ai-to-beat-pros-in-6-player-poker/
[15] Julien Perolat, et al. (2022-12-02). Mastering the game of Stratego with model-free multiagent reinforcement learning. Science 378:990-996 https://doi.org/10.1126/science.add4679
[16] Meta Fundamental AI Research Diplomacy Team (FAIR) (2022-11-22). Human-level play in the game of Diplomacy by combining language models with strategic reasoning. Science 378(6624):1067-1074 https://doi.org/10.1126/science.ade9097
[17] Meta AI (2022). CICERO. Meta AI research project page https://ai.meta.com/research/cicero/
[18] Meta Platforms (facebookresearch) (2022). diplomacy_cicero. GitHub https://github.com/facebookresearch/diplomacy_cicero
[19] Nenad Tomašev, Ulrich Paquet, Demis Hassabis, Vladimir Kramnik (2020-09-09). Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess. arXiv https://arxiv.org/abs/2009.04374
[20] Julian Schrittwieser, Ioannis Antonoglou, Thomas Hubert, Karen Simonyan, Laurent Sifre, Simon Schmitt et al. (2020-12-23). Mastering Atari, Go, chess and shogi by planning with a learned model. Nature 588:604-609 https://doi.org/10.1038/s41586-020-03051-4
[21] Wikipedia contributors (2026). MuZero. Wikipedia https://en.wikipedia.org/wiki/MuZero
[22] Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, et al., Peter Battaglia (2023-11-14). GraphCast: Learning skillful medium-range global weather forecasting. Science 382 (2023) https://arxiv.org/abs/2212.12794
[23] Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, et al., Matthew Willson (2024-12-04). Probabilistic weather forecasting with machine learning. Nature 637, 84-90 https://www.nature.com/articles/s41586-024-08252-9
[24] ECMWF (2025-02-25). ECMWF's AI forecasts become operational. ECMWF news https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational
[25] Alexander Mordvintsev, Ettore Randazzo, Eyvind Niklasson, Michael Levin (2020-02-11). Growing Neural Cellular Automata. Distill https://distill.pub/2020/growing-ca/
[26] Shyam Sudhakaran, Djordje Grbic, Siyan Li, Adam Katona, Elias Najarro, Claire Glanois et al. (2021-03-15). Growing 3D Artefacts and Functional Machines with Neural Cellular Automata. ALIFE 2021 https://arxiv.org/abs/2103.08737
[27] Pietro Miotti, Eyvind Niklasson, Ettore Randazzo, Alexander Mordvintsev (2025-03-03). Differentiable Logic Cellular Automata: From Game of Life to Pattern Generation. Google Paradigms of Intelligence https://google-research.github.io/self-organising-systems/difflogic-ca/
[28] Bert Wang-Chak Chan (2018-12-13). Lenia: Biology of Artificial Life. Complex Systems 28(3) https://arxiv.org/abs/1812.05433
[29] Akarsh Kumar, Chris Lu, Louis Kirsch, Yujin Tang, Kenneth O. Stanley, Phillip Isola et al. (2024-12-23). Automating the Search for Artificial Life with Foundation Models. https://arxiv.org/abs/2412.17799
[30] Matthew C. Fontaine, Scott Lee, L. B. Soros, Fernando De Mesentier Silva, Julian Togelius, Amy K. Hoover (2019-04-24). Mapping Hearthstone Deck Spaces through MAP-Elites with Sliding Boundaries. arXiv https://arxiv.org/abs/1904.10656
[31] Matthew C. Fontaine, Ruilin Liu, Ahmed Khalifa, Jignesh Modi, Julian Togelius, Amy K. Hoover et al. (2020-07-11). Illuminating Mario Scenes in the Latent Space of a Generative Adversarial Network. arXiv https://arxiv.org/abs/2007.05674
[32] Google DeepMind (2025-05-14). AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms. Google DeepMind blog https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/
[33] Alexander Novikov, Ngan Vu, Marvin Eisenberger, et al., Matej Balog (2025-06-16). AlphaEvolve: A coding agent for scientific and algorithmic discovery. https://arxiv.org/abs/2506.13131
[34] Joel Dyer, Patrick Cannon, J. Doyne Farmer, Sebastian Schmon (2024). Black-box Bayesian inference for economic agent-based models. Journal of Economic Dynamics and Control https://arxiv.org/abs/2202.00625
[35] Joon Sung Park, Carolyn Q. Zou, Jonne Kamphorst, Niles Egan, Aaron Shaw, Benjamin Mako Hill et al. (2024-11-15). Generative Agent Simulations of 1,000 People. arXiv https://arxiv.org/abs/2411.10109
[36] Joon Sung Park, Carolyn Q. Zou, Aaron Shaw, Benjamin Mako Hill, Carrie J. Cai, Meredith Ringel Morris et al. (2025-05-20). Simulating Human Behavior with AI Agents. Stanford HAI policy brief https://hai.stanford.edu/policy/simulating-human-behavior-with-ai-agents
[37] Marcel Binz, Elif Akata, Matthias Bethge, et al., Eric Schulz (2025-07-02). A foundation model to predict and capture human cognition. Nature 644, 1002-1009 https://www.nature.com/articles/s41586-025-09215-4
[38] Wei Liu, Nai Ding (2025-12-11). Can Centaur truly simulate human cognition? The fundamental limitation of instruction understanding. National Science Open 5(1) https://www.nso-journal.org/10.1360/nso/20250053
[39] Mark Orr, Drew Cranford, Ken Ford, Kevin Gluck, Will Hancock, Christian Lebiere et al. (2025-09-30). Not Even Wrong: On the Limits of Prediction as Explanation in Cognitive Science. arXiv https://arxiv.org/abs/2510.03311

Dataset dfde7895aa6d.

mechanics in this chapter

65 linked or named

AdvisorsAgent pathfindingAuctionsBlueprintsCare-dependent developmentCreature careCreature learningCrowd preferencesDaily schedulesDecoration and beautyDelegation to AI managersDiagnosis and treatmentDilemma cardsDirect avatar controlElectionsEmbodied skill controlFaction diplomacyFalling-sand materialsFluid and gas simulationFog of warForecastingGenerative content on demandHand craftingHauling jobsHistory chronicleInterest groupsMachine boostingMaterial interactionsMemories and thoughtsMilitary operationsModding APIModular assemblyNatural selectionNatural-language interactionNegotiationPlayer legislationPlayer programmingPolicy influence networkPolicy togglesProcedural build assistProcedural charactersProcedural terrainProduct developmentProduction chainsProgramming by demonstrationReal-world data feedsRelationship valuesRigid-body physicsRival firmsRomance and familyRule-checking inspectionScenario objectivesSimulated history generationSmart objectsSocial exchangesSoft-body deformationSpoilageStoryteller directorSupply-demand pricingTech treeTriggered event chainsUnit commandsUtility action selectionVehicle dynamicsWeather