---
title: "Minds, brains and artificial worlds · Simulations · Simulation Hypothesis"
url: https://simulation-hypothesis.com/simulations/minds-computation
description: "Neurons, brains, artificial life, and the simulated worlds in which machines learn. Part of a taxonomy of the simulations people actually run."
license: https://creativecommons.org/licenses/by/4.0/
credit: "Ross Allan Dahl, Simulation Hypothesis (simulation-hypothesis.com)"
---

[Simulations](https://simulation-hypothesis.com/simulations)domain 11 of 18

# Minds, brains and artificial worlds

Neurons, brains, artificial life, and the simulated worlds in which machines learn.

Simulations of brains, and of worlds built for artificial agents to live in and learn from. In one limited way, the simulation hypothesis is already real here: we've built simulated worlds with learning agents inside them.

## Brains

### Detailed neuron models

A single neuron as a branching electrical cable with ion channels, following Hodgkin and Huxley.

Question it answers: How does this neuron’s shape change what it computes?

Leaves out: One neuron's fine chemistry is still simplified into a handful of currents.

Runs on: about 10^12 FLOPS rough a laptop, real time for one neuron Since: 1952 Huxley integrated the Hodgkin–Huxley equations by hand on a mechanical calculator, about three weeks per action potential; computers took over later [source](https://www.physoc.org/about-us/history-archives/historicalhighlights/hodgkin-and-huxleys-modelling-of-the-action-potential/) In the simulations tree: [Hodgkin–Huxley by hand](https://simulation-hypothesis.com/trees/simulations?n=min-hh)

[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Cells

### Large-scale brain circuits

Millions of spiking neurons wired into a circuit resembling a patch of cortex.

Question it answers: Can we reproduce brain rhythms and responses from wiring alone?

Leaves out: Wiring is often statistical, not the animal's actual measured connections.

Runs on: about 10^15 FLOPS rough a university cluster, hours for a cortical patch Since: 2006 by 2006, early large-scale simulations of a cortical column In the simulations tree: [Neocortical microcircuit](https://simulation-hypothesis.com/trees/simulations?n=min-markram)

[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)floor: Cells

### Whole-brain network models

Brain regions as nodes connected by measured wiring, fitted to scans.

Question it answers: How does damage in one region change activity across the brain?

Leaves out: Regions are treated as single points, missing all detail inside them.

Runs on: about 10^14 FLOPS rough a university cluster, minutes per simulated scenario Since: 2008 by 2008, early whole-brain network models fitted to diffusion MRI wiring

Models: [Human](https://simulation-hypothesis.com/entities/human)

[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Populations

### Simulated small organisms

A tiny animal’s complete nervous system, such as a roundworm’s 302 neurons, connected to a simulated body in a virtual dish.

Question it answers: Is the wiring diagram enough to produce the behaviour?

Leaves out: The wiring diagram alone still does not fully explain the animal's behaviour.

Runs on: about 10^12 FLOPS rough a laptop, real time for the whole nervous system Since: 1997 by 1997, early simulations built on the roundworm's 1986 connectome map In the simulations tree: [OpenWorm](https://simulation-hypothesis.com/trees/simulations?n=min-openworm)

[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)[Real time](https://simulation-hypothesis.com/simulations/methods#m-real-time)floor: Cells

### Neuromorphic hardware simulation

Chips built with circuits that behave like neurons and synapses directly in hardware, instead of calculating the equations step by step in software.

Question it answers: Can neuromorphic hardware compute the same things a brain does, using far less power?

Leaves out: Real neurons have far richer chemistry than a circuit built to mimic their spikes.

Runs on: about 10^13 FLOPS rough a neuromorphic research chip, real time Since: 1989 Carver Mead's early silicon neuron and retina circuits

[Real time](https://simulation-hypothesis.com/simulations/methods#m-real-time)[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Cells

### Brain disease and seizure spread

Abnormal electrical activity or misfolded proteins spreading through a simulated brain network, as in epilepsy or dementia.

Question it answers: Where does a seizure start, and how far will it spread?

Leaves out: Every real brain's disease path is different; the model shows a pattern, not a prediction for one patient.

Runs on: about 10^13 FLOPS rough a hospital or university cluster, minutes per patient model Since: 1997 by 1997, early network models of seizure spread through the brain

[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Populations

### Connectome-constrained whole-brain models

A neural simulation uses the full measured wiring of an insect brain with simple neuron models to predict how stimuli produce behaviour.

Question it answers: Does wiring alone predict how the brain turns senses into action?

Leaves out: Synapse strengths and neuron dynamics are assumed from simple rules.

Runs on: about 10^12 FLOPS rough Spiking network of about 130,000 neurons Since: 2024 A Drosophila computational brain model reveals sensorimotor processing (2024) [source](https://doi.org/10.1038/s41586-024-07763-9)

[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Cells

### Neural mass and field models

A population of neurons reduced to one or two average activity variables, giving the rhythms and waves of whole regions.

Question it answers: What collective rhythms can excitation and inhibition produce between populations?

Leaves out: Individual neurons vanish into an average.

Runs on: about 10^9 FLOPS rough a laptop, seconds Since: 1972 by 1972, Wilson and Cowan, Excitatory and inhibitory interactions in localized populations of model neurons [source](https://doi.org/10.1016/s0006-3495%2872%2986068-5)

[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)[Continuum fields](https://simulation-hypothesis.com/simulations/methods#m-continuum)floor: Cells

### Functional spiking brain models

A network of spiking model neurons built to carry out cognitive tasks such as recall, counting and copying a drawn digit.

Question it answers: Can a model made of neurons carry out tasks a brain handles?

Leaves out: Its task set is small and wiring is designed, not measured.

Runs on: about 10^15 FLOPS rough a university cluster, hours Since: 2012 by 2012, Eliasmith and others, A large-scale model of the functioning brain (Spaun) [source](https://doi.org/10.1126/science.1225266)

[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Cells

### Attractor memory networks

A network of simple units whose connections store patterns as stable states, so a partial cue settles into the full memory.

Question it answers: How can a network recall a whole memory from a fragment?

Leaves out: Capacity is small and real neurons are far more complex.

Runs on: about 10^8 FLOPS rough a laptop, seconds Since: 1982 by 1982, Hopfield, Neural networks and physical systems with emergent collective computational abilities [source](https://en.wikipedia.org/wiki/Hopfield%5Fnetwork)

[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Cells

### Connectome-based insect brain models

A whole fly brain simulated from its measured wiring diagram, with each neuron and synapse taken from the reconstruction.

Question it answers: Which behaviours follow from wiring alone, once every connection is known?

Leaves out: Synapse strengths and neuron chemistry are inferred, not measured.

Runs on: about 10^14 FLOPS rough a workstation, minutes per simulated second Since: 2024 by 2024, Dorkenwald and others, the adult fly brain wiring diagram, and Shiu and others, a computational model built on it [source](https://doi.org/10.1038/s41586-024-07763-9)

[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Cells

## Artificial life and complexity

### Cellular automata

Grids of cells following simple rules, such as John Conway’s Life, which grow gliders, replicators and even working computers.

Question it answers: How much complexity can grow from how simple a rule?

Leaves out: Simple rules rarely map cleanly onto real physical or biological systems.

Runs on: about 10^12 FLOPS rough a laptop, huge grids in seconds Since: 1970 John Conway's Game of Life In the simulations tree: [Life, played by hand](https://simulation-hypothesis.com/trees/simulations?n=life-life)

[Run the interactive model](https://simulation-hypothesis.com/simulations/run#toy-life)

[Cellular automata](https://simulation-hypothesis.com/simulations/methods#m-cellular-automata)floor: Rules

### Digital evolution

Self-copying computer programs mutating and competing for processor time in a simulated machine.

Question it answers: Can evolution invent complex features we can watch step by step?

Leaves out: Programs evolve in an artificial computer world, not real biochemistry.

Runs on: about 10^12 FLOPS rough a desktop, millions of generations overnight Since: 1953 Barricelli’s numerical evolution experiments on the Institute for Advanced Study computer [source](https://www.ias.edu/ideas/2012/george-dyson-ecp)

[Evolutionary](https://simulation-hypothesis.com/simulations/methods#m-evolutionary)floor: Rules

### Evolved virtual creatures

Bodies and brains evolved together in a physics simulation, learning to swim, walk or jump.

Question it answers: What bodies and behaviours does evolution invent when we are not steering it?

Leaves out: Evolved gaits often look very different from biological ones, because the simulated physics differs subtly from the real world.

Runs on: about 10^13 FLOPS rough a desktop or small cluster, hours to evolve a population Since: 1994 Karl Sims evolved creatures that swim and walk in a physics simulation

[Evolutionary](https://simulation-hypothesis.com/simulations/methods#m-evolutionary)[Real time](https://simulation-hypothesis.com/simulations/methods#m-real-time)floor: Bodies

### Self-organised criticality

Grains added one at a time to a sandpile until avalanches of every size appear.

Question it answers: Why do earthquakes, fires and extinctions follow the same statistics?

Leaves out: Real earthquakes and fires have specific physics a sandpile only caricatures.

Runs on: about 10^11 FLOPS rough a laptop, seconds for a large sandpile Since: 1987 Bak, Tang and Wiesenfeld's sandpile model

[Run the interactive model](https://simulation-hypothesis.com/simulations/run#toy-sandpile)

[Cellular automata](https://simulation-hypothesis.com/simulations/methods#m-cellular-automata)floor: Bodies

### Swarm intelligence algorithms

Simple virtual agents that leave trails or copy their neighbours, borrowed from ants and birds to solve abstract problems like routing and scheduling.

Question it answers: Can simple copying and trail-following solve a hard optimisation problem?

Leaves out: Works well on the well-posed problems it was tested on. Real, irregular problems need more tuning.

Runs on: about 10^11 FLOPS rough a laptop, seconds to minutes per run Since: 1992 Marco Dorigo's ant colony optimisation algorithm

[Run the interactive model](https://simulation-hypothesis.com/simulations/run#toy-flock)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Evolutionary](https://simulation-hypothesis.com/simulations/methods#m-evolutionary)floor: Rules

### Random Boolean networks

Switches wired together at random, each turning on or off based on a handful of others, studied for when the whole network settles down or churns forever.

Question it answers: How much wiring does it take before a network of switches turns chaotic?

Leaves out: Random wiring is a crude stand-in for the carefully evolved wiring of a real cell or brain.

Runs on: about 10^11 FLOPS rough a laptop, seconds per network Since: 1969 Stuart Kauffman's random networks of genetic switches

[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)[Cellular automata](https://simulation-hypothesis.com/simulations/methods#m-cellular-automata)floor: Rules

### Continuous cellular automata

Cellular automata whose cell states, space and time are continuous, producing smooth self-organising creatures.

Question it answers: What life-like patterns appear in a continuous cellular automaton?

Leaves out: The rules are artificial and do not model real chemistry.

Runs on: about 10^11 FLOPS rough GPU grid updates Since: 2018 Lenia: biology of artificial life (2018) [source](https://arxiv.org/abs/1812.05433)

[Cellular automata](https://simulation-hypothesis.com/simulations/methods#m-cellular-automata)floor: Rules

### Neural cellular automata

Each cell holds a small neural network that learns the local rule so a grid grows and repairs a target pattern.

Question it answers: Can a local learned rule make a global pattern grow and self-repair?

Leaves out: Target patterns are small, and rules are trained, not discovered.

Runs on: about 10^11 FLOPS rough Small differentiable training Since: 2020 Growing Neural Cellular Automata (2020) [source](https://distill.pub/2020/growing-ca/)

[Cellular automata](https://simulation-hypothesis.com/simulations/methods#m-cellular-automata)[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Rules

### Self-replicating program soups

A pool of random programs interacts by chance until self-copying ones emerge from the mix.

Question it answers: Can replicators appear without being designed in?

Leaves out: Language choice sets the result, and it is far from chemistry.

Runs on: about 10^10 FLOPS rough Millions of tiny program interactions Since: 2024 Computational Life: how well-formed, self-replicating programs emerge from simple interaction (2024) [source](https://arxiv.org/abs/2406.19108)

[Evolutionary](https://simulation-hypothesis.com/simulations/methods#m-evolutionary)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Rules

### Artificial chemistry

A pool of abstract molecules that react by rules, so self-maintaining structures can appear without any real chemistry.

Question it answers: Can self-sustaining chemical organisation arise from simple reaction rules?

Leaves out: The molecules are symbols; real bond energies and geometry are absent.

Runs on: about 10^9 FLOPS rough a laptop, hours Since: 1991 by 1991, Fontana's AlChemy, presented at the Second Artificial Life Conference [source](https://en.wikipedia.org/wiki/Artificial%5Fchemistry)

[Rules and states](https://simulation-hypothesis.com/simulations/methods#m-rules)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Molecules

### Evolutionary robotics

Controllers or bodies of simulated robots are evolved over many generations by selecting for task performance.

Question it answers: Can a robot's brain and body be bred instead of designed?

Leaves out: Controllers evolved in simulation often fail on real hardware.

Runs on: about 10^14 FLOPS rough a university cluster, days Since: 1993 by 1993, Cliff, Husbands and Harvey, Evolving visually guided robots [source](https://doi.org/10.7551/mitpress/3116.003.0051)

[Evolutionary](https://simulation-hypothesis.com/simulations/methods#m-evolutionary)[Real time](https://simulation-hypothesis.com/simulations/methods#m-real-time)[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Bodies

### Continuous artificial life

A cellular automaton with continuous states, space and time that produces moving, self-organising lifelike patterns.

Question it answers: What lifelike forms appear when the rules are smooth rather than discrete?

Leaves out: The patterns resemble organisms but follow no real biology.

Runs on: about 10^11 FLOPS rough a gaming PC, real time Since: 2018 by 2018, Chan, Lenia: Biology of Artificial Life (arXiv 1812.05433) [source](https://arxiv.org/abs/1812.05433)

[Cellular automata](https://simulation-hypothesis.com/simulations/methods#m-cellular-automata)[Continuum fields](https://simulation-hypothesis.com/simulations/methods#m-continuum)floor: Rules

### Artificial immune models

Populations of interacting antibody-like agents that learn to recognise patterns by selection and network effects.

Question it answers: How does a recognition system adapt using only local interactions?

Leaves out: It borrows ideas from immunology for computing, not a calibrated model of a real immune response.

Runs on: about 10^8 FLOPS rough a laptop, hours Since: 1986 by 1986, Farmer, Packard and Perelson, The immune system, adaptation, and machine learning [source](https://doi.org/10.1016/0167-2789%2886%2990240-x)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)[Evolutionary](https://simulation-hypothesis.com/simulations/methods#m-evolutionary)floor: Cells

### Core War battles

Programs written in a tiny assembly language fight for control of a shared simulated memory, overwriting each other.

Question it answers: Which self-copying program survives when all of memory is a battlefield?

Leaves out: The machine is an invented instruction set, so lessons do not map to real malware.

Runs on: about 10^7 FLOPS rough a laptop, seconds per battle Since: 1984 by 1984, Core War by Jones and Dewdney, in Scientific American [source](https://en.wikipedia.org/wiki/Core%5FWar)

[Rules and states](https://simulation-hypothesis.com/simulations/methods#m-rules)[Discrete event](https://simulation-hypothesis.com/simulations/methods#m-discrete-event)[Evolutionary](https://simulation-hypothesis.com/simulations/methods#m-evolutionary)floor: Rules

## Worlds for artificial minds

### Reinforcement learning environments

Simulated worlds where an artificial agent acts, is rewarded and improves over millions of attempts.

Question it answers: Can an agent learn this skill from trial and error alone?

Leaves out: Skills learned in a simplified world do not always transfer to the complexity of reality.

Runs on: about 10^15 FLOPS rough a cluster of GPUs, days of simulated experience Since: 1992 by 1992, TD-Gammon learned backgammon strategy through self-play

[Real time](https://simulation-hypothesis.com/simulations/methods#m-real-time)[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Bodies

### Multi-agent AI societies

Many AI agents sharing a world, with their own goals, memories and conversations.

Question it answers: What social behaviour appears when many artificial agents live together?

Leaves out: Behaviours that look social can be narrow tricks tuned to the training setup.

Runs on: about 10^15 FLOPS rough a cluster of GPUs, running many agents in parallel Since: 2017 by 2017, early experiments with many learning agents sharing one world

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Bodies

### Synthetic training data

Photorealistic scenes rendered in endless variations to train vision systems.

Question it answers: Can a system trained only on rendered images see the real world?

Leaves out: Rendered scenes are cleaner than reality, so trained systems can be misled by real-world noise.

Runs on: about 10^14 FLOPS rough a render farm or GPU cluster, hours for a dataset Since: 2016 by 2016, early rendered datasets train image recognition systems

[Real time](https://simulation-hypothesis.com/simulations/methods#m-real-time)floor: Bodies

### Learned world models

A neural network that predicts what happens next in a world, well enough to generate playable scenes of it.

Question it answers: Can a network learn to simulate a world just by watching it?

Leaves out: Long-term predictions drift away from anything physically consistent.

Runs on: about 10^15 FLOPS rough a large GPU cluster, days of training Since: 2018 Ha and Schmidhuber trained a network to predict and dream game-like scenes

[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Fields

### Self-play strategy training

An artificial agent playing repeated matches against copies of itself in an abstract strategy game, improving with every round.

Question it answers: How strong can a strategy get with no teacher but its own past self?

Leaves out: Strategies tuned against a copy of the same agent can have blind spots that a different opponent would exploit.

Runs on: about 10^16 FLOPS rough a large GPU cluster, days of self-play Since: 1959 Arthur Samuel's checkers program learned by playing itself

[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)[Game theory](https://simulation-hypothesis.com/simulations/methods#m-game-theory)floor: Rules

### AI safety red-teaming environments

A simulated setting built to tempt an artificial agent into unwanted behaviour, so the failure can be found before deployment.

Question it answers: Under what conditions does this system exhibit behaviour that it should not?

Leaves out: Only tests for the failures the red team thought to look for.

Runs on: about 10^14 FLOPS rough a research team's GPU server, hours per test suite Since: 2016 by 2016, early adversarial testing environments for reinforcement learning agents

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Rules

### Embodied household simulators

3D virtual homes where an agent sees, moves and manipulates objects, used to train and test embodied AI.

Question it answers: Can an agent learn tasks in a virtual home that transfer to real ones?

Leaves out: Object physics and scene variety are limited.

Runs on: about 10^12 FLOPS rough Real-time rendering, many parallel instances Since: 2017 AI2-THOR: an interactive 3D environment for visual AI (2017) [source](https://arxiv.org/abs/1712.05474)

[Real time](https://simulation-hypothesis.com/simulations/methods#m-real-time)[Rules and states](https://simulation-hypothesis.com/simulations/methods#m-rules)floor: Bodies

### Open-ended environment generation

Training worlds are created and mutated automatically as agents learn so challenges keep growing in difficulty and variety.

Question it answers: Can an environment generator keep producing useful new challenges?

Leaves out: The search is limited by what the environment family can express.

Runs on: about 10^15 FLOPS rough Population of environments and agents Since: 2019 Paired Open-Ended Trailblazer (POET) (2019) [source](https://arxiv.org/abs/1901.01753)

[Evolutionary](https://simulation-hypothesis.com/simulations/methods#m-evolutionary)[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Rules

### Simulated web environments for agents

Self-hosted copies of real websites where a language-model agent completes tasks, scored by checking the final state.

Question it answers: Can an agent complete multi-step tasks on realistic websites?

Leaves out: Sites are static copies, and task variety is limited.

Runs on: about 10^13 FLOPS rough Environment resets and many model calls Since: 2023 WebArena (2023) [source](https://arxiv.org/abs/2307.13854)

[Rules and states](https://simulation-hypothesis.com/simulations/methods#m-rules)[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Rules

### Recommender-system simulators

A configurable simulated user population and item catalogue used to train and test recommendation algorithms.

Question it answers: How would a recommendation algorithm behave over many interactions with users?

Leaves out: User behaviour is assumed, not measured.

Runs on: about 10^10 FLOPS rough Cheap parallel runs Since: 2019 RecSim: a configurable simulation platform for recommender systems (2019) [source](https://arxiv.org/abs/1909.04847)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Rules

### Neural game engines

A diffusion model trained on gameplay recordings generates the next frame from the player's actions, replacing the game's own engine.

Question it answers: Can a neural network run an interactive game at playable speed?

Leaves out: Memory is short, and details drift over long play.

Runs on: about 10^15 FLOPS rough Heavy training; real-time inference on one accelerator Since: 2024 Diffusion models are real-time game engines (2024) [source](https://arxiv.org/abs/2408.14837)

[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)[Real time](https://simulation-hypothesis.com/simulations/methods#m-real-time)floor: Rules

### Interactive generative worlds

A model learns from unlabelled video to create controllable worlds in which a user can act frame by frame.

Question it answers: Can a playable world be generated from images alone?

Leaves out: Consistency and frame rate are limited, and physics are not guaranteed.

Runs on: about 10^17 FLOPS rough Large-scale video training Since: 2024 Genie: generative interactive environments (2024) [source](https://arxiv.org/abs/2402.15391)

[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Rules

### Learned real-world interactive simulators

A video model conditioned on actions predicts what a robot or agent would see after each command, replacing a hand-built simulator.

Question it answers: Can a learned model simulate the real-world results of actions?

Leaves out: Long rollouts hallucinate, and physical accuracy is not verified.

Runs on: about 10^17 FLOPS rough Training on large video and action datasets Since: 2023 Learning interactive real-world simulators (2023) [source](https://arxiv.org/abs/2310.06114)

[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Rules

### Generative driving world models

A video model conditioned on speed and steering generates plausible future scenes from a car's view for training and testing autonomy.

Question it answers: What traffic scenes could follow from a given driving action?

Leaves out: Generated scenes may violate physics or rare-case behaviour.

Runs on: about 10^17 FLOPS rough Large video model training Since: 2023 GAIA-1: a generative world model for autonomous driving (2023) [source](https://arxiv.org/abs/2309.17080)

[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Rules

### Simulated robot soccer

Two teams of software players, each seeing only part of the pitch through noisy senses, play football in a simulated stadium to test cooperation and strategy.

Question it answers: Can independent software players coordinate well enough to beat a rival team?

Leaves out: The pitch, ball and robot bodies are idealised, so success shows tactics rather than skill with real hardware.

Runs on: about 10^9 FLOPS rough real time on a workstation for a match. Since: 1998 Noda and colleagues' Soccer Server, a tool for research on multi-agent systems. [source](https://doi.org/10.1080/088395198117848)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Real time](https://simulation-hypothesis.com/simulations/methods#m-real-time)[Game theory](https://simulation-hypothesis.com/simulations/methods#m-game-theory)floor: Populations

### Software engineering agent sandboxes

An agent is given a real code project and a problem report, edits files and runs tests in a sandbox, and is scored on whether the fix passes.

Question it answers: Can an AI agent fix real bugs in real software projects without help?

Leaves out: Tests passing is not the same as a correct fix, and the projects are a small slice of all software.

Runs on: about 10^15 FLOPS rough thousands of model runs, each with many calls and test runs. Since: 2023 Jimenez and colleagues' SWE-bench of real GitHub issues for language models. [source](https://arxiv.org/abs/2310.06770)

[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)[Real time](https://simulation-hypothesis.com/simulations/methods#m-real-time)[Rules and states](https://simulation-hypothesis.com/simulations/methods#m-rules)floor: Rules

### Desktop computer agent environments

A full virtual computer with real programs is given to an agent that sees the screen and clicks and types, and is scored on whether the task was done.

Question it answers: Can an AI agent carry out tasks on a computer the way a person does?

Leaves out: The tasks are fixed and scripted, while real work is open-ended, interrupted and shared with other people.

Runs on: about 10^15 FLOPS rough thousands of runs, each running a virtual machine and many model calls. Since: 2024 Xie and colleagues' OSWorld, benchmarking agents in real computer environments. [source](https://doi.org/10.52202/079017-1650)

[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)[Real time](https://simulation-hypothesis.com/simulations/methods#m-real-time)[Rules and states](https://simulation-hypothesis.com/simulations/methods#m-rules)floor: Rules

[◂ Cities, infrastructure and operations](https://simulation-hypothesis.com/simulations/cities-operations)[Mathematics and computation ▸](https://simulation-hypothesis.com/simulations/mathematics-computation)