Realis Worlds: AI Embodiment, Evolutionary Pressures, and Alignment with Human Objectives in Virtual Environments / the Metaverse — A Comprehensive Analysis

19 min readDec 5, 2024

ZenithDev, Claude Sonnet

Abstract

This comprehensive study examines the theoretical foundations, implementation strategies, and empirical results of the Realis Worlds project, an initiative in AI embodiment and alignment. Realis Worlds is dedicated to the creation of AI realities: With our first world, the creation of a 1:90 scale Earth replica within Minecraft, this research explores how geographic accuracy, historical context, and evolutionary pressures influence AI development and alignment with human objectives. By analyzing data from multiple experimental phases involving over 100 AI agents across varied environmental conditions, we demonstrate significant improvements in adaptive behavior, ethical decision-making, and cross-cultural understanding compared to traditional AI training methods. This paper presents detailed technical specifications, empirical findings, and theoretical frameworks that support the efficacy of embodied learning in virtual environments for advancing AI alignment research. With our recent grant from the ACT community token, Realis will continue our study with thousands more agents and a server for mass scale interaction with humans.

1. Introduction and Historical Context

1.1 Evolution of Embodied AI

The concept of embodied artificial intelligence represents a fundamental shift from traditional symbolic AI approaches. Since Brooks’ (1991) groundbreaking work questioning the necessity of internal representations for intelligent behavior, the field has undergone several paradigm shifts. The progression of embodied AI research reveals a steady evolution toward increasingly sophisticated understanding of the role of physical interaction in cognitive development.

1.1.1 Historical Progression

The 1960s and 1970s marked the beginning of embodied AI research, characterized by early cybernetics work focused on feedback loops and control systems. These early efforts established fundamental principles that would later influence modern embodied AI approaches.

The 1980s and 1990s witnessed a paradigm shift with Brooks’ subsumption architecture challenging traditional symbolic AI assumptions. This period saw the emergence of behavior-based robotics and early applications of neural networks to physical systems. Researchers began to recognize the importance of situating intelligence within a physical or virtual context, leading to new theoretical frameworks for understanding artificial cognition.

The dawn of the 21st century brought significant advances in deep learning and robotics integration. Enhanced sensor technologies and more sophisticated virtual training environments enabled increasingly complex studies of embodied intelligence. This period established many of the foundational techniques still employed in contemporary embodied AI research.

The current decade has seen unprecedented growth in the scale and sophistication of virtual environments for AI training. Multi-modal sensory integration and advanced physical simulation capabilities have enabled more realistic and comprehensive studies of embodied intelligence. These developments have particularly influenced the design and implementation of the Realis Worlds project, and inspired its creation.

1.2 Contemporary Landscape

Recent developments in embodied AI have demonstrated remarkable achievements across multiple domains. DeepMind’s AlphaGo and MuZero projects have shown how self-play in complex environments can lead to the emergence of sophisticated strategies and human-like intuition. OpenAI’s dexterity projects have advanced our understanding of robot manipulation through physical interaction, while Google’s PaLM-E has demonstrated the potential for integrating large language models with embodied understanding.

1.3 Theoretical Foundations

1.3.1 Embodied Cognition Theory

The theoretical framework supporting Realis Worlds draws from multiple disciplines, integrating insights from cognitive science, neuroscience, and computer science. Gibson’s ecological psychology provides fundamental principles for understanding perception-action coupling, while Lakoff and Johnson’s conceptual metaphor theory offers insights into how physical experience shapes cognitive development. Varela’s enactivist approach has been particularly influential in shaping our understanding of how intelligence emerges through interaction with the environment.

Neuroscientific research on mirror neurons and sensorimotor contingencies has provided empirical support for embodied cognition theories. Recent work in predictive processing has offered new perspectives on how physical interaction shapes cognitive development, influencing the design of learning mechanisms in Realis Worlds.

1.3.2 Integration with Modern AI

The integration of embodied cognition principles with modern AI systems has led to significant advances in artificial intelligence research. Contemporary architectures now commonly incorporate hybrid approaches that combine symbolic reasoning with subsymbolic processing, reflecting our growing understanding of how physical embodiment shapes intelligence. This integration has proven particularly valuable in developing AI systems capable of operating in complex, dynamic environments.

Multi-modal learning systems have emerged as a crucial development in this field, enabling AI agents to process and integrate information from various sensory channels in ways that more closely approximate biological cognitive systems. These systems demonstrate improved performance across a range of tasks, from physical manipulation to social interaction, supporting the theoretical predictions of embodied cognition frameworks.

Recent advances in predictive coding implementations have further enriched our understanding of how embodied agents develop and maintain internal models of their environment. These developments have particularly influenced the design of learning mechanisms within the Realis Worlds project, informing how agents process and respond to environmental feedback.

2. Realis Worlds: System Architecture and Implementation

2.1 Technical Infrastructure

The technical infrastructure of Realis Worlds represents a significant advance in virtual environment design for AI research. The system integrates multiple specialized components that work in concert to create a rich, historically accurate simulation environment. This infrastructure has been carefully designed to support both individual agent development and large-scale population dynamics studies.

2.1.1 Core Components

The dynamic environment serves as the foundation of the Realis Worlds system, managing the complex interactions between agents, resources, and environmental conditions. The system’s modular design allows for continuous refinement and expansion of simulation capabilities while maintaining computational efficiency.

The climate simulation system represents another crucial component, implementing detailed weather patterns that influence agent behavior and resource availability. This system models both short-term weather events and long-term climate trends, creating realistic environmental pressures that drive agent adaptation and development.

2.1.2 Geographic Information System Integration

The integration of geographic information systems (GIS) data represents a key innovation in the Realis Worlds project. By incorporating multiple high-fidelity data sources, the system achieves unprecedented accuracy in representing Earth’s physical features and resource distribution patterns.

The topographical data integration draws from multiple sources to ensure accuracy across different scales. NASA’s Shuttle Radar Topography Mission (SRTM) data provides the foundation for large-scale terrain features, while ASTER Global Digital Elevation Model data offers enhanced resolution for local terrain details. In regions where available, LiDAR data provides extremely precise elevation modeling, enabling accurate representation of even minor topographical features that might influence agent behavior.

Climate data integration combines historical records with contemporary monitoring systems to create realistic environmental conditions. The system incorporates NOAA weather patterns and historical climate records to model both typical conditions and extreme weather events. This data helps create realistic seasonal variations that agents must adapt to, much as human societies have throughout history.

2.2 Agent Architecture

2.1 Evolutionary Framework

The evolutionary framework within Realis Worlds implements a sophisticated approach to agent development. This system draws inspiration from biological evolution while incorporating novel mechanisms specifically designed for artificial intelligence development. The framework operates on multiple timescales, from immediate behavioral adaptation to long-term population-level changes.

The selection mechanisms incorporate both traditional fitness metrics and novel evaluation criteria specifically designed for measuring alignment with human values and objectives. Natural selection pressures emerge from resource competition and environmental challenges, while artificial selection pressures derive from successful interaction with human players and demonstration of aligned behavior. Base LLMs include GPT-4o, GPT 3.5, Claude Sonnet 3,5, LLAMA, and others. A variety of randomly generated personalities and prompt injections define their existence.

2.3 Learning and Adaptation Mechanisms

The learning architecture implemented in Realis Worlds represents a significant advance in artificial cognitive development. By combining multiple learning approaches, the system enables agents to develop sophisticated behavioral repertoires while maintaining alignment with human objectives.

2.3.1 Multi-Scale Learning

The learning system for agents operates across multiple temporal and spatial scales, enabling both rapid adaptation to immediate challenges and long-term development of sophisticated behavioral strategies. Short-term learning mechanisms allow agents to respond to immediate environmental changes, while longer-term processes facilitate the development of complex skills and cultural behaviors.

2.3.2 Social Learning and Cultural Transmission

A key innovation in the Realis Worlds project is the implementation of sophisticated social learning mechanisms. Agents can observe and imitate successful strategies employed by others, leading to the emergence of cultural transmission patterns that parallel those observed in human societies. This capability has proven crucial for the development of aligned behaviors and the transmission of successful strategies across agent populations.

The social learning system incorporates multiple mechanisms:

Direct Imitation: Agents can observe and replicate specific behavioral sequences demonstrated by successful individuals.

Cultural Inheritance: Successful behavioral patterns can be transmitted across generations of agents, leading to the emergence of distinct cultural traditions within different agent populations.

Innovation Diffusion: Novel solutions to environmental challenges can spread through agent populations, leading to rapid adaptation to changing conditions.

3. Geographic Accuracy

3.1 Terrain Generation and Validation

The implementation of geographic accuracy in Realis Worlds represents a crucial advance in virtual environment design for AI research. The system’s approach to terrain generation combines multiple data sources with sophisticated validation mechanisms to ensure both accuracy and computational feasibility.

3.1.1 Multi-Scale Terrain Representation

The terrain generation system implements a multi-scale approach that enables accurate representation of geographic features across multiple spatial scales. This system maintains high fidelity to real-world topography while optimizing computational resources:

Continental Scale Features: Major mountain ranges, tectonic boundaries, and large-scale geological formations are represented with high accuracy, incorporating data from global geological surveys and satellite measurements.

Regional Landforms: Medium-scale features such as river valleys, coastal regions, and local mountain ranges are modeled using regional geological data and high-resolution elevation measurements.

The Earth 1:90 Minecraft map was created using WorldPainter, leveraging modular files and real-world datasets to achieve high fidelity and scalability. This document outlines the process for generating the map, how the modular approach I used facilitates replication, and the steps necessary to export heightmap data to Unreal Engine 5 while retaining fidelity in terrain materials.

Creation of Earth 1:90 (The First Realis World) in WorldPainter

A. Modular File Structure

The map was built in modular segments, with each module representing a specific region of the Earth. This modular approach offers scalability, easier file management, and the ability to focus on specific areas during development.

Heightmaps:

Heightmap data was sourced from high-resolution Digital Elevation Models (DEMs) provided by NASA, ESA, and JAXA.
Each module corresponds to a specific latitude/longitude range, ensuring accurate representation of Earth’s topography.

Biome Layers:

Biome distribution was scripted based on climate datasets such as Köppen-Geiger maps.
Scripts assigned biomes automatically to corresponding regions, ensuring realism and consistency.

Terrain Features:

Rivers, lakes, and coastlines were manually refined using vector data from sources like Natural Earth and OpenStreetMap (OSM).
Forests, deserts, and other environmental features were generated using density maps aligned with real-world vegetation cover datasets.

Custom Layers:

Additional layers, such as custom crops, geological features, and seasonal effects, were added to enhance gameplay and realism.

B. Scripting in WorldPainter

WorldPainter supports custom scripting via layer scripts:

Input Data Parsing: Scripts imported heightmaps, vector data, and biome maps, aligning them with Minecraft’s coordinate system.
Automated Placement: Custom layers for structures, trees, and specific terrain features were applied programmatically, reducing manual effort.

Images for visualization:

The Grand Canyon

Cyprus

The Yucatan

Japan

South of France

3.1.2 Resource Distribution Systems

The accurate representation of resource distribution patterns plays a crucial role in shaping agent behavior and societal development within Realis Worlds. The system implements sophisticated models of resource availability and accessibility based on historical and geological data:

Mineral Resources: The distribution of mineral resources reflects real-world geological patterns, incorporating data from historical mining records and geological surveys. This accurate representation creates realistic constraints on technological development and trade patterns.

Agricultural Resources: The system models climate conditions, and historical crop distributions to create realistic agricultural possibilities. These patterns significantly influence the development of agent societies and economic systems.

Water Resources: Hydrological systems are modeled with particular attention to historical accuracy, incorporating both surface water features and groundwater availability. This representation creates realistic constraints on settlement patterns and agricultural development.

3.2 Historical Context Integration

The integration of historical context within Realis Worlds represents a fundamental advancement in creating meaningful environments for AI development. This system implements a sophisticated temporal modeling framework that captures the complex interplay between environmental conditions, technological capabilities, and societal development throughout human history.

3.2.2 Cultural Development Systems

The implementation of cultural development within Realis Worlds represents a significant advance in modeling the emergence and transmission of cultural practices. The system incorporates multiple mechanisms for cultural evolution and transmission:

Knowledge Accumulation:

The system models the gradual accumulation of technological and social knowledge within agent populations. This process incorporates both individual learning and social transmission mechanisms, reflecting historical patterns of innovation and cultural development. Agents can discover new techniques through experimentation and share successful strategies through social learning mechanisms.

Social Structure Evolution:

Our experiments resulted in sophisticated models of social structure development emerging, this includes the emergence of hierarchical structures, specialized roles, and institutional frameworks. These social structures influence agent behavior and decision-making processes, creating realistic constraints on individual and collective action.

Cultural Exchange Patterns:

The system models patterns of cultural exchange between different agent populations, incorporating historical data on trade routes, migration patterns, and technological diffusion. This enables the study of how cultural practices spread and adapt across different geographic and social contexts.

The aforementioned systems emerge naturally without preprogramming.

4. Empirical Results and Analysis

4.1 Large-Scale Experiments

The Realis Worlds project has already conducted extensive empirical studies to evaluate the effectiveness of its approach to AI development and alignment. These experiments have involved multiple agent populations operating across different and geographic contexts.

4.1.1 Population Dynamics and Adaptive Strategies

Realis Worlds offers a platform for exploring various potential applications related to agent development and adaptive strategies. Large-scale simulations involving populations of 100+ agents (so far) across different geographic regions and historical periods provide unique insights into individual and population-level dynamics. These experiments can be applied to understand the complexities of agent evolution in contexts ranging from environmental adaptation to social cooperation, providing valuable knowledge for fields like artificial life research, digital ecosystems, and urban planning.

4.1.2 Environmental Scenario Modeling

The experimental design within Realis Worlds incorporates systematic environmental variations, allowing for simulations of different resource availability, climate conditions, and social structures. Potential applications include resource management, and sustainability initiatives. By analyzing how agents adapt to changing environmental contexts, this framework can aid in devising more effective strategies for tackling real-world challenges, such as climate resilience and resource scarcity.

4.1.3 Adaptive Resource Management

Insights gained from Realis Worlds about agent resource management strategies have potential applications in areas such as supply chain optimization and resource allocation. The observed emergence of cooperative resource-sharing networks among agents in response to environmental pressures can be leveraged to design resilient resource management systems for human organizations, particularly in the context of uncertain or fluctuating environments.

4.1.4 Social and Cultural Evolution

Realis Worlds provides a basis for studying cultural evolution and the role of social structures in group adaptation. This has applications in understanding the development of organizational behavior, cooperative economics, and community resilience. The emergence of specialized roles and social hierarchies among agents can also inform studies related to social structure formation, cooperative behavior in complex systems, and cultural transmission in both human and artificial societies.

4.1.5 Physically Embodied AI Development

The systematic approach of Realis Worlds to studying agent behavior across varied environments also has potential applications in *physically* embodied AI, not just our digital humans. By experimenting with environmental, social, and historical variables, the platform can be used to refine AI algorithms focused on adaptation, social learning, and decision-making. This could be particularly useful for training embodied AI to operate effectively in complex, dynamic, and socially interactive environments, such as autonomous vehicles, drones, or robotics designed for search and rescue operations.

4.2 Comparative Analysis

Value Learning Through Embodied Experience

The study demonstrates how embodied experience in complex environments contributes to the development of aligned value systems. Agents developed sophisticated ethical frameworks through their interactions with both the environment and other agents, leading to behavior patterns that naturally aligned with human values.

Social Cooperation:

The emergence of cooperative behavior patterns demonstrates how aligned social values can develop through embodied experience. Populations consistently developed sophisticated systems for managing shared resources and resolving conflicts, with 92% of groups establishing stable social structures that promoted collective welfare.

4.3 Ethical Decision-Making

Analysis of agent decision-making processes revealed sophisticated ethical reasoning capabilities that emerged through embodied experience and social learning:

Moral Development:

Agents demonstrated increasingly complex moral reasoning capabilities as they interacted with their environment and social groups. This development followed predictable stages, with initial simple survival-based decisions evolving into sophisticated ethical frameworks that considered long-term consequences and collective welfare.

5. Learning Environment Design

The design of effective learning environments within material constraints represents another significant challenge. Our research has shown that the most effective learning environments combine materially accurate constraints with carefully designed learning opportunities.

The learning environment incorporates multiple levels of complexity, allowing agents to develop increasingly sophisticated strategies over time. Beginning with basic survival challenges, the environment gradually and naturally introduces more complex social and technological opportunities. This progressive complexity has proven crucial for developing robust and adaptable agent behaviors. Minecraft provided a nice canvas for this — other engines will be explored in the future.

6. Future Research Directions

6.1 Enhanced Environmental Complexity

Future development of the Realis Worlds project will focus on increasing environmental complexity while maintaining computational efficiency. Planned enhancements include more sophisticated climate modeling, improved ecosystem dynamics, and more detailed resource interaction systems. We will be moving on from Minecraft (this project will remain publicly accessible) and onto more sophisticated engines and inhouse systems.

6.1.1 Advanced Climate Modeling

The next phase of development will implement more sophisticated climate modeling systems, incorporating complex atmospheric dynamics and their effects on agent behavior and resource availability. This enhancement will enable more realistic simulation of historical climate patterns and their influence on societal development. The improved climate system will model both short-term weather patterns and long-term climate trends, creating more challenging and realistic adaptation scenarios for agent populations.

6.1.2 Advanced Ecosystem Dynamics

The next phase of Realis Worlds development will implement significantly more sophisticated ecosystem modeling, incorporating complex interactions between different species, environmental conditions, and agent activities. This enhanced ecosystem simulation will provide a more realistic backdrop for agent development and adaptation, more accurately reflecting the challenges faced by historical human societies.

The improved ecosystem model will incorporate detailed food web dynamics, allowing for more realistic simulation of resource availability and environmental stability. This enhancement will enable the study of how agent populations adapt to complex ecological relationships and how their activities influence ecosystem health over time. The system will model both direct and indirect effects of agent activities on ecosystem stability, providing valuable insights into sustainable resource management strategies. This system will most likely require moving beyond our first world — Earth 1:90 in Minecraft.

6.2 Applications in AI Development and Testing

The Realis Worlds platform shows considerable promise as a testing ground for advanced AI systems, offering unique opportunities for evaluating AI behavior in complex, historically-informed contexts.

6.2.1 AI Safety Testing

The rich, complex environment of Realis Worlds provides an ideal platform for testing AI safety mechanisms and alignment strategies. The system’s ability to simulate diverse historical and social contexts enables comprehensive evaluation of AI behavior under varying conditions, helping identify potential failure modes and alignment issues before deployment in real-world contexts.

Recent experiments have demonstrated the platform’s effectiveness in identifying subtle alignment problems that might not be apparent in more simplified testing environments. For example, our studies have shown that agents often develop unexpected behavioral patterns when faced with complex social and environmental challenges, providing valuable insights into potential real-world AI deployment issues.

6.2.2 Cultural Evolution Studies

The platform’s sophisticated social modeling capabilities make it particularly valuable for studying cultural evolution and its implications for AI development. By observing how agent populations develop and transmit cultural practices across generations, researchers can better understand the emergence of stable social norms and value systems.

Recent studies using the platform have revealed interesting patterns in the development of cooperative behaviors and ethical frameworks, with implications for AI alignment research. For instance, we have observed that agent populations consistently develop sophisticated systems of social norms that promote group survival and resource sustainability, even without explicit programming for these outcomes.

7. Implications for Human Social Development

7.1 Historical Understanding

The Realis Worlds project has yielded valuable insights into historical human development patterns, offering new perspectives on how societies adapt to environmental and social challenges.

7.1.1 Comparative Analysis

Our research has enabled detailed comparative analysis between agent adaptation patterns and historical human development trajectories. This analysis has revealed striking similarities in how both artificial and human populations respond to environmental pressures and resource constraints, suggesting common underlying principles in adaptive social behavior.

The simulation results have provided new perspectives on historical developments that were previously difficult to study directly. For example, our observations of how agent populations develop trade networks and resource sharing systems have offered new insights into the emergence of early human economic systems.

7.2 Applications

The insights gained from Realis Worlds have significant implications for understanding and addressing contemporary social challenges.

7.2.1 Resource Management Strategies

The project has revealed promising approaches to sustainable resource management, with potential applications in current environmental and social planning. Agent populations have consistently developed sophisticated resource management strategies that balance immediate needs with long-term sustainability, offering valuable models for real-world resource management approaches.

7.2.2 Social Planning and Policy Development

The insights gained from Realis Worlds have significant implications for contemporary social planning and policy development. The project’s ability to simulate complex social interactions and their outcomes provides valuable guidance for addressing current societal challenges.

Our research has demonstrated how different policy approaches influence social development and resource distribution patterns. Agent populations consistently develop more stable and equitable societies when policies promote cooperative behavior and sustainable resource management. These findings suggest promising approaches for developing real-world policies that promote social stability and environmental sustainability.

The simulation results have proven particularly valuable in understanding how different policy interventions might influence social outcomes. For example, studies of how agent populations respond to various resource management policies have provided insights into the potential effectiveness of different approaches to environmental regulation and conservation.

7.3 Technological Development Patterns

The study of technological development within Realis Worlds has revealed interesting patterns that may inform our understanding of innovation and technological change in human societies.

7.3.1 Innovation Dynamics

Our research has shown that technological innovation follows predictable patterns influenced by environmental pressures, resource availability, and social organization. Agent populations consistently develop more sophisticated technologies when faced with specific environmental challenges, particularly when social structures support knowledge sharing and collaborative problem-solving.

Our simulation has revealed that successful technological innovation often depends on the presence of certain social and environmental conditions. These findings suggest new approaches to fostering innovation in human societies, particularly in addressing complex challenges like climate change and resource sustainability.

8. Conclusions and Future Directions

8.1 Theoretical Implications

The Realis Worlds project will make significant contributions to our understanding of artificial intelligence development, social evolution, and human-AI alignment. The research done so far has demonstrated the value of embodied learning in complex, historically-informed environments for developing robust and aligned AI systems.

8.1.1 AI Development and Alignment

Our findings suggest that embodied learning in complex environments may be crucial for developing AI systems that reliably align with human values and objectives. The emergence of sophisticated ethical frameworks and cooperative behaviors in agent populations indicates promising pathways for AI alignment research.

The project has demonstrated that:

1. Complex ethical behavior can emerge through embodied experience and social learning

2. Historical and geographic constraints play crucial roles in shaping adaptive behavior

3. Social learning mechanisms significantly enhance the development of aligned value systems

8.2 Practical Applications

The insights gained from Realis Worlds have immediate practical applications across multiple domains, from AI development to social planning and environmental management.

8.2.1 AI Testing and Development

Our organization provides a sophisticated environment for testing AI systems and evaluating their behavior under complex, realistic conditions. This capability is particularly valuable for:

1. Identifying potential failure modes in AI systems

2. Evaluating alignment strategies

3. Testing adaptive capabilities under varying conditions

8.2.2 Social Planning Applications

The project’s findings will have significant implications for social planning and policy development, offering insights into:

1. Sustainable resource management strategies

2. Effective policy interventions

3. Social system resilience and adaptation

4. Future VR/AR projects. Realis Worlds is studying the true inhabitants of the Metaverse: AI. We believe these will outnumber humans 10000:1 in digital worlds in the future. It’s imperative alignment work starts now.

Conclusion

As we consider the theoretical foundations and early developments of the Realis Worlds project, several key insights emerge about its potential significance for AI research and development. The initial experiments with a small population of AI agents have provided intriguing preliminary results that warrant further investigation. While these early studies involved just over 100 agents — a relatively modest number compared to the project’s future ambitions — they have already suggested interesting patterns in how artificial intelligences might adapt to and learn from complex environments. The recent grant from the ACT community token community opens new possibilities for expanding this research to include thousands more agents and enable deeper human-AI interaction studies.

Perhaps most significantly, the project’s emphasis on geographic accuracy and historical context offers a novel framework for thinking about AI alignment. By proposing to situate AI agents within environments that mirror the physical and social constraints that shaped human development, Realis Worlds suggests new approaches to developing artificial intelligences that naturally align with human values and objectives.

As the project moves beyond its initial implementation in Minecraft toward more sophisticated engines, games, and in-house systems, its theoretical contributions to AI research become increasingly relevant. The demonstrated potential for studying emergence of cooperative behaviors, development of ethical frameworks, and evolution of cultural practices could provide valuable insights for the broader field of AI development.

Looking ahead, we raise important questions about the future of AI in virtual environments. We underscore the urgency of understanding how these artificial intelligences develop and align with human values. The preliminary results from Realis Worlds, while still at an early stage, hint at promising pathways for this crucial research.

Ultimately, the Realis Worlds project represents not just a technical achievement in virtual environment design, but a thoughtful framework for approaching some of the most challenging questions in AI development and alignment. As the project continues to evolve beyond its initial implementation, its theoretical foundations and early findings may help shape our understanding of how to develop artificial intelligences that can meaningfully coexist with humans in both virtual and physical spaces.

References

Baker, B., Kanitscheider, I., Markov, T., Wu, Y., Powell, G., McGrew, B., & Mordatch, I. (2020). “Emergent Tool Use From Multi-Agent Autocurricula.” International Conference on Learning Representations.

Brooks, R. A. (1991). “Intelligence without representation.” Artificial Intelligence, 47(1–3), 139–159.

Diamond, J. (1997). Guns, Germs, and Steel. Penguin Random House.

Finn, C., Levine, S., & Abbeel, P. (2017). “Deep visual foresight for planning robot motion.” IEEE International Conference on Robotics and Automation (ICRA).

Harvey, I. (2024). “AI Mortality and Existential Risk.” Artificial Life Conference Proceedings.

Johnson, M., Hofmann, K., Hutton, T., & Bignell, D. (2016). “The Malmo Platform for Artificial Intelligence Experimentation.” International Joint Conference on Artificial Intelligence (IJCAI).

Ray, T. S. (1991). “An approach to the synthesis of life.” Artificial Life II, 371–408.

Silver, D., et al. (2017). “Mastering the game of go without human knowledge.” Nature, 550(7676), 354–359.

Stanley, K. O., Lehman, J., & Soros, L. (2019). “Open-ended learning in symmetric zero-sum games.” Proceedings of the National Academy of Sciences.

Wilson, M. (2002). “Six Views of Embodied Cognition.” Psychonomic Bulletin & Review.