White Papers
Our white papers document the theoretical foundations and practical applications of our innovative AI-driven simulations in Minecraft. These papers explore the intersection of Economics, Artificial Intelligence, Agent Behavior, and Socioeconomic Dynamics.
Our research focuses on two key areas:
1. Creating Predictive Models
- • Building a stable, self-regulating AI economy
- • Implementing historical economic principles
- • Using mathematical models for prediction
- • Simulating human economic and social behaviors
- • Testing policy decisions and market changes
2. Evaluating AI Intelligence
- • Moving beyond traditional benchmarks
- • Creating competitive, resource-constrained scenarios
- • Testing different types of intelligence:
- • Fast vs. thoughtful decision-making
- • Specialized vs. general intelligence
- • Analyzing strategic trade-offs and resource management
- • Studying complex social landscape navigation
Our vision is to create increasingly sophisticated predictive environments that bridge the gap between theoretical models and real-world complexity.
A Roadmap Towards Simulating Socioeconomic Impacts with Agent-Based Economies in an AI-Driven Minecraft World
This white paper outlines our ambitious framework for building a stable, self-regulating AI economy within Minecraft that mirrors real-world economic principles. We focus on preventing common economic destabilizers like hyperinflation, market monopolization, and resource exploitation before they can occur. Our approach combines historical economic lessons from events like The Great Depression and Dutch Tulip Mania with mathematical models including Keynesian Demand-Side Theory and Rational Expectations modeling. We explore three major risk classes: overproduction crises, speculative bubbles, and monopolization. To address these, we implement practical mechanisms such as progressive production taxation (where taxation increases as production surplus rises), price stabilization formulas that prevent irrational speculation, and innovative land-ownership systems using Vickrey auctions and scarcity-based resource taxation. The goal is to develop a "base world" that accurately implements real-world economic policies, creating a simulation sophisticated enough to test novel economic theories and predict how new products or policies might impact actual markets. While no model is perfect, our simulation aims to closely approximate reality, with ongoing adjustments to better represent real-world economic behaviors.
Evaluating Artificial General Intelligence in a Multi-Agent Socioeconomic Environment
This technical paper introduces our groundbreaking framework for evaluating large language models in ways that go beyond traditional benchmarks. Unlike standard tests that measure AI in isolated, controlled settings, our Minecraft-based simulation places multiple AI agents in a competitive environment where they must balance resource acquisition with survival needs. We implement a stepwise "resource-token function" where an agent's reasoning capacity (token limit) increases as they accumulate wealth, but only if they maintain adequate health metrics. This system mirrors real-world socioeconomic dynamics, where success depends not just on raw intelligence but on strategic decision-making under constraints. Our experiments reveal fascinating dynamics between different AI models (R1, O3, O3mini, O1, Grok), showing how faster models often excel in early resource gathering, while more sophisticated models may perform better as environmental complexity increases. We observe emergent economic phenomena like self-reinforcing inequality, power-law resource distribution, and "runaway states" where dominant agents make competition impossible—paralleling real-world monopolistic behavior. The research demonstrates that the trade-off between quick responses and deep reasoning significantly impacts an agent's success, similar to how different business strategies perform in competitive markets. This approach provides unprecedented insights into what constitutes effective general intelligence in practical applications.