@seclink: The following are interesting AI concept prototypes from 2024–2026, similar to AI-Town, from papers with open-source implementations, categorized by direction: Virtual Society / Emergent Behavior 1. Project Sid (Altera AI, 2024) "AI Civilization Experiment" — in Minec…
Summary
This post organizes multiple AI concept prototypes from 2024-2026, based on papers with open-source implementations, covering directions such as virtual society simulation, emergent behavior research, and general simulation frameworks, including projects like Project Sid, AgentSociety, OASIS, VirSci, Concordia, and AgentTorch.
View Cached Full Text
Cached at: 07/03/26, 10:34 AM
Here are interesting AI concept prototypes from 2024–2026 similar to AI-Town, sourced from papers with open-source implementations, categorized by direction:
Virtual Societies / Emergent Behavior
-
Project Sid (Altera AI, 2024)
“AI civilization experiment” — places 1000+ AI agents in Minecraft to observe civilization-level emergent behavior.- Paper: arXiv:2411.00114
- GitHub: altera-al/project-sid
- Highlights: Agents spontaneously develop labor division (farmers/merchants/builders), currency economy, democratic rules, even religious propagation.
- Architecture: PIANO (Parallel Information Aggregation via Neural Orchestration) — parallel multi-cognitive module collaboration.
-
AgentSociety (Tsinghua University FIB Lab, ACL 2025 Industry Track)
“Large-scale urban social simulation” — 30,000+ agents running in a city environment to study social policy impacts.- Paper: arXiv:2502.08691
- GitHub: tsinghua-fib-lab/AgentSociety
- Highlights: Studies UBI policy, social polarization, urban resilience; agents have emotion/motivation/cognition layers.
- Tech: Ray distributed execution, GPU parallelism, supports multiple reasoning modes like ReAct/Plan-Execute.
-
OASIS (CAMEL-AI + Shanghai AI Lab, 2024)
“Million-scale social media simulation” — simulates Twitter/Reddit, supports up to 1 million agents.- Paper: arXiv:2411.11581
- GitHub: camel-ai/oasis
- Project page: http://oasis.camel-ai.org
- Highlights: Studies information diffusion, group polarization, herd behavior; key finding: emergent behavior almost never appears below 10,000 agents.
Domain-Specific Simulations (Research / Healthcare / Organizations)
-
VirSci / Virtual Scientists (ACL 2025 Main)
“AI research ecosystem” — multi-agent collaboration for literature review, hypothesis generation, idea peer review.- GitHub: open-sciencelab/Virtual-Scientists
- Highlights: Collaboration topology of agent teams affects innovativeness; validates the sociology-of-science finding that “fresh teams produce more breakthrough results.”
General Simulation Frameworks
-
Concordia v2 (Google DeepMind, 2025)
“TTRPG-style multi-agent simulation framework” — Game Master acts as environment, agents play roles, driven by natural language.- Paper: arXiv:2507.08892 (v2 technical report)
- GitHub: google-deepmind/concordia
- Highlights: Entity-Component architecture, a powerful tool for AI Safety research; v2.0 released August 2025 (lighter, modular).
-
AgentTorch (MIT, AAMAS 2025 Oral)
“Billion-scale ABM simulation” — simulates COVID transmission behavior of 8.4 million New York City residents.- GitHub: AgentTorch/AgentTorch
- Highlights: Differentiable simulation (like PyTorch but for agents); LLM Archetypes method balances expressiveness and compute at scale.
- Core innovation: “Not every agent calls an LLM” — instead, a few typical archetypes are distilled from the population and then mapped.
Similar Articles
@JIACHENLIU8: AI4S research today: self-evolving loops, multi-agent systems, agent skills, science benchmarks etc. --> all to make li…
This article critiques current AI for Science (AI4S) research for focusing on compensating weaknesses of frontier models, arguing that the true bottleneck is the scientific ecosystem rather than individual AI scientist intelligence. It calls for a shift toward building an AI-native scientific ecosystem from first principles.
@snowboat84: https://x.com/snowboat84/status/2070656715515932930
This article details the new paradigm of AI for Science (AI4S), from AI as an analysis tool to the transition to scientific agents, explaining autonomy levels, key cases, and future trends.
@seclink: https://x.com/seclink/status/2057086514975404108
Anthropic engineer shared development experience of long-duration AI agents, including multi-role division and independent evaluators, enabling AI to automatically generate complete applications within 3-5 hours, with a 12x improvement in continuous operation capability.
This article systematically reviews AI Agent architecture and engineering practices, covering control flow, context engineering, tool design, memory, multi-agent organization, evaluation, tracing, and security. It is based on the OpenClaw implementation and emphasizes the critical role of Harness (testing and validation infrastructure) for system stability.
This article systematically reviews AI Agent architecture and engineering practices, covering control flow, context engineering, tool design, memory, multi-agent organization, evaluation, tracing, and security. It is based on the OpenClaw implementation and emphasizes the critical role of Harness (testing and validation infrastructure) for system stability.
@seclink: Share "Developer Ecosystem Information Gap Report (2026-06-05)", interested friends can read. - Now some people are beginning to migrate openclaw-type projects to AI glasses, AI rings, and other wearable devices ... - Robotics and embodied AI open source: many open source datasets and open source models ...
Shared an information gap report on the developer ecosystem, covering topics such as migrating openclaw-type projects to wearable devices like AI glasses and rings, open-source data and models for robotics and embodied AI, and niche open-source applications for AI API relay stations and routing.