project-sid game verified
Project Sid
An Altera research demonstration that runs 10 to 1,000-plus language-model agents with the concurrent PIANO architecture in Minecraft to study role specialization, collective rule-making, and cultural transmission.
parts
8 mechanics, as Project Sid does themcore · 3
Elections
Constituents give feedback on a constitution, amendments are generated, agents vote, and the amended constitution is distributed on a fixed schedule (feedback at 300 s, vote at 420 s, distribution at 600 s).
Gossip propagation
Memes and a religion (Pastafarianism, seeded by 20 priests in one town) spread through agent conversations and travel; conversion was tracked by keyword use and had not saturated after two hours.
Taxation
Agents deposit a share of inventory into community chests every 120 seconds at the constitutional rate, starting at 20 percent; when amendments cut it to 5-10 percent, payments fell to about 9 percent.
supporting · 5
Generative content on demand
Goal generation, talking, and the cognitive controller run on language-model calls, with GPT-4o as the base model in the strongest configuration.
Hand crafting
Agents gather and craft Minecraft items; a 49-agent group plateaued at about 320 unique items over four hours.
Memories and thoughts
A memory module stores conversations, actions, and observations across short and long timescales.
Relationship values
A social-awareness module infers other agents' sentiment, and the paper compares perceived with true likeability.
Resource extraction
core loop
seed agents with personas and community goals -> concurrent modules perceive, remember, talk, and act in Minecraft -> cognitive controller commits a coherent decision -> roles specialize and rules are paid, debated, and amended -> beliefs and memes spread between towns
design notes
PIANO runs about ten modules concurrently at different speeds (memory, action awareness, goal generation, social awareness, talking, skill execution), with fast non-language reflex networks and a cognitive controller that acts as a bottleneck and broadcasts one decision so speech and action stay consistent; GPT-4o was the strongest base model. In 30-agent runs agents settled into persistent roles (farmer, miner, engineer, guard, explorer, blacksmith) that collapsed without the social module. In a 29-agent constitution experiment with a 20 percent tax paid into community chests every 120 seconds, pro- and anti-tax influencers shifted votes, amendments passed, and payments followed the new rate. In a 500-agent run, Pastafarianism seeded by 20 priests spread to other towns; 1,000-agent attempts exceeded the Minecraft server's capacity. The preprint states no public code release.
appears in
the paper, 4 sections · 1 chapterrelations
1 stored edgesbacklinks
computed from other records
Many-agent LLM societiesclaims
1 with sourcesExperiments ranged from 10 to 500 agents in reported results; runs above 1,000 agents were attempted but exceeded the Minecraft server's capacity.
We have also simulated societies with over 1000 agents, but these runs exceeded the computational constraints
Section 5.3, Cultural Transmission