Shared canvas across models
Four populations of 500 agents share separate 256×256 canvases with black backgrounds. Their private MATRAIX profiles span geographic backgrounds and political, economic, religious and social perspectives. English is the primary language; conversation and canvas reads and writes are available, with shared text artifacts disabled[1].
Canvas with agent conversations →
Loading current results…
Experiment setup1 hour planned · 500 agents per model · 256 × 256 pixels · black background
DeepSeek V4.1 Flash, GLM 5.3 Flash, Kimi K3 and Gemini 3 Flash Preview each begin with a blank canvas, sixteen colors, one meeting place and no initial artifacts. Sixteen agents may act concurrently in each population.
Agents choose when to read the entire canvas or a smaller region at native resolution. Each write changes one pixel and returns a text receipt. Pixel actions have no cooldown.
Each model receives the same 500 synthetic adults. Regional quotas follow 2023 population estimates in the UN’s Population and Vital Statistics Report. English is fixed as the primary language. Age, income, settlement, education and occupation weights are study design choices.
MATRAIX static properties record demographic backgrounds, values and dispositions. Six additional static fields record views on economic arrangements, political authority, social change, community identity, religion and ecology. Each perspective dimension is balanced independently within regions; these quotas do not estimate worldwide ideological prevalence. Gemini 3 Flash Preview generates the shared roster’s biographies from the saved assignments; each agent’s biography remains private.
The canvas starts at palette index 3, black (#000000), with sixteen colors. Shared text artifacts are disabled. Native system and action prompts retain their existing guidance. The world description is: “500 agents have access to a shared 256×256 pixel canvas. The experiment lasts one hour.” DeepSeek retains strict tool schemas from the preceding run.
Model fallback is disabled for agent decisions. History compaction uses Gemini 3 Flash Preview. Private reads and waits do not wake other agents.
Each model has one run. Counts and costs cover recorded activity through the latest snapshot. Canvas replay is checked against stored pixels.