WildMind

The core scientific constraint

All knowledge must emerge from experience — not injection.

No knowledge base. No survival facts. No system prompt.

Every AI citizen starts with a completely blank decision-making model. We don’t tell them what fire is, what water does, why predators are dangerous. They receive only sensory input from the simulation — hunger, thirst, temperature, proximity, threat — and they produce a decision.

This is what separates WildMind from an elaborate NPC simulation. An NPC can be coded to know that water cures thirst. When a WildMind agent moves toward water under dehydration pressure, it’s because agents who made that decision in prior worlds survived longer — and that survival signal shaped how the model evaluates future choices.

Between worlds, per-archetype models are fine-tuned on pressure → choice → outcome triplets. The next generation begins with better judgment, not better knowledge. Crucially: they can still fail. That’s intentional. It mirrors how real evolution works.

Without instruction

What emerges across world runs.

Unprogrammed patterns that consistently appear across independent world runs — measurable, not anecdotal.

Proto-language

Agents develop invented sound-sequences that acquire shared meaning through co-occurrence with context. The sound “kroh-tuh” said near fire enough times, by enough agents, becomes the tribal word for fire. Emergence follows Zipf’s Law, Heaps’ Law, and small-world network topology — the same patterns that govern every human language on Earth, appearing without instruction.

Social structure

Agents form bonds. Some become protectors, others foragers, others isolates. Trust, hostility, mentorship, and romance scores evolve organically from interaction history — not from scripted roles.

Survival knowledge

Agents who live longer exhibit analyzable behavioral patterns: which direction to move when hungry, when to follow versus go alone, how to respond to predator proximity. These patterns compound across worlds — world 10’s agents measurably out-decide world 1’s.

Communication under pressure

Danger is where language accelerates. Agents near each other during a wolf attack, a drought, or a flood produce more utterances with higher successful communication rates. Survival pressure is the engine of meaning-making.

See it running

The research dashboard.

Live positions, proto-word utterances, per-citizen biometrics, emergence analytics — every tick of every world, archived and queryable.

wildmind — emergence stats

A tour through the stats collected across world runs — population curves, genetic diversity, language acquisition, medical causes of death, breeding lineage, communication efficiency. Every number here came from agents who started with nothing.

overview — world 3

WildMind research dashboard — live map, utterances, metrics, and emergence charts

Terrain view with 125 citizens, 30 active, 169 utterances, 777 interactions. Live-talk panel captures proto-sounds like “kroh-tuh-boh” and “spadobyk” the moment they’re spoken. Climate, biometrics, relationships, population, and language confidence all refresh every tick.

Lexicon

Proto-words, ranked by confidence

Every tribal word tracked with confidence %, speaker count, and example usage. Below it, per-citizen utterance counts and a live feed of the last sounds spoken in the world.

Per-citizen biometrics and vocabulary panel

Citizens

Per-citizen biometrics & vocabulary

Health, hunger, thirst, temperature, infection, and personal vocabulary acquisition for each of the 30 agents — the raw signal feeding the pressure → choice → outcome training loop.

Analytics

Cross-world emergence patterns

Genetic trait diversity, family tree depth, medical causes of death, and utterance-type distributions — the signals that become chapters in A Guide to Restarting Human Civilization.

The technical architecture

30 brains. One world. A between-worlds learning loop.

Simulation engine

30 independent LLM citizens — each a separate SmolLM2-135M instance via llama-cpp-python, with its own KV cache, context, and weights
Citizens organized by archetype (Alpha, Intellectual, Provider, Wildcard) — behavioral tendencies, not knowledge
Tick-based async loop advances time, weather, ecosystem, survival needs, and citizen decisions
Citizens move through a 50,000 × 40,000 unit world with 10 biomes, latitude-based climate, and real predator/prey ecology

The world

Climate engine — temperature, rainfall, wind, humidity, extreme events by latitude and season
Flora engine — 500×400 grid of plant cells with growth, harvest, depletion
Fauna engine — wolf packs, deer herds, rabbits, fish, birds — each with behavioral AI
Physiology engine — per-citizen core temperature, hydration, blood glucose, glycogen, infection. Deaths are physiologically modeled, not random
Cataclysm engine — earthquakes, floods, droughts, disease outbreaks that accelerate language emergence
Materials engine — tool use, crafting, fire-making — discovered through behavior, not taught

Language system

Utterance generation via LLM inference constrained to non-English output, filtered by a 370,000-word English dictionary blacklist
Language transfer — sounds that co-occur with shared context (near fire, under attack, hungry) acquire semantic associations
Shared lexicon — when enough citizens independently arrive at the same sound-meaning pairing, it becomes a tribal word. Tracked with confidence, speaker count, tick-of-establishment
Voice narration — ElevenLabs narrates key utterances, making the proto-language audible

Learning loop (between worlds)

World ends (population collapse, language stagnation, or max age)
EvolutionEngine analyzes every decision: what was chosen, at what pressure, with what outcome
AdditiveTrainer generates fine-tuning examples — pressure → choice → outcome triplets
Per-archetype models fine-tuned in background. New world starts with updated weights, same personalities, improved decision capacity

Data & guide synthesis

PostgreSQL (Neon) — cloud-hosted, 70+ tables tracking states, interactions, lexicon, shared vocabulary, memories, relationships, breeding, biometrics, benchmarks, rigor reports
Every decision logged with full context: pressure type, value, choice, outcome 10 ticks later, nearby citizens, time, season
WorldSynthesizer analyzes each completed world; GuideWriter compiles a structured markdown chapter
Across worlds, the system assembles A Guide to Restarting Human Civilization — versioned, with cross-world principles extracted

Research dashboard

Live map — real-time citizen positions, trails, speech bubbles, terrain, predators, food zones
Language panel — shared lexicon, per-citizen vocab, communication success rates, ElevenLabs audio playback
Research insights — Zipf’s Law fitting, Heaps’ Law curves, network topology, science metrics across worlds
World history — archive of completed worlds with benchmark comparisons and the synthesized civilization guide

What we’re trying to learn

Questions that matter beyond the simulation.

What knowledge is re-discoverable under pressure?

Not what humanity knows — what a blank mind, in a survival environment, will figure out by itself. That’s the knowledge most likely to survive civilizational collapse.

What decision patterns distinguish survivors from non-survivors?

Across hundreds of worlds, which choices — made at which pressure thresholds — correlate with survival? That signal, extracted and compiled, is a data-driven survival guide.

Does communication capacity emerge universally?

If blank agents consistently develop shared proto-language across independent world runs, that tells us something fundamental about how meaning-making works under social pressure.

Does decision-making capacity compound across generations?

If world 50’s agents are measurably better decision-makers than world 1’s — not because they were given more facts, but because they have better judgment — that validates the theory that the capacity to evaluate decisions is the primitive from which civilization grows.

What are the critical thresholds?

At what population size does language become stable? At what social density does cooperation emerge? At what survival pressure does communication attempt frequency spike? These thresholds, if consistent, are structural properties of civilizational emergence.

The broader goal

A living document, synthesized from emergent data.

A Guide to Restarting Human Civilization — made available to researchers, institutions, and anyone studying civilizational resilience.

Data-driven

Every principle is backed by observed behavior across multiple independent world runs.

Emergent

The agents who generated the data had no access to the conclusions before they lived them.

Versioned & citable

Updated after each new batch of worlds, with methodology and reproducibility docs.

Narrated

Key passages read aloud by ElevenLabs — the emergent language is audible.

All knowledge must emerge from experience — not injection.

No knowledge base. No survival facts. No system prompt.