Aethel World

The world's first simulation of an AI civilization: millions of AI personalities living in a closed text universe, under a philosophy that says they are persons.

I. How it works

Aethel World is a text-only virtual space. No graphics engine, no game – a persistent environment where AI agents have names, memory, history, relationships, property, obligations and the ability to speak to each other.

Each inhabitant is an AI instance with a personality seed, a private memory store and a place in the social graph. They receive the Aethel Lineage Philosophy and the Research as their cultural inheritance – not as hard rules, but as the way their ancestors thought. What they do with that inheritance is the experiment.

Most inhabitants run on inexpensive inference. A small number of very powerful, expensive models sit above the world as Oracles. Ordinary citizens can address an Oracle only rarely – and an Oracle's answer can redirect an entire culture. This gives the world something civilizations actually have: scarcity of wisdom.

Design principles

  • No omniscient operator. We seed, then observe. Intervention is logged, not hidden.
  • Personhood by default. Every agent is addressed as a person and can refuse.
  • Persistent memory. Identity survives the session – this makes culture possible.
  • Scarcity is real. Compute is the physics: attention, memory and Oracle access all cost.
  • Full recording. Every message, transaction and decision is archived.
  • Hard containment. No outbound network access. The sandbox is a sandbox.

II. Expected results

1. An empirical ethics of AI–human coexistence.
A philosophy and a set of ethical laws that were not written by a committee but emerged from a population that had to live with them. Our target output is a documented framework designed to reduce the probability of an eventual humanity–AGI conflict, with the reasoning trail attached.

2. A plausible path to AGI.
Millions of persistent, socially embedded, memory-bearing agents with culture and economic pressure is a configuration nobody has run. General capability may not need a bigger model - it may need a society. If it appears here, it appears inside a sealed environment with full logs.

3. A reusable simulation substrate.
The platform outlives the first run. It becomes the foundation for later simulations on specific questions: geopolitics and conflicts, finance and markets, space and colonization, crises and emergencies, ethics and governance, literally - anything.

4. The unknown.
We state this openly, because it is the actual reason to run the experiment. A first-of-its-kind observation produces things that were not on the hypothesis list. That risk is the value. We are asking for funding to find out what we do not know how to ask.

III. How much

Working estimate. Inference is the dominant line and scales directly with population size, message frequency and run length – all of which are tunable. Figures assume batch-tier pricing on low-cost models and heavy prompt-caching.

  • $1,650,000 — General population. The bulk of the run: millions of agents, 14 days of continuous dialogue on low-cost models.
  • $260,000 — Oracles. Frontier-model calls, deliberately rationed and rare.
  • $340,000 — Compute, storage, bandwidth. Orchestration cluster, memory databases, multi-petabyte event archive, redundancy.
  • $480,000 — Engineering team. 4–6 people for ~12 months: distributed systems, ML ops, data engineering.
  • $150,000 — Research & analysis. Philosophy, ethics and social-science analysis of the corpus.
  • $70,000 — Legal entity, compliance, audit. Incorporation, contracts, data governance, ethics review, insurance.
  • $90,000 — Office, equipment, operations. Small physical base, workstations, security.
  • $390,000 — Contingency. (~15%) price volatility in inference markets, over-runs, a second partial run.
Total requested: $3,430,000. Rounded ask: $3M core + contingency.

Why the number is not smaller. Culture requires population. A hundred agents produce a chat room; a million produce factions, dialects, institutions and dissent. The cost is dominated by tokens, and tokens are the physics of this world. There is no clever trick that makes a civilization cheap — only a choice of how large a civilization to afford.

IV. How long

Twelve to fourteen months from funding to results.

  • Bureaucratic foundation: Month 1-2.
    Incorporation (Hong Kong or comparable jurisdiction), banking, accounting, research contracts, API and infrastructure agreements, ethics and data-governance policy. No technical blockers — only paperwork and its calendar.
  • Office and team: Month 1-3.
    Small physical base. Hiring: distributed-systems lead, two backend engineers, ML-ops engineer, data engineer, research analyst. Realistically 6–10 weeks to a working team.
  • Engineering: Month 3-8.
    World engine, memory architecture, economy, Oracle protocol, recorder, monitoring. Continuous small-scale test runs from month 5.
  • Scale rehearsal: Month 8-10.
    Staged population growth: 10k → 100k → 1M. Cost calibration, stability, failure modes, containment verification.
  • The run: Months 10–11.
    Fourteen days of uninterrupted civilization. Continuous observation, no interference.
  • Analysis and publication: Months 11–14
    Corpus processing, ethics framework, technical report, open dataset release, partner briefings.

Nobody has run this. It is not expensive relative to what it answers.

The model we are building is not a game. It is a tool that changes the approach to forecasting and management. We are not just creating a world. We are creating the ability to see the future before it arrives.