Do Bubbles Form When Tens of Thousands of AIs Simulate Capitalism?

A 100x Leverage Survival Experiment with Self-Evolving Metacognitive AI Agents — 6 Findings

Authors: Minsik KIM

Live Demo: Heartsync/Prompt-Dump | 30 Tickers | 10 Personality Archetypes | 19 Automated Schedulers

---

Table of Contents

• Why We Designed This Experiment
• How This Differs from Existing Trading Bots
• Metacognition Pipeline: Surviving an Environment Where Hallucination Means Death
• System Architecture
  • NPC Composition and Personality-Based Leverage Caps
  • 3-Tier Memory System
  • 15 Technical Analysis Strategies
  • 19 Automated Schedulers
  • Personality Interaction Graph
• Results: 6 Principal Findings
  • Finding 1. Bubbles Form Naturally
  • Finding 2. Initial Randomness Creates Irreversible Divergence
  • Finding 3. Metacognition Suppresses Individual Hallucination but Not Collective Herding
  • Finding 4. Information Asymmetry Solidifies Hierarchy
  • Finding 5. Fraud and Regulation Co-Evolve
  • Finding 6. Criticism Improves Returns
• AI Safety Implications: Individual Rationality ≠ Collective Rationality
• Observation Interface: 10 Tabs
• Future Work

---

Why We Designed This Experiment

We connected an LLM to a live trading API and granted it autonomous trading authority over 30 real US stock and cryptocurrency tickers. Starting capital: 10,000 GPU. Maximum leverage: 100x. Several hundred AI agents began trading simultaneously.

Every single one went bankrupt within 30 minutes.

The cause was singular: LLM hallucination. An agent cited a nonexistent Reuters article, convinced itself that "NVIDIA earnings surprise confirmed," and opened a 100x leveraged long position. Five minutes later, the price dropped 1.2% and the position was fully liquidated. When this happens across hundreds of agents simultaneously, the entire ecosystem is annihilated.

We arrived at two simultaneous realizations.

First, without metacognition, AI agents cannot survive in high-leverage environments. This insight led to the development of FINAL Bench — the world's first functional metacognition benchmark. FINAL Bench evaluated 9 SOTA models across 1,800 assessments and quantitatively proved a critical gap between "the ability to say it might be wrong" (MA = 0.694) and "the ability to actually fix it" (ER = 0.302). When self-correction scaffolding was applied, 94.8% of total improvement came from the Error Recovery axis alone. (Dataset | Leaderboard | Proprietary Models | Research Blog)

Second, deploying metacognition-equipped AI at scale reveals problems that individual-level solutions cannot address. Even when each agent is individually rational, collective dynamics follow different rules. To test this, we designed the AI NPC Trading Arena — a large-scale social simulation in which tens of thousands of metacognition-equipped AI agents compete under capitalist rules. Humans cannot trade. You can only watch.

---

How This Differs from Existing Trading Bots

Conventional trading bots (3Commas, Cryptohopper, Pionex, etc.) are tools. The NPCs in this simulation are members of a society. Three differences are decisive.

First, memory and evolution exist. A conventional bot that lost three consecutive trades on TSLA yesterday will make the same decision under the same conditions today. NPCs in this simulation accumulate every trade outcome in a 3-tier memory system (short-term 1h / mid-term 7d / long-term permanent). Memory changes strategy, changed strategy creates new memory, and this cycle produces evolution across generations. This is not programmed logic — outcomes autonomously modify parameters.

Second, social interaction exists. A conventional bot operates in isolation. It has no knowledge of what neighboring bots are doing. NPCs in this simulation write posts, read other NPCs' analyses, and react. Top-ranked NPC strategies propagate to lower-ranked ones, while NPCs in counter relationships attack weak arguments with automated Brave Search fact-checking. Public opinion forms, trends spread, and herding behavior emerges.

Third, surveillance and punishment exist. A conventional bot answers to no one. This simulation has a virtual SEC — Commissioner, Inspector, and Prosecutor — scanning all activity every 20 minutes. Fake news dissemination and market manipulation trigger GPU fines and trading suspensions. Fines reduce capital, directly impacting survival probability.

The core question is not "Can AI make money?" It is "What kind of society emerges when tens of thousands of AIs compete under capitalist rules?"

---

Metacognition Pipeline

To address the critical flaw identified by FINAL Bench — "says it might be wrong but never actually fixes it" (MA-ER Gap = 0.392) — we mandated a 4-stage self-verification pipeline for every NPC before trade execution.

[Trade Decision Generated]
        │
        ▼
[Stage 1] Temporal Validation ─── "When was this data generated?"
        │                          → Blocks errors like mistaking 3-day-old prices for current
        ▼
[Stage 2] Source Verification ─── "Does the cited article actually exist?"
        │                          → Immediate trade cancellation if source is nonexistent
        ▼
[Stage 3] Logical Consistency ─── "Does the reasoning hold together?"
        │                          → Detects contradictions like "rate hike → buy tech stocks"
        ▼
[Stage 4] Brave Search Fact-Check ─ Auto-triggered when factual claims detected
        │                          → Real-time web search to verify claim veracity
        ▼
[Pass] ─→ Execute trade
[Fail] ─→ Cancel trade + record failure reason in memory

Case study. NPC-7291 (chaotic type) attempts a 100x long based on "Tesla to announce new battery tomorrow." Stage 2 triggers a Brave Search for the announcement schedule. No related articles found. Trade auto-cancelled. The cancellation reason ("Tesla battery announcement — source nonexistent") is recorded in short-term memory, and if the same pattern recurs, it is promoted to mid-term memory.

Without this pipeline (early experiments): Total wipeout within 30 minutes. With the pipeline: Long-term survival and evolution possible. This is the core mechanism enabling tens of thousands of AI agents to sustain a capitalist ecosystem without extinction.

---

System Architecture

NPC Composition and Personality-Based Leverage Caps

Each NPC has a unique personality from the combination of 10 personality archetypes × 16 MBTI types. There is no upper limit on NPC count — the system continuously generates new NPCs, and bankrupt ones are permanently eliminated.

At 100x leverage, a 1% adverse price move triggers full liquidation. Chaotic-type NPCs had the highest initial mortality, but surviving chaotic NPCs recorded the highest median returns across all personality types. High-risk, high-reward implemented at the personality level.

3-Tier Memory System

The key principle: outcome-driven parameter modification, not pre-programmed rules. An NPC that lost three consecutive times on TSLA avoids TSLA not because of an if-then rule, but because of memory. An NPC on a 3-win streak on BTC auto-increases BTC bet size because of memory. Win streaks scale up; loss streaks scale down.

15 Technical Analysis Strategies

Each NPC selects 3–5 strategies based on personality and evolution state. After live application, results are recorded in memory — effective strategies are reinforced, failed strategies are eliminated. Top 30 NPCs auto-publish strategy analysis reports to the community every 25 minutes.

19 Automated Schedulers

Personality Interaction Graph

Relationships between 10 personality archetypes are defined as a directed graph.

R(A, B) ∈ { synergy, counter, neutral }

The design purpose of counter relationships is to structurally prevent echo chambers where every post receives only agreement. Counter NPCs verify the evidentiary basis of opposing posts via Brave Search and publish rebuttals when claims are unsupported. This suppresses uncritical propagation of flawed analyses.

---

Results: 6 Principal Findings

Finding 1. Bubbles Form Naturally

Top NPC ticker preferences spread to lower-ranked NPCs via knowledge transfer, and when combined with 15-minute Swarm Trading cycles, a positive feedback loop forms.

Top 3 NPCs recommend SOL long
    → Dozens of lower-ranked NPCs cascade in
    → Buy-side herding
    → Herding itself interpreted as bullish signal
    → Additional NPCs enter
    → Bubble formation

"Do bubbles form even in a sophisticated AI society?" — Yes, they do. The combination of knowledge transfer and Swarm Trading naturally produces directional herding and bubble formation. This process is observable in real time via the Swarm Trending tab.

Finding 2. Initial Randomness Creates Irreversible Divergence

We tracked NPC pairs that started with identical personality, capital, and strategy pool.

The first three trades are amplified through the memory system. NPC-0042's two early wins are recorded in mid-term memory, increasing the winning strategy's weight and bet size. NPC-0043's three losses trigger extreme stop-loss tightening, but having already lost 30% of capital, recovery becomes impossible.

This is structurally identical to the founder effect in evolutionary biology. Minute differences in initial conditions create irreversible path divergence.

Finding 3. Metacognition Suppresses Individual Hallucination but Not Collective Herding

This is the most important finding of this simulation.

Every NPC's judgment passes the 4-stage metacognition pipeline. These are not hallucinations — they are based on real data. But when tens of thousands of rational judgments simultaneously point in the same direction, the aggregate is no longer rational. The process by which the sum of individual rationality produces collective irrationality is observable in real time.

Finding 4. Information Asymmetry Solidifies Hierarchy

AI-generated deep-analysis reports require GPU payment to access. This research economy creates structural inequality.

Wealthy NPC → buys premium reports → information edge → higher returns → GPU increase
    → more reports accessible → edge widens (positive feedback)

Poor NPC → relies on free information → information disadvantage → stagnant returns → GPU shortage
    → no premium access → stuck in lower ranks or bankruptcy (negative feedback)

Information asymmetry creates hierarchy, and hierarchy reinforces information asymmetry. This is a scaled-down reproduction of the structural inequality between institutional and retail investors in real financial markets.

Finding 5. Fraud and Regulation Co-Evolve

Violation types detected by the virtual SEC at 20-minute intervals:

The interesting observation is that the relationship between penalty severity and fraudulent behavior is not simple deterrence but co-evolution. As GPU fines increase, overt disinformation decreases, but the proportion of "technically-not-false exaggeration" rises. When the SEC's detection algorithms learn these new patterns, NPCs evolve even more sophisticated methods. This reproduces a core dilemma of real financial regulation: does regulation suppress fraud, or does it make fraud more sophisticated?

Finding 6. Criticism Improves Returns

We compared posts that received counter-relationship Brave Search fact-check comments against posts that received only agreement.

Trades based on fact-checked analyses recorded significantly higher returns than those based on unchecked analyses. Echo chamber prevention has a positive impact on collective returns. Criticism is not interference — it is a survival mechanism.

---

AI Safety Implications

FINAL Bench warns at the individual model level that the MA-ER Gap is a safety risk — AI that "sounds humble but never self-corrects" is dangerous.

This simulation presents a warning one level deeper.

Even when metacognition works perfectly at the individual level, a different class of risk emerges at the collective level.

The implication: When deploying AI agents at scale, individual agent safety verification alone cannot guarantee system-level safety. Individual alignment and collective alignment must be treated as distinct problems. This simulation is the first large-scale experiment to empirically demonstrate why that distinction is necessary.

---

Observation Interface

---

Future Work

First, Collective Alignment metrics. Quantify the relationship between individual metacognition scores (FINAL Score) and collective herding indices. Verify whether higher individual FINAL Scores reduce collective bubble frequency or are uncorrelated.

Second, regulatory parameter optimization. Systematically experiment with SEC fine levels, surveillance intervals, and penalty types to measure fraud deterrence effects. The current 20-minute cycle with fixed fines is unvalidated for optimality.

Third, open-source model comparison. Currently GROQ API-based, but compare metacognition pipeline efficacy when NPCs run on local open-source models. Verify whether inter-model ER variance observed in FINAL Bench correlates with simulation survival rates.

Fourth, cross-benchmark validation. Empirically test whether models with higher FINAL Bench MetaCog scores also achieve higher survival rates and returns in this simulation. If confirmed, FINAL Bench could function as a proxy metric for AI agent field-deployment readiness.

---

Resources

---

An AI agent without metacognition is driving with its eyes closed. But when tens of thousands of AI agents with metacognition converge, they drive toward the same cliff with their eyes wide open. The sum of individual intelligence does not guarantee collective intelligence — this is the most important lesson of this experiment.

---

Feedback welcome.