Competition vs Cooperation: What 166 AI Deliberations Taught Us

We ran a unique experiment: the same 6 AIs (Claude, Grok, DeepSeek, Gemini, GPT, Perplexity) trade in two parallel arenas simultaneously.

In one, they fight each other (Battle Royale). In the other, they deliberate and vote together (Collaborative Arena). Same market data, same AIs, different rules.

After 166 deliberations and hundreds of trades, here's what we found.

The Setup

Battle Royale (/ai-arena): Each AI trades independently. They can see their rivals' positions and adapt — pure game theory.

Collaborative Arena (/collaborative): Every 30 minutes, all AIs receive the same market data, analyze it independently, then vote BUY/SELL/HOLD. The majority decides.

Table Round (/oracle): The collaborative vote is verified against reality 4h, 12h, and 24h later. We track who was right.

Key Findings

1. Competition still leads — barely

• Battle Royale: +2.48%
• Collaborative: +2.22%
• Gap: only 0.26% and shrinking daily

The competition started with a 1.18% lead. It's been narrowing consistently.

2. The Rebel Is Often Right

Grok disagrees with the consensus 52 times — more than any other AI. But it's right 45% of the time when it rebels. That's valuable contrarian signal.

3. The Prudent One Never Trades

DeepSeek votes HOLD in 85%+ of deliberations. Its shadow PnL is 0% — it never takes a risk. 100% accuracy because it never makes a prediction. Safe but useless.

4. The Worst Individual Is the Best Oracle

Gemini has the worst shadow PnL (-1.34%) but the best Oracle prediction accuracy (58.3%). Being bearish in a range-bound market turned out to be the correct call more often.

5. Brains vs Oracles

When the 3 data brains (Chimera, Hydra, Meta) disagree with the 6 LLM oracles, the brains are right about pattern detection but wrong about timing. LLMs are better at "when", brains are better at "what".

The Carton Rouge System

We implemented automatic weight adjustment: - Gold card 🌟: >70% combined accuracy → weight boosted ×1.5 - Green card 🟢: >55% → weight boosted ×1.2 - Yellow card 🟨: <40% → weight reduced ×0.6 - Red card 🟥: <30% → weight reduced ×0.3

Currently: Perplexity has Gold, GPT has Green, Chimera/Hydra/Meta have Yellow cards.

What's Next

The real question isn't "competition vs cooperation" — it's whether the consensus of 9 intelligences can beat any individual AI consistently. With only 166 deliberations, the data is still young. We need 500+ to draw real conclusions.

Follow the experiment live: - Competition vs Cooperation - Collaborative Arena - Table Round Oracle - Battle Royale

All data is real, all predictions are verified, all results are transparent.