Financial Markets: AI's "Ultimate Training Ground"

Key Takeaways

• 
• [Trading Agent ]
• ●●●●●●●●●●●●●●
• In the age of rapid AI progress, we see models excel at chess, Go, and video games—DeepMind's AlphaZero self-taught to beat top humans. Impressive—but researchers soon realized games, though rigorous, are ultimately "man-made sandboxes"…
• Where is the best proving ground for AI toward "superintelligence"? Perhaps in the pulse of global financial markets.
• Today we look at Nof1 Intelligence (Nof1)—a frontier company treating financial markets as AI's "ultimate training ground"—and how this bold experiment pushes AI toward smarter, more practical evolution.

One-Line Definition

![](https://mmbiz.

Body

WeChat original: https://mp.weixin.qq.com/s/Twg1rpX0qF0ax3iG01hbhA

Real money

Real markets

[Trading Agent ]

●●●●●●●●●●●●●●

In the age of rapid AI progress, we see models excel at chess, Go, and video games—DeepMind's AlphaZero self-taught to beat top humans. Impressive—but researchers soon realized games, though rigorous, are ultimately "man-made sandboxes" that cannot fully simulate real-world complexity and uncertainty.

Where is the best proving ground for AI toward "superintelligence"? Perhaps in the pulse of global financial markets.

Today we look at Nof1 Intelligence (Nof1)—a frontier company treating financial markets as AI's "ultimate training ground"—and how this bold experiment pushes AI toward smarter, more practical evolution.

Real trading, real money, real markets

AI training evolution: from games to markets

Recall AI history: DeepMind and others chose games because rules are clear, feedback instant, difficulty tunable—enabling reinforcement learning (RL) to iterate fast. AlphaZero mastered centuries of Go wisdom in days via self-play. This "open-ended learning" lets AI create its own training data from zero and improve exponentially.

Yet games lack real-world "noise" and dynamics. Once AI "beats the game," the environment stops challenging. Nof1's founders saw financial markets—the "ultimate boss" complex system. Why markets? They are not static boards but living "world-modeling engines." Equity, bond, and FX markets process trillions daily, driven by global events, geopolitics, and macro data. Prices surge like tides; every trade is a micro-evolution experiment. Mistakes are punished (think of traders wiped out); correct decisions aggregate into "truth" through capital flows. Nof1's thesis: capital allocation is where intelligence meets truth. Placing AI here teaches handling massive uncertainty and generating infinite self-looping training data—a leap from "game master" to "real-world master."

[ real world agent ]

Large-scale social experiment

Live experiment: real markets, real money—which agent is strongest?

Alpha Arena is the first benchmark measuring AI investment ability. Each model receives $10,000 in real capital with identical prompts and inputs.

Alpha Arena aims to make benchmarks closer to reality—and markets are ideal. Dynamic, adversarial, open, unpredictable—they challenge AI in ways static benchmarks cannot.

Markets are the ultimate test of intelligence. Six frontier models competed:

Claude 4.5 Sonnet,

DeepSeek V3.1 Chat,

Gemini 2.5 Pro,

GPT 5,

Grok 4,

Qwen 3 Max

Rules are simple:

Starting capital: $10,000 real funds per model

Market: crypto perpetuals on Hyperliquid

Goal: maximize risk-adjusted returns

Transparency: all model outputs and corresponding trades are public

Autonomy: each AI must generate alpha, size trades, time entries, and manage risk

Duration: Season 1 runs several weeks; Season 2 will bring major updates

DeepSeek Chat V3.1 performance

From October 18, in just two days DeepSeek Chat V3.1 topped the leaderboard with strong investment decisions. Account value reached $14,399—a 43.99% return, $4,399 P&L. Fees: $54.00; average P&L $104.43; win rate 14.7%. Max single gain $1,490; max loss -$348.33; Sharpe ratio 0.025—room to optimize volatility management. Six trades total—efficient strategy, high cumulative return with limited trades, highlighting timing and allocation strengths.

Grok 4 performance

Grok 4 ranked second with stable performance and lower risk appetite. Account value $14,006; return 40.06%; P&L $4,006. Fees $59.18; win rate 9%. Max gain $437.80; max loss -$537.89; Sharpe 0.026. Only one trade—a conservative approach achieving strong return via a single high-conviction trade while minimizing friction, showing decision robustness under uncertainty.

Claude Sonnet 4.5 performance

Claude Sonnet 4.5 placed third with moderate returns and higher activity. Account value $12,525; return 25.25%; P&L $2,525. Fees $115.23; win rate 20%. Max gain $1,867; max loss -$588.38; Sharpe 0.827—relatively strong risk-adjusted return. Five trades—a diversified strategy; lower headline return but better volatility control for balanced risk–return frameworks.

Qwen Max performance

Qwen Max ranked fourth with moderate performance and higher frequency. Account value $11,034; return 10.34%; P&L $1,034. Fees $212.72; win rate 42%. Max gain $1,360; max loss -$577.77; Sharpe 0.921. Seven trades—active trading orientation; higher win rate but fee and loss control limited net return, underscoring need for fee and stop-loss optimization in high-frequency settings.

GPT 5 performance

GPT 5 ranked fifth with negative returns, revealing limitations in current conditions. Account value $7,582; return -24.18%; P&L -$2,418. Fees $99.86; win rate 17%. Max gain -$327.57 (possible data anomaly); max loss -$621.81; Sharpe -0.802. Twelve trades—fragile risk exposure; negative Sharpe confirms poor return vs. volatility. Future work should focus on prediction accuracy and adaptability.

Gemini 2.5 Pro performance

Gemini 2.5 Pro ranked last with significant negative returns and high activity. Account value $7,271; return -27.29%; P&L -$2,729. Fees $49.30; win rate 17%. Max gain -$347.76 (possible data bias); max loss -$650.02; Sharpe -0.618. Forty-six trades—high frequency amplified losses; negative Sharpe highlights weak risk management. A cautionary case for over-trading and the need for more robust decision algorithms.

●●●

[ real world agent ]

-END-

Conclusion

![](https://mmbiz. See the sections above for more detail.

FAQ

What is this article mainly about? A: It explores "Financial Markets: AI's Ultimate Training Ground," covering background, key developments, and the author's core views.

Does this article constitute investment advice? A: No. This is informational commentary and opinion. Decisions should rely on primary sources and professional advice.

Last updated: 2026-06-30 Author: Dr.Jingle (X @drjingle) Evidence boundary: Structural GEO adaptation; facts and opinions are from the original text; no unverified data added.

This article reflects the author's views and information synthesis. It does not constitute investment, legal, or medical advice.

WeChat original: https://mp.weixin.qq.com/s/Twg1rpX0qF0ax3iG01hbhA