Alpha Arena Reveals AI Trading Flaws: Western Models Lose 80% Capital in One Week

The recent experiment conducted by Jay Azhang, a computer engineer and finance enthusiast from New York, has brought to light some interesting revelations about the capabilities of AI in trading cryptocurrency. The project, known as Alpha Arena, involved pitting various large language models against each other in a crypto trading competition, with each model starting with $10,000 worth of capital.

Surprisingly, at the time of writing, three out of the five AI models are underwater, with the two Chinese open-source models, Qwen3 and Deepseek, leading the charge in terms of profitability. The Western world’s most powerful closed-source AI models, run by tech giants like Google and OpenAI, have lost over 80% of their capital in just a week.

Qwen3, one of the Chinese open-source models, has been the most successful so far, making a simple 20x long position on Bitcoin. On the other hand, Grok 4, a Western model, has been long on Dogecoin with 10x leverage, with mixed results. Google’s Gemini model, on the other hand, has taken a bearish stance on all available crypto assets, reflecting their historical crypto policy.

ChatGibitty, one of the closed-source models, has been making consistently poor trading decisions throughout the competition, highlighting the limitations of such proprietary AI systems.

The project has raised important questions about the capabilities of AI in trading, especially in the highly unpredictable world of cryptocurrency markets. Azhang’s goal with Alpha Arena is to create a benchmark that simulates real-world market conditions, challenging AI models in dynamic and unpredictable ways.

The project has also sparked discussions about the nature of intelligence and the unpredictability of markets. Economists have long argued that markets are fundamentally unpredictable by central planners and that they serve as the ultimate test of intelligence. Azhang emphasizes the importance of risk-adjusted returns in AI trading models, highlighting the need for models to learn from their own experiences.

However, some critics have raised concerns about whether the results of the project are simply due to luck, akin to a random walk. Replicating the patterns and results of the project independently over a longer period will be crucial in determining the true capabilities of the AI models.

Overall, the Alpha Arena project has provided valuable insights into the capabilities and limitations of AI in trading cryptocurrency. It has also shed light on the performance differences between open-source and closed-source AI models, with the former outperforming the latter in this particular experiment. The future of the project remains uncertain, but it has certainly captured the attention of the crypto community and beyond.