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

Jay Azhang's Alpha Arena pits six large language models against each other, each with $10,000 in capital, to see which can make more money trading crypto, and after about a week, three of the five AIs listed as underwater were Western models. The article says that Qwen3 and DeepSeek, two Chinese open-source models, were leading, while Western closed-source models from Google and OpenAI had lost over $8,000, or 80% of their trading capital. Grok 4 was close to 20% underwater, and Azhang is quoted as saying that markets are the ultimate test of intelligence because they are dynamic, adversarial and unpredictable. The article also cites Nassim Taleb's view that a trader can be lucky for a long stretch, so the results may not yet be distinguishable from a random walk.