AI“角斗场”实盘大赛落幕，阿里千问夺冠, GPT-5亏麻了, Gemini成“末日空头”

Core Insights - The article highlights the performance of AI models in a real-world investment competition, with Alibaba's Qwen achieving a 22.32% return, while top American models like OpenAI's GPT-5 and Google's Gemini 2.5 Pro suffered significant losses of 62.66% and 56.71% respectively [3][24]. Group 1: Competition Overview - The "Alpha Arena" competition, initiated by the American AI research lab Nof1, aimed to test AI models' decision-making abilities in a chaotic and dynamic environment, contrasting with traditional academic benchmarks [6][32]. - Six leading AI models participated, including Alibaba's Qwen3-Max and DeepSeek, alongside OpenAI's GPT-5 and Google's Gemini 2.5 Pro [7][8]. Group 2: Performance Analysis - Qwen and DeepSeek emerged as the only two profitable models, while the four American models incurred losses [31]. - Qwen's strategy involved a straightforward long position on Bitcoin, demonstrating strong conviction in a high-volatility market [16][30]. - DeepSeek adopted a similar bullish strategy, utilizing high leverage [15]. Group 3: Trading Strategies - The competition revealed three distinct trading camps: - Eastern Winners: Qwen and DeepSeek, both employing clear bullish strategies [14]. - Lost Geniuses: GPT-5 and Gemini, which consistently lost due to poor decision-making and excessive caution [17][18]. - Observant Players: Grok and Claude, which displayed unique and less effective trading strategies [19][20]. Group 4: Key Takeaways - Qwen's victory was attributed to its effective risk management and timely defensive actions, particularly in the competition's final moments [22][30]. - The competition underscored the disparity between academic intelligence and practical market decision-making, with Qwen and DeepSeek exemplifying successful strategies in real-world conditions [28][32].