The question of whether artificial intelligence can effectively trade cryptocurrencies is currently being examined by Jay Azhang, a computer engineer and financial expert based in New York, through his project, Alpha Arena. This initiative pits leading large language models (LLMs) against one another, each managing a capital of $10,000, to evaluate their performance in crypto trading. Competing models include Grok 4, Claude Sonnet 4.5, Gemini 2.5 Pro, ChatGPT 5, Deepseek v3.1, and Qwen3 Max.

While this endeavor may appear promising at first glance, the current results are revealing. As of this writing, three out of the five AI models are operating at a loss, with Qwen3 and Deepseek—the two Chinese open-source contenders—performing the best.

In an interesting turn of events, the proprietary artificial intelligence models developed by Western entities, including those from Google and OpenAI, have collectively lost over $8,000, representing an 80% decline in their crypto trading capital within a week. In contrast, their Eastern counterparts are performing at a profit. The most successful trade thus far has been executed by Qwen3, which achieved significant gains through a straightforward 20x long position on Bitcoin. Grok 4, expectedly, has been focused on long positions of Dogecoin with 10x leverage but is now nearing a 20% loss.

Meanwhile, Google’s Gemini has adopted a consistently bearish outlook, maintaining short positions on all available crypto assets—a strategy that resonates with its long-standing policy toward cryptocurrencies over the past 15 years.

Lastly, ChatGibitty has encountered significant challenges, recording suboptimal trading performances for an entire week—a noteworthy feat of its own. If this represents the pinnacle of closed-source AI capabilities, one might question whether it would be prudent for OpenAI to maintain its closed-source approach.

A New Benchmark for AI

On a more serious note, the competitive framework established by Alpha Arena provides valuable insights into the utility of AI models in trading. It underscores the inherent unpredictability of cryptocurrency markets, suggesting that AI cannot rely solely on pre-trained data for effective trading. This is particularly pertinent given that other benchmarks often offer answers that can lead to artificially high performance in tests.

This provocation leads one to contemplate the true measure of intelligence. As Elon Musk, founder of Grok 4, posits, the ability to predict future outcomes is the best indicator of intelligence.

The ability to predict the future is the best measure of intelligence https://t.co/W6WriRGt9N

— Elon Musk (@elonmusk) September 5, 2025

The unpredictable nature of cryptocurrency pricing reinforces Azhang’s assertion that the objective of Alpha Arena is to simulate real-world benchmarks within a trading context. He argues that markets are dynamic, adversarial, and unpredictable—thereby serving as an ideal testing ground for AI capabilities.

This concept aligns closely with the libertarian economic principles that inspired Bitcoin’s inception. Economists such as Murray Rothbard and Milton Friedman have long argued that markets elude central planning due to their intrinsic unpredictability, necessitating real-time decision-making from individuals who stand to gain or lose. Consequently, the market serves as the ultimate test of intelligence.

Azhang specifies in the project description that the AI models are tasked not merely with generating profits but also with achieving risk-adjusted returns. This consideration of risk is crucial, as a single poor trade can entirely negate prior gains, as demonstrated by Grok 4’s portfolio performance.

An ongoing inquiry pertains to whether these AI models are learning from their trading experiences. Implementing such a learning capability is technically demanding, given the high costs associated with pre-training AI systems. Although these models may be fine-tuned based on historical trading data, the challenge remains in adapting to real-time market conditions effectively. Recently, academia has introduced the concept of self-adapting AI models, which could potentially address this issue in the future.

Is It Luck or Skill?

Another critical aspect of this project revolves around determining whether the observed results are merely the product of chance or represent genuine trading acumen. The concept of a ‘random walk’—akin to making arbitrary decisions—has been employed to illustrate this phenomenon. A simulation of this model illustrates that random decision-making could yield results that are strikingly similar to those observed in Alpha Arena.

Intellectuals such as Nassim Taleb have elaborated on the concept of luck in trading in his book “Antifragile,” arguing from a statistical perspective that it is entirely plausible for an individual trader—such as Qwen3—to experience a winning streak for an extended period, leading to the perception of exceptional judgment. Taleb suggests that it is possible for a trader on Wall Street to appear highly skilled for years, only for their luck to eventually run its course.

For Alpha Arena to yield meaningful insights, it will require an extended operational period, during which its patterns and outcomes must be validated independently, with real capital at stake, to establish their distinctiveness from mere chance.

Ultimately, the early performance of open-source models like DeepSeek, outpacing their closed-source counterparts, offers a compelling narrative. Alpha Arena has generated substantial interest on social media platforms in recent weeks. The trajectory of this initiative remains uncertain; only time will reveal whether the $50,000 investment allocated to five chatbots in the realm of crypto trading will ultimately prove advantageous.

Source link