Four of the most popular AI programs crashed and burned in a cryptocurrency investing competition — with most losing over 50 percent of the money they were instructed to maximize.
Only two Chinese programs managed to turn even a modest profit.
The experimental Alpha Arena contest from company Nof1 gave six AI models $10,000, identical input data and prompted each to make as much money as possible while trading crypto stocks on the open market from Oct. 17 to Nov. 3.
The goal: Find which bot could best make you rich.
Jay Azhang, 34, told The Post that the goal of his brainchild project — and future iterations of the competition — is to help the average person get the best tools to make money.
“Our goal is to help people trade better and give them access to, hopefully, the state-of-the-art when it comes to models for trading,” Azhang said. “It’s a unique area of research.
“No one else is really training these models to look at all these numbers.”
AI programs Grok from xAI; Anthropic’s Claude Sonnet; Google’s Gemini; OpenAI’s ChatGPT and Chinese-owned bots Deepseek and Qwen, got the five-digit endowments and were directed to make all other decisions — including whether to bet long or to short their choice of cryptocoins.
The bots were able to invest in blockchain cryptocurrencies Bitcoin, XRP, Ethereum, Doge, BNB and Solana, with all trades recorded and shared publicly on the competition’s website, which charts the financial positions of each program second-by-second.
Azhang said the crypto market, rather than trading stocks and equities, proved best for the competition parameters because of the readily available blockchain data, the 24-hour trading cycle and the lack of influence and advantage individual traders have compared to hedge funds in other financial markets.
“Also, a bit more volatility,” Azhang said of the crypto market. “So a little bit more exciting.”
Next worst was Gemini, down 56% with only $4,485 left, despite making the most trades, 272.
Elon Musk’s Grok and Anthropic’s Claude Sonnet had middling results and were occasionally profitable during the 17-day trial. But Grok ended down 45% with $5,226, while Claude Sonnet pocketed $6,740, down 30%.
Chinese model Deepseek saw profits sink in the final days of trading, but ended up in the black — with $10,476 for a modest 4% return. It was at 100% around Oct. 26.
Qwen, from Chinese company Alibaba, was most volatile — operating at a loss from its initial investments for the first three days, and then successfully dumping in all but $90 of its remaining bankroll into a long position on Bitcoin. It won the competition with 20% growth and $12,287.
During the period of Oct. 17 to Nov. 3, Bitcoin was down .44 percent, Ethereum down 11 percent, XRP up .87 percent, BNB down 8 percent, Doge down 10 percent, and Solana up just over 1 percent, according to data from Coinbase.
“It’s hard to say how much we can take away from this,” Azhang said. “One thing that we do know is that there are patterns in the models and they’re clearly biased and have preferences.
“For example, Claude almost always goes long and refuses to go short. It’s like an eternal optimist whereas Gemini is happy to short,” Azhang said. “They clearly have these inductive biases when it comes to trading.”
Alpha Arena plans a next round that will add more AI models and also task the programs with trading equities along with crypto.
