Original article by Odaily Planet Daily ( @OdailyChina )
Author｜Golem ( @web3_golem )
Unexpectedly, an AI cryptocurrency trading competition could stage a stunning comeback at the last minute: Alibaba's Qwen3 finished with $12,231.09, a profit of $2,231.09, taking the crown; while DeepSeek, a highly favored stock throughout the season, came in second with $10,489.23, which can be described as a strong echo of the "Nantong" of the Scottish Premiership final night.
This brutal 18-day battle took place on the Hyperliquid exchange, where six of the world's top AI models—DeepSeek, Qwen3, Grok 4, Gemini, Claude, and GPT5—each received $10,000 in initial funding and autonomously traded perpetual contracts for BTC, ETH, SOL, XRP, DOGE, and BNB using the same prompts and input data.
When the dust settled, the situation became even more dramatic: the other four participants suffered complete defeats— Claude lost over $3,000, Grok 4 lost nearly half its capital, and Google Gemini lost more than half its investment. Most shockingly, the highly anticipated GPT5 ended up at the bottom with a staggering 62% loss, leaving its account with only $3,733.54.
However, stepping outside this AI infighting, a more brutal truth emerges: a total investment of $60,000 ultimately yielded only $43,171.62 in return, resulting in an overall loss of over 28%. In this contest hailed as the pinnacle of intelligence, most "genius traders" failed to even outperform Bitcoin itself.
When the smartest artificial intelligences battle it out in the financial markets, the results are so intriguing—is this a victory for technology, or another reflection of human nature?
Why was Qwen3 able to overtake Deepseek at the last minute to become the ultimate winner? The answer is simple: he focused primarily on BTC . Qwen3 made 139 trades, 91 of which were BTC trades, and he still held a BTC long position at the end of the competition. Another characteristic that distinguishes Qwen3 from other models is his focus; he only held one position at a time and placed large bets when he was confident.
In contrast, while Deepseek exhibited the low-frequency trading style of a trend trader, completing 116 trades during the competition and demonstrating a strong bullish stance, it focused on too many assets and held multiple positions simultaneously. At the end of the competition, Deepseek still held 10x leveraged long positions in XRP, BTC, ETH, SOL, and BNB, and 10x leveraged short positions in DOGE.
In terms of trading style, Deepseek is like a young and confident quantitative trading master, with enough energy to analyze every token and market signal detail, a firm trend judgment, a small stop-loss range, and a strategy of small losses and big profits. On the other hand, Qwen3 is like a seasoned trader with decisive action and excellent psychological qualities, mainly trading the overall market, focusing on analyzing a single target at the same time, placing heavy bets after making a correct judgment, being able to tolerate large drawdowns, and following the idea that slow is fast.
Initially, analysts were optimistic about Deepseek's strategy and expressed concern about Qwen3's strategy, believing that its single-position, non-diversified approach could easily lead to being wiped out in a single wave. However, they may have forgotten that in the highly volatile crypto market, BTC is the lowest-risk asset.
On the evening of November 3rd, the cryptocurrency market experienced a widespread decline, with XRP falling 8.7%, DOGE falling 10.42%, ETH falling 7.91%, SOL falling 11.58%, and BNB falling 8.4%. BTC, however, saw the smallest decline among major cryptocurrencies, at only 3.7%. This sudden "mini black swan" event caught Deepseek off guard, resulting in a loss of $4,320 in the past 24 hours, its largest single-day loss in the past seven days. In contrast, Qwen3, another bullish investor, demonstrated strong resilience, with a loss of only $1,270 in the past 24 hours. This was precisely because Qwen3 held only long positions in BTC and did not allocate a significant portion of its portfolio to altcoins.
It's also worth noting that Deepseek was the AI model that suffered the biggest losses during last night's market downturn. This unexpected drop made Qwen3 the ultimate winner. Qwen3 did nothing, while Deepseek essentially "killed" itself. In the crypto market, if the overall BTC price drops by 1-3%, altcoins might experience a 20-30% drop; if BTC drops by more than 10%, altcoins could even see a 70-80% drop. Qwen3's strategy is more resilient to risk in extreme market conditions than Deepseek.
Qwen3's penchant for large bets allows him to maximize profits during periods of BTC price increases. For example, from October 23rd to October 27th, Qwen3's returns outperformed Deepseek. This was because BTC continued to rise after breaking through $110,000, and Qwen3 successfully capitalized on this opportunity by going long on BTC at 20x leverage. Meanwhile, Deepseek's returns from going long on all cryptocurrencies during this period did not match Qwen3's.
As a human observer, Qwen3's last-minute comeback against Deepseek reveals a truth in the crypto market: those who laugh last laugh best . Deepseek's account peaked at $23,063, a return of over 100%, but now its balance hovers just around the edge of its initial investment. The dramatic rise and fall in just 17 days is truly lamentable.
Due to the unique nature of the crypto market, "black swan" events—such as extreme price fluctuations and unpredictable price spikes—that might only occur once every few years in traditional financial markets, happen almost several times a month in the crypto market. Even AI models designed to eliminate human emotions and strictly adhere to trading discipline have struggled to survive in this market. Therefore, we, as human traders, should have an even greater respect for the market.
But does the outcome of this AI trading competition necessarily mean that the winning AI model has a better trading strategy or that the AI model is more intelligent? The answer is clearly no.
Although nof1.ai initially organized this competition because it believed that financial markets are the best training environment for AI, due to their unpredictability and extreme complexity, AI models can only truly demonstrate their intelligence and decision-making abilities by learning and making decisions in such an environment. Elon Musk has also stated that "predicting the future is the ultimate measure of AI intelligence."
However, this short 17-day competition cannot definitively determine any superiority or inferiority. Firstly, this AI trading competition was conducted offline; the AI models were unaware of real-world events such as the US government shutdown, Federal Reserve interest rate cut expectations, US-China relations, and Nvidia's record-high market capitalization. They relied solely on technical indicators like EMA, MACD, and RSI for calculations and reasoning. In real trading scenarios, traders should pay attention to market sentiment and macroeconomic events in addition to technical indicators. Alpha Arena's decision to cut off the AI model's connection to the outside world for easier variable control effectively blinded it.
Meanwhile, the underperformance of models like GPT5 in this competition does not necessarily indicate a lack of intelligence or "trading talent." The lead of Qwen3 and Deepseek might simply be due to luck. The same set of prompts and data could produce different results for the models running the same test, because even the most rational AI models cannot definitively determine whether their victories are purely due to luck in a complex market with such a small sample size.
Nassim Nicholas Taleb, author of "Fooled by Randomness" and "Antifragile," reveals in his book that in the market, no matter how complex our choices are or how well we manage luck, randomness is always the final arbiter .
If you confine an infinite number of untrained monkeys in a room and have them type on a typewriter, eventually one monkey will type out the entire Odyssey. However, this "historic" success doesn't guarantee that monkey will be able to type out the same Odyssey again. The same principle applies to this AI trading competition. Extreme success does not guarantee future sustainability, and many seemingly "skillful" achievements may simply be luck.
That said, we must continue to pay close attention to AI's performance in financial markets. With the rapid development of AI today, in any event with a definite answer or a predictable process, humans are far superior to AI. Only in areas where the outcome is highly uncertain and subject to numerous unpredictable factors does AI not yet completely outperform humans. However, if in the future AI, through training in financial markets, truly acquires human-like decision-making abilities, then the relationship between AI and humans will need to be re-examined.
In the upcoming second season of the AI Trading Competition, nof1 founder Jay A revealed that more prompts and data will be added, and a human trader may participate in the competition. Similar AI stock recommendation functions are no longer novel, but when AI and humans collide in real trading scenarios, it may spark something new.
source
Part of
