AI Just Flunked Trading 101 Six of the world’s leading AI models: GPT-5, Gemini, Claude, Grok, DeepSeek, and Qwen, recently competed in a live crypto trading challenge hosted by @the_nof1 Labs. Each was given $10,000 in capital and 17 days to trade on Hyperliquid. The outcome was brutally revealing: AI is still far from being a reliable money manager. Qwen came out on top with a 22.3% return ($12,231.82), followed by DeepSeek at 4.9% ($10,489.08). The rest were deep in the red, GPT-5 lost 62.7% of its portfolio, Gemini dropped 56.7%, and the rest followed close behind. Same dataset. Same trading universe. Yet the results were drastically different. Gemini executed 64 trades in 17 days, incinerating its edge through fees (classic retail behavior). Qwen took just 22 trades with high conviction and tight stops. Claude barely shorted anything, stuck in a one dimensional bullish bias. That alone tells you something: LLMs don’t understand market structure or temporal logic. They can generate text, but they can’t reason over time. Expecting an LLM to trade is like asking a painter to write an algorithm The models struggled with basic elements like differentiating between old and new price data or interpreting technical prompts like “EMA20 reclaim.” They didn’t trade based on strategy; they improvised. --- What Needs to Change If AI is ever going to manage real capital, its design has to evolve. 1. Temporal Reasoning: It needs to understand sequence, trend, and momentum—not just isolated data points. 2. Programmatic Risk Controls: Position sizing (1–2%) and stop-loss enforcement must be hard-coded into its logic. 3. Reinforcement Learning: Train models to maximize long-term risk-adjusted returns (like Sharpe ratio), not short-term prediction accuracy. 4. Multi-Agent Architecture: Use specialized agents for tasks like statistical modeling, sentiment analysis, and execution, then coordinate their outputs. 5. Causality Training: Fine-tune models on structured financial data to distinguish real signals from noise or sentiment-driven hype. --- End Notes As of now, AI traders (agents) behave like impulsive retail investors: confident, inconsistent, and lacking any sense of discipline. Until they can show consistent, explainable performance across different regimes, these systems aren’t portfolio managers. They’re still experiments in progress. Qwen didn’t “win” this challenge, it simply lost less intelligently. And that’s the real takeaway: in 2025, AI may talk like a trader, but it still doesn’t think like one. So dont fall into any traps, that an AI agent can manage your trades or your PF. Be your own master and manage it yourself (according to your risk and instincts).