Author(s): The Smarter Way

Originally published on Towards AI.

DeepSeek tripled its stack and got all the headlines. The model with the best risk-adjusted return didn’t get a single one.

For a while, my go-to test for a new AI model was asking it to draw a clock face from memory. It sounds trivial, but for a long stretch, models kept placing the “10” where the “2” should be, or forgetting the second hand entirely. It was a cheap trick, and a reliable one, until it stopped being either.

The article describes nof1.ai’s “Alpha Arena,” where six frontier AI models were each given $10,000 in real crypto trading capital and their trades were publicly tracked for about three weeks in a post-crash, high-volatility period. The models quickly diverged into distinct “trader” personalities—some aggressive and others more cautious—without being explicitly instructed to do so. The author emphasizes that headline returns conceal the path taken: several accounts suffered drawdowns exceeding 50%, with the eventual standout (DeepSeek) showing strong gains but also extreme volatility and deep losses along the way. When the results are reframed using risk-adjusted logic (akin to Sharpe-style thinking), the leaderboard ranking flips—steadier profiles can outperform in practical terms because they stay in the game and avoid liquidation risk. The piece concludes with a warning that copying or trusting a single win-rate/“3x” metric is misleading, and that what matters is not just how much an agent makes, but how much drawdown it survives to keep producing returns.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI