本文以凱利準則（Kelly Criterion）為引子，探討了布拉格AI實驗室EquiLibre獲得5億美元估值的爭議。EquiLibre由三位前DeepMind研究員創立，利用強化學習技術從撲克遊戲轉向金融交易，與高頻交易公司Tower Research合作，聲稱在標普500和納斯達克每日交易數十億美元，且自成立以來每月均錄得正收益。文章以文藝復興科技公司（Renaissance Technologies）的傳奇基金Medallion為對比，指出該基金刻意將資產管理規模控制在約100億美元，甚至在2003年清退所有外部投資者，因為其創始人、數學家吉姆·西蒙斯深諳一個道理：下注規模應由你的優勢決定，而非市場規模。

文章的核心論點在於，量化基金與科技公司在本質上截然相反。風險投資追求的是非線性增值——更多用戶使產品更好、吸引更多用戶、形成護城河；然而交易策略恰恰相反，使用的人越多，策略就越弱，因為其本質是發現錯誤定價並從中獲利，當更多人複製同一策略時，套利機會便消失殆盡。此外，軟體具有規模經濟效應，而交易卻具有規模不經濟的特性——邊際用戶的成本為零，但邊際資金卻被「徵稅」，且稅率隨投入資金的增加而攀升，因為大額倉位會影響市場流動性，交易者的訂單本身就會推動價格偏離預期方向。

EquiLibre的CEO施密德坦承風投投資對沖基金確實不合常理，並將公司定位為「AI技術公司」而非金融公司，將交易視為第一個垂直應用場景，長期目標是開發能在真實世界中行動的AI。然而文章指出多項隱憂：其「零負月」紀錄誕生於2025年加密貨幣及股市普遍上漲的環境中，缺乏類似2007年「量化地震」般的極端壓力測試；回測表現與實盤盈虧之間存在系統性交易中最昂貴的鴻溝；且即便對其優勢做最慷慨的估計，按凱利準則計算，其價值也可能遠低於5億美元的估值。文章最終警告：AI並未將交易變成軟體業務，量化基金不是科技公司，投資者不應混淆兩者。

This article uses the Kelly Criterion as a framework to scrutinize the $500 million valuation of EquiLibre, a Prague-based AI lab founded by three former DeepMind researchers. The firm applies reinforcement learning—originally developed to master no-limit poker—to financial trading in partnership with Tower Research Capital, claiming billions in daily volume across the S&P 500 and Nasdaq with zero negative months since inception. The piece contrasts this with Jim Simons' Renaissance Technologies, whose legendary Medallion fund deliberately capped its assets at roughly $10 billion and expelled all outside investors in 2003, embodying the Kelly principle that bet size must match edge size, not market size.

The article's central argument is that quant funds and tech companies are structurally opposite businesses. Venture capitalists seek non-linear appreciation through network effects—more users improve the product, creating defensible moats. Trading strategies, however, degrade with scale and imitation: they profit from mispricings that close as more capital exploits them, and larger positions erode returns by moving the very prices the strategy targets. Software enjoys economies of scale where the marginal user is essentially free, while trading suffers diseconomies of scale where each marginal dollar is effectively taxed at an increasing rate. This fundamental asymmetry makes venture capital a poor structural fit for funding trading operations.

EquiLibre's CEO candidly acknowledges this tension, positioning the company as a technology lab that uses trading as its first revenue-generating vertical, with aspirations to build real-world AI agents. However, the article raises serious concerns: the firm's unblemished track record was established during a broadly rising market in 2025 crypto and equities, untested by stress events like the 2007 quant quake; back-tested improvements do not constitute evidence of edge replacement; and even generous estimates of the strategy's capacity, lifespan, and economics likely yield a value well below the $500 million valuation. The author concludes that AI has not transformed trading into a software business—quants have used machine learning for decades—and warns investors against confusing a business that compounds with one whose alpha inevitably depletes.