Jev 用 RLCD 替代 RLHF 做自动化决策
作者研究 Jev 后认为其价值不在快和便宜，而在把 AI 目标从讨好人类转向结构化决策与校准概率，适合交易、支付等高风险自动化场景。
Jev has been widely discussed over the past few days, so I spent some time researching it as well. What interests me most is not simply that it is faster or cheaper, but that it changes what AI is optimized to do. Its founder argues that RLCD will outperform RLHF for automation. I think the direction makes sense. RLHF made AI extremely good at understanding, responding to and even pleasing humans. But human preference is not the same as truth. When AI is an assistant, a human can still catch its mistakes. When AI begins acting autonomously in trading, payments or other high-stakes environments, the instinct to please—and to sound confident—becomes a liability. RLCD shifts the objective toward structured decisions and calibrated probabilities. AI should not only provide an answer; it should communicate how certain it is, when it can act independently and when the decision should return to a human. Making AI better at pleasing us may be the wrong kind of acceleration. Once sufficient context and data exist, enabling AI to make measurable judgments under uncertainty may be the more valuable path. The next stage of AI may not be about sounding more human. It may be about making decisions that can be measured, verified and safely used. https://www.youtube.com/watch?v=cJ0EOzey--o