RLHF 共同发明人谈 Jev：不生成文本的决策层
播客预告：Diogo Almeida 隐身两年后带 4000 万美元和 Jev 现身，讨论专用决策层如何改变 agent 的延迟与成本约束，以及 System One 是否分裂市场。
🚨 Episode 97 of The Intelligence Journal is here. Jev: The Model That Doesn't Talk 🤐⚡ Diogo Almeida co-invented RLHF, the technique that made ChatGPT work. He spent two years in stealth. This week he emerged with $40 million and a model that does the opposite of everything he helped build. Jev generates no text. None. You give it unstructured input, you hand it the options in advance, and it returns a typed decision with a calibrated confidence score. 🎯 It threw out autoregressive generation entirely and evaluates every option in parallel instead. Join us as we break down: - Why the man who taught models to talk decided the next bottleneck is models that talk too much. - What a dedicated decision layer changes for agents, where latency and cost per call are the entire constraint on what you can actually let an agent do. - Whether "System One" splits the market into models that think out loud and models that just act. 🧠 🗓️ Date: Sep 21, 2026 ⏰ Time: 3:00 PM UTC 📍 Set your reminder: https://x.com/i/spaces/1yGBePrZqMkKN?s=20