Jev 用 RLCD 承接 agent 里的小决策
介绍 Jev 以 RLCD（校准决策强化学习）处理「调不调工具、是否完成、走哪条分支」这类高频小判断，几乎零延迟、零输出 token 成本，大模型只负责重推理。
Have you heard of RLCD? it might be the most interesting idea in AI right now and almost nobody outside the space is talking about it yet the apps already using it run way faster while paying close to nothing for output here is the shape of it big models like Claude, GPT and Gemini are great at the hard reasoning but an agent makes hundreds of tiny decisions in between, should I call this tool, is this done, which branch next running those through a full model is slow and burns tokens on every single one a new model called Jev handles just those small calls it uses RLCD, reinforcement learning for calibrated decisions, so the little choices come back with almost no latency and zero output token cost the heavy model thinks, this handles the reflexes it was built by Diogo Almeida, a former OpenAI researcher who co authored the InstructGPT paper so this is not a random side project tomorrow I will break down how to wire a fast decision model like this into your own agent loop so the big model only gets called when it actually has to think
这条是引用着另一条写的：x.com/skrilla_git/status/2101641339175641467