对 Jev 的批评：真实管线里用不上
作者认为 Jev 只给类型化答案，不校验输入真伪，脚本和廉价模型已能做同样的事，还多出延迟与成本。
My offical take and review of JEV. BY @typesafeai Jev and RLCD are useless in real world applications. Quick read: 1 it is a Wasteful entity and product. A money grab from all angels. This training method is stupid and was already prove to fail in large scale training environments many many moons ago. (Its in a deepseek paper) It does not help the actual problems in AI. Not compute - dosnt solve. Not token cost. - dosnt solve. Not intelligence. - dosnt solve. Not acceptance. - dosnt solve. Jev only gives you a typed answer. That’s it. It does not check if the input is real. It does not fix the record. It does not verify the world. If the state you hand it is already wrong, Jev still answers. Clean looking. Schema legal. Still wrong. So the input has to be accurate first. Already checked. Already filtered. Already trusted by something that is not Jev. In real pipelines that job is already done. Scripts run 24/7. Free. Gateways. Runtimes. Validators. Schema checks. Dedupe. Dead letter queues. They already drop the bad file. They already reject the broken event. They already enforce the contract. Putting Jev there is useless. You pay a model to redo work code already finished. You add latency. An API. A bill. A new way to fail. Same thing at the model layer. Haiku. DeepSeek. The basically-free models. They already classify. Route. Score. Yes or no. Structured enough to branch on. And add our above already built scripts and your own memory and boom no hallucinations. That’s the job RLCD was sold as. You don’t need a special “system one” model for it. You already have it. Worse: Jev still needs the content made for it. Someone still has to extract the record, clean the text, assemble the evidence, produce the “verified” input. Scripts do that. Cheap models do that. Jev sits after it and picks from a closed list. That’s not a breakthrough. That’s an extra hop on finished work. Jev does not create verified inputs. It does not solve compute. It does not solve cost. It does not get you closer to real intelligence. RLCD is a money grab. Jev is a step backwards. ACTUALLY! Your runtime already verifies. ITS NATIVE! Your free model already judges. Jev is just the receipt they want you to pay for.