Diogo Almeida, a former researcher at OpenAI, has unveiled a groundbreaking new model known as Jev. This innovative model eschews the conventional approach of free text generation, opting instead to directly produce structured judgments with predefined types, accompanied by their respective probabilities and confidence levels. Leveraging parallel computing techniques, Jev achieves a remarkable speed boost of 20 to 200 times compared to traditional autoregressive methods employed by large-scale models. This not only slashes costs by up to 1/400th of the original but also offers the added benefit of free output tokens.

Jev's training methodology, dubbed RLCD, centers on calibrated decision-making. Unlike RLHF, RLCD places a premium on ensuring that the probabilities assigned by the model genuinely mirror its level of certainty. TypeSafe, the entity behind Jev, asserts that the model boasts a 0% hallucination rate. In practical terms, this signifies that the model's output will remain strictly within the predefined boundaries. Despite facing some skepticism regarding its perceived lack of technical innovation and the redefinition of hallucination, the underlying premise of Jev—utilizing just the right amount of capability to automate a vast array of repetitive, latency- and cost-sensitive tasks—still holds significant promise.