The top comment, from jacobgold, opens by congratulating the team and then questions two parts of the launch. It calls the speed comparison misleading, since a generative model that writes code in a Turing-complete language can do anything a computer can, while Jev only returns structured output. It suggests a more accurate title: “Jev: Trading general purpose generation for fast typed inference”. Another commenter adds that the thread’s original title, “Jev: New frontier model 40-400x cheaper and 20-200x faster”, was misleading.
The “can’t hallucinate” claim draws the most pushback. jacobgold points out that Jev cannot emit an invalid type but can still return a valid value that is completely wrong, and that an ordinary language model can be forced into structured output with the right harness. A reply puts it as “Type safety is not factual correctness.” Another commenter notes Jev can still forward a billing query to the dev department by mistake. One more says they have done similar constrained generation with local models, reading a confidence score off the logits, the raw scores a model assigns to each possible next token.
Several commenters link to TypeSafe’s demo of Jev playing Doom, which is also in the launch article.
