Jev isn't new tech. Its marketing targets people who think AI started with LLMs.

I keep seeing Jev presented as some new class of decision model, but most of what’s being advertised is just normal classifier behavior with modern zero-shot capabilities. It outputs probabilities over constrained choices, doesn’t generate autoregressively, can’t output an invalid class, and can use labels defined at inference time. None of that is new. Zero-shot/NLI classifiers, embedding models, cross-encoders and rerankers have been doing variations of this for years. The weird part is that most of the impressive Jev comparisons are against LLMs. Of course a specialized classifier is faster and cheaper than making an autoregressive LLM generate an answer. That doesn’t establish a new paradigm. The meaningful comparison is against strong existing classifiers. There are already benchmarks like BTZSC evaluating dozens of zero-shot classifiers across 22 datasets, including NLI models, embedding models and rerankers. I haven’t seen Jev properly benchmarked across that landscape yet. (https://proceedings.iclr.cc/paper_files/paper/2026/hash/417e1c15b3d49852fceded8aa104107d-Abstract-Conference.html) Where people have compared Jev with conventional classifiers, the story is much less magical. One Banking77 experiment got 93.3% from BGE-small + logistic regression versus 83.2% for Jev, at about 9ms locally. (https://github.com/ickma2311/jev-baselines-eval) So color me a skeptic. Look, Jev might even be a good product. Maybe their unpublished architecture or RLCD training method is genuinely novel. But nothing we've seen so far establishes that "System One Models" are a new class of AI. What the public evidence mostly establishes is that using a specialized classifier for classification can be much cheaper and faster than using an autoregressive LLM, which we already knew. It only sounds novel if your idea of AI begins and ends with LLMs.