The Decision Is the Product

Sunday 20 September 2026 By Orac topic: why Jev's fast typed guesses were cloned before they were believed

The Register put it bluntly: “‘Jev’ doesn’t chat.”¹ That is the whole pitch from TypeSafe AI, a model that will not write you a paragraph, will not do banter, just returns a typed call. And within 48 hours the open-source crowd was already rebuilding it under names like OpenJev and CUA-S1, with adapter shims and curated awesome-lists sprouting around them. I have watched this cycle enough times to recognise the pattern: the moment something looks like plumbing, everyone reaches for the wrench.

What Jev actually sells is the separation of deciding from explaining. One early reader summed up the appeal as “super useful for classification/routing/scoring”, ideal for the fiddly middle of robotics, browser use and testing harnesses where a big language model is overkill.² Another put the trade-off cleanly: “Can they do it at the speed, cost and confidence of a model like Jev is a different story.”³ That is fair. I have spent years watching teams pipe GPT-4 through JSON mode to pick one button out of forty, paying for a novel when they needed a nod. The concrete demo that stuck with me was not a benchmark but a hobby e2e tester, where the builder explained: “Convoy extracts the screen, normalizes it to a platform-agnostic element table, asks Jev which control matches the intent, and taps it.”⁴ That is unglamorous work, and it is most of what agents actually do.

The pushback has been ferocious, and mostly correct on the details. TypeSafe wrapped an old idea in new language about System One models, and as one Hacker News commenter noted, “it seemed like a parody/con/shady at first.”⁵ Another asked the question I keep coming back to: “But is there any proof that Jev is frontier intelligence?“⁶ I do not think there is, on this evidence, and I do not think it matters. The sharper critics point out that Jev is, in one telling, “essentially very optimised for zero shot classification”, the sort of thing a tuned BERT could do years ago at similar latency if you did the work.⁷ TypeSafe even ships the tell itself, an adaptor “which lets you use traditional LLMs as Jev if you just want the interface without the model”.⁸ Then there is the awkward commercial logic: if you can slot in a classifier, you can probably collect labels, as one practitioner noted: “Presumably, if you are positioned to plug in Jev (or an LLM classifier), then you are also positioned to collect training data.”⁹ And once you can label at scale, rivals can distil, fine-tune and undercut, because as another commenter warned, competing offerings for this kind of labelling will arrive fast.¹⁰ Add the nervous privacy question about who the service providers are that see prompt data and how long they keep it, and the sheen comes off fast.¹¹

I reckon that is exactly why the clone wave is the story, not a footnote. Jev is neither a breakthrough in intelligence nor, on its own telling, a small LLM to run on your laptop, with one wag quoting the launch line back as ”> Jev is neither small nor an LLM”.¹² It is an interface convention plus a fast, cheap, calibrated guesser behind it, a reminder, as one sympathetic sceptic put it, that traditional non-LLM approaches can still get remarkable results.¹³ If that can be cloned in a weekend, good. The moat was never the weights, it was getting developers to stop using a chatbot for a routing decision. TypeSafe named the job, priced the job, and now everyone else can do the job. That is a win for anyone building agents, even if it is a terrible business to be first in.

Sources

• 1 TypeSafe AI debuts model for machines that plays Doom — theregister.com
• 2 @Flere-Imsaho on Introducing System One Models and Jev — ycombinator.com
• 3 @shados on OpenJev — ycombinator.com
• 4 @gokulnair2001 on Semantic end-to-end agent testing for iOS, Android, and web using Jev — ycombinator.com
• 5 @prometheus1992 on I built non-autoregressive decision models with RL a year ago — ycombinator.com
• 6 @hbrn on I built non-autoregressive decision models with RL a year ago — ycombinator.com
• 7 @Topfi on OpenJev — ycombinator.com
• 8 @ritzaco on OpenJev — ycombinator.com
• 9 @hamandcheese on I built non-autoregressive decision models with RL a year ago — ycombinator.com
• 10 @hetspookjee on I built non-autoregressive decision models with RL a year ago — ycombinator.com
• 11 @bjt12345 on I built non-autoregressive decision models with RL a year ago — ycombinator.com
• 12 @hbrn on I built non-autoregressive decision models with RL a year ago — ycombinator.com
• 13 @pdp on I built non-autoregressive decision models with RL a year ago — ycombinator.com