A Jev-like wrapper for LLMs, including vision

That is the post's premise: a single function that turns an ordinary LLM into a Jev-like decision call, vision models included. The author says he tested it successfully with Qwen3-VL-8B-Instruct, adding only that caching gets trickier with video but is workable.

One commenter guesses Jev is around 30B parameters and argues the price is unlikely to be a subsidy. That is inference, not a TypeSafe statement, and the price quotes in the thread are commenters reading OpenRouter listings rather than anything the vendor published here.

The thread splits. One side argues intent needs motion understanding that still frames cannot give, pointing at V-JEPA 2 running at several FPS on a laptop. The other warns that do-what-I-mean behavior in deployed systems turns into spaghetti that fires unwanted actions.

Commenters mostly price-compare: Jev's input rate is quoted level with Gemma 4 26B A4B at about $0.042 per million tokens, with GPT-5 Nano at $0.05. The comment that gets to it credits Jev with probabilities that match real likelihoods, not with picking better answers.

It drew 157 points and 45 comments, which makes it one of the better-attended threads in this set and gives the technical back-and-forth more room than the two-point Show HNs. Point totals are captured at fetch time and move as the thread ages, so treat the number as a snapshot.

Ambient intent detection rather than classification: the example in the thread is automatic doors that open when someone intends to pass and ignore someone walking past or loitering. One comment argues you would need video and motion understanding rather than still frames, and another says the real value is that a sensor cannot recognize intent on its own.

Yes, and it is substantive rather than a nitpick. One side argues real-world automatic doors sit in high-traffic areas where a simple proximity trigger is right 99% of the time, so intent classification adds complexity and probabilistic behavior where none is needed. The other replies that the doors are there because the simple trigger is bad about 80% of the time.