On this page
The most interesting news this week, to me, was the JEV model: extremely fast, and it only does probability judgments and formatted output. After seeing all kinds of applications on X, I got early access, and actually running it has been very impressive. A lot of the semantic-judgment gates I used to jam luna or flash-lite into: benchmarks come out close, but JEV is faster and cheaper, so I’m swapping them out.
Want AI working in your own job, not just in demos? See the 3-month 1-on-1 program →
Four Places I Use It
A few applications I think are genuinely great:
- Reorganizing and slimming down my harness. It’s up to 15,000 segments across the whole machine at this point. My approach: (1) Regex-scan for keywords that shouldn’t be there (dates, negative constraints, incident logs, etc.) (2) Use semantic vectors to catch conflicts and duplication (3) Have JEV judge each entry one by one against the official harness recommendations (4) Only hand it to an LLM for the final call when confidence is low (5) Whatever failures get caught eventually turn into hooks that block automatically
-
My daily work-progress and dashboard automation pipeline: JEV does fast classification and filtering, plus the contextual call on “is this to-do actually a good fit to hand to AI, with a human reviewing at the end.”
-
Every night I crawl the web for a big batch of AI-industry news, and JEV scores all of it fast so I can tell which items are just noise and filter them out.
- My education product platform needs to generate a large volume of exam questions, and I use JEV to gate the creation quality and answer consistency of those questions.
Three Criteria to Judge By
As for the flashier stuff on X, using it to check flight tickets, play Mario, whatever, it looks flashy but you don’t have to chase it. I’d suggest judging with these three criteria instead:
- Is there something you can’t just filter with plain code, something that actually needs semantic judgment?
- Is there a judgment that only needs a fixed-format answer, not a big wall of output?
- Is there something that doesn’t need multi-layer, high-level reasoning — just a quick confidence-based routing decision?
Aside: one AGI, everyone reads it their own way
Beyond serious engineering deployments, the discussions in the community have been pretty entertaining too. Someone said JEV is finally our glimpse of AGI, and I found that pretty funny. These days everything gets “one thing, everyone reads it their own way,” and AGI is unsurprisingly no exception. Too bad JEV can’t take images. There’s also this one going around online, probably the most creative imagined use case for JEV I’ve seen:
