Made a Jev-based daily guessing game that's like a crossover between Wavelength and Family Feud (or Family Fortunes if you're from the UK or Ireland).

It's called Vaguely Right (https://vaguelyright.com).

Each round the you get a spectrum and a target on that spectrum, (e.g. Calm ----X-- Stressful). You then have to name something that Jev will place on that target on the spectrum (a job interview, a walk in the park, a pelican riding a unicycle?). The closer you are, the higher your score.

I actually posted about it on Show HN earlier today: https://news.ycombinator.com/item?id=50071000

My one experiment with these models so far hasn't worked. I was hoping I could use them to find my "best" photos - most aesthetically pleasing, best composed, etc.

I tried using GPT-6 Astra Decisions since they accepts multiple image attachments, but it's not clear to me that the order of attachments is available to the model. I couldn't find a prompt for "which of these photos is best" that wasn't obviously affected by the order in which I attached the images.

Have you tried giving these models questions about coin flips and dice rolls? Or happen to catch any of the Jev plays Pokemon?

It completely undermines any trust I have in their outputs. Comparing to a little llm is amusing, coins are still 50/50 with 100% confidence for them. I didn't follow up with any questions about landing on the side.

It's used extensively in my own agentic products. basically as a filter for various purposes, one example, to reduce the main agent's cognitive load.

there was a dynamic context loader (mixer) that combined results from embedding top k, regex filters etc and then there was a sub agent as a ranker to filter out the useless ones not matching the queries (wrong results from regex, and useless ones from top k etc). it was the job of a 5.6 luna(which worked very well ngl) but then replaced by jev.

Saved 5-10s in each round, no precision loss, and saved the bill a bit more. net win for me

experimenting with a bunch of things:

- reimplemented claude code auto mode for pi

- prompt injection detections refactored from llm judges to classifier judge

- as above various first pass detections for generated and untrusted files, such as moderation and malware detection

- ranking search results in web_search tools to reduce context overload

I found that a lot of what i was using a one shot haiku llm call for can now be done with jev or another decision model.

I am about to try using it as the last step in an email spam filter.

I work on a video editor. It exposes 100+ possible tools to LLMs. We use Jev to decide which subset of tools to send to the LLM. We also use Jev to classify a user's request, are they saying "Hello" or do they want some work done. I looked into OAI decision api, it's strongest advantage right now is multimodal but we don't need that yet. The makers of Jev just announced something similar.

I'm using the openai decision model mostly because I already have an API key. I use it to categorize which existing topic/location/company/etc is a best match for some new content that comes in.