Post

Technical correction/clarification: the choice response is not a probability distribution "plus" a confidence level. The "confidence" is a summary measure derived from the probability distribution: docs.typesafe.ai/confidence

“It doesn’t imitate writing!” [crowd cheers] “It makes decisions!” [crowd boos]

“ I really hope nobody uses Jev to rank job applicants” That, and triaging email inboxes, are the two things I’ve seen everyone leap on so far.

Do you see decision models fitting best as a fast routing layer before handing work to a heavier agent?

it's not really a new idea, but the way they focused on the dev experience right out of the gate definitely helped it explode

"If Jev marks something as spam, which content signals tipped it off? This also means that concerns about bias should be front and center."

Regarding the opaqueness, one way to mitigate would be, after it says "yes" or "no", ask it "why" and have it choose from a list of categories. Also, multiple choice questions might have a "none of the above" or "it's complicated" answer where it can report bugs and other problems.

Are there competing services that you would recommend using instead?

It's a large model, so it has a good amount of general knowledge and pre-trained understanding, meaning it doesn't need fine-tuning. And it's easily accessible over API. I don't think that has existed since forever.

Interesting writeup, though IMHO you missed the BERT connection - tech is different (maybe? we don't know!), the instructions-following part is a big improvement, the use cases are similar, as is performance. Interesting tech, great marketing; my own take: blog.senko.net/analyzing-je...

Decision models make the latency/quality tradeoff easier to reason about. A useful production split is routing classification and triage to the fast path, then escalating ambiguous cases to a slower reasoning model. The eval should measure escalation rate, not just accuracy.

Bookmarked, need to read this properly. The "system one" framing for decision models is a good hook, been trying to name that category myself while building agent stuff.

Good explainer. Thanks.

i suppose you would have to hook it up to an llm and drawing program and make it decide the length and direction of each successive stroke to generate the pelican on a bicycle drawing thanks for the writeup

Calling them "decision models" makes so much more sense conceptually than forcing the System One framing. Focusing on fast, low-latency execution for structured choices rather than full generative reasoning is going to be massive for agentic workflows.

Several of these made public now, some much better or worse than others. I did a write up on the latest one bsky.app/profile/jona...

Thread includes a repo that can be used to demo Jev against any model. The speed up numbers are real, even in the area least likely to benefit as much (web search) bsky.app/profile/jshe...