Post on X
- Views
- 163,291
- Likes
- 1,941
- Reposts
- 98
Read from X on . Counts change daily.
kimmonismus’s post is a skeptical-but-curious summary of what Jev is, rather than a project built on top of it, and it circulated widely. The framing: Jev, introduced by Diogo Almeida, is a model built to make decisions instead of generating text, trained with a method the post calls RLCD.
The claims kimmonismus lists are TypeSafe’s own: 20 to 200 times faster and 40 to 400 times cheaper than comparable large language models, at $0.042 per million input tokens with free output tokens. The trade-off stated is that Jev cannot generate text at all, only pick from options or score them. The post points to two demos as evidence of the range of tasks this covers: controlling a game and choosing between Wikipedia links. Its closing line frames the pitch plainly: fast, cheap decisions embedded directly into software rather than a chat interface.
