TypeSafe's Jev ditches text generation for instant, calibrated numerical answers — and it plays Doom at 7 req/s
JnBrymn · x · 2026-09-17
TypeSafe announced Jev, a new take on LLMs: instead of generating text token by token, it takes a state (text or JSON) plus a list of questions and returns calibrated, consistent numerical answers almost instantaneously. Blogger James Brooks digs into how it works.
Three question types:
- choice: picks from a fixed list, returns {choice, probabilities, confidence}
- score: positions on an ordered rubric, returns {score, legend, probabilities, confidence}
- noul: true/false probability from 0 to 1
Why it matters: conventional LLMs are orders of magnitude slower, and their self-reported confidence often bears little relation to reality. Jev wins on speed and calibrated numeric prediction — in the official Doom demo it processes 7 requests per second in real time, navigating the map and shooting enemies, something a standard LLM can't do.
Related event: ChatGPT co-inventor launches Jev, a decision-only model 200x faster(80 posts)→
More from coding & agent
- Claude Code Tells Dev a Feature Is Too Simple to Code, Go Do It Yourself — bendee983 · 2026-09-17
- Microsoft's MarkItDown Passes 100K Stars: Turn Almost Any File Into Clean Markdown — mdancho84 · 2026-09-17
- What breaks when an LLM agent moves from demo to production? — Substantial_Bus_5237 · 2026-09-17
- 'When Context and Code Live Together': Devs Revive Knuth's Literate Programming for AI — carsonfarmer · 2026-09-17
- Switching tweet classification to Jev: 6x faster, ~40x cheaper than fastest LLMs — altryne · 2026-09-17
- Dev says Jev classifies tweets 6X faster and ~40X cheaper than fastest LLM — altryne · 2026-09-17