Jev replaces LLM decision calls with typed questions
A video covers TypeSafe's Jev: no text generation, calibrated probabilities in one call, reported 193.6x faster and 444.6x cheaper than LLM calls on four decision workflows.
🚨holy f*ck, this broke my brain. Jev replaces every LLM decision call for $0.042 per million. Output is free. Diogo Almeida built it at TypeSafe. No text generation. The model takes typed questions and returns calibrated probabilities in one call. Jev against LLM calls on four decision workflows: 193.6x faster, 444.6x cheaper, upper bound. The worked example on the home page clocks 75x and 171x. SPLIT - no KV cache, no long transcript, typed questions in parallel SCORE - choice, yes/no, probability answered at once in one call GATE - guardrail scores a shell command before the harness runs it One long transcript is the tax. The split pipeline runs below ↓
This post quotes another post: x.com/6aw0ron0k/status/2114358843685609544