Jev 宣称让 Astra 推理快 193 倍、便宜 444 倍

darkzodchi @zodchiii原推发布于 2026.09.20

在 X 上打开原推

Jev is the "ChatGPT" moment for AI infrastructure: up to 193x faster and 444x cheaper in tests with Claude Fable 5.1 and GPT-6 Astra What Jev is, how the three primitives work, and how to start using it, in my 2-page working note: step 1 → meet the constraint: it generates no prose, no explanation, no reasoning trace. a bounded question returns a value directly, so a decision stops being a prompt to engineer and becomes a contract to declare step 2 → learn the three shapes: Choice picks one winner from 2 to 255 options and reports the full distribution, Score places state on a rubric of 2 to 10 levels, Noul returns one probability step 3 → treat state as evidence: goal, verified facts, rules in force, the options actually available right now, and what has already been done. keep it as small as can still change the answer step 4 → stop hiding instructions in the state: the questions carry the instruction. the state carries what happened step 5 → batch what is independent: every question in a request runs against the same state in one round trip, so ask everything answerable from the current evidence. if B needs A, that is a second step, not a second question step 6 → read confidence correctly: it reports how concentrated the distribution is. it is not correctness and it is not permission. a low figure usually means your criteria were ambiguous step 7 → use bands, not a cutoff: escalate below, auto-route in the middle, and require a deterministic check or a person for anything that cannot be taken back step 8 → keep the boundary sharp: Jev selects among allowed options, your code owns schema, threshold, permission and budget, the executor takes one observable action, the verifier reads the external system again step 9 → keep it out of the wrong places: anything computable, open option spaces, multi-label questions and free-form writing all belong elsewhere then run the first week: one repeated decision, ten examples including two ambiguous and one with no valid answer, expected outcomes written down before anything runs the result: a hallucinated tool name cannot execute because it was never in the candidate set, and a completion claim cannot close a loop because fresh evidence is required Copy this complete Jev blueprint - then read full Jev setup below ↓

这条是引用着另一条写的：x.com/zodchiii/status/2101243146596384854