TypeSafe Jev returns typed decisions, not text: 70ms latency, free output tokens
mostly_idempotent · reddit · 2026-09-18
An independent AI engineer shares early access impressions of TypeSafe Jev, a 'decision model' that breaks from the LLM paradigm:
- How it works: A hybrid of classifier and LLM. You feed app state plus typed questions; it returns choices, scores, and probabilities your code acts on—no prose.
- Speed: 70–500ms end-to-end inference; questions are evaluated in parallel, so adding more barely changes latency.
- Pricing: $0.042 per million input tokens, output free (metering output reportedly costs more than producing it).
- Tradeoffs: You must decompose fuzzy judgment into discrete questions and keep weighting in your own code; no weights or paper yet, so benchmark claims are self-reported.
- Philosophy: 'Build prod, not God'—models are smart enough; what blocks automation is building systems on top. Typed, calibrated judgments can replace human-in-the-loop deep in production stacks.
Use cases include classification, routing, approval, and escalation scoring.
Related event: Hands-on with TypeSafe's Jev: A Fast Decision Model, Not an LLM Replacement(5 posts)→
More from coding & agent
- 15 harnesses, same result: informal test concludes agent harnesses don't matter — airesearch12 · 2026-09-18
- 4 deployment strategies explained via 4 visuals: feature toggle, blue-green, canary — _jaydeepkarale · 2026-09-18
- ServerKit: open-source server panel for apps, DBs and Docker hits 1.3k stars — tom_doerr · 2026-09-18
- Gary Marcus: Agents hold production credentials in a security gap nobody owns — GaryMarcus · 2026-09-18
- Amp's design philosophy: simple primitives, no tricks, let models improve it — HankYeomans · 2026-09-18
- You.com wraps AI Agentic Hackathon with self-improving agents challenge — PolarBearby · 2026-09-18