FULL STORY

TypeSafe Emerges From Stealth With Jev and RLCD, Sparking Debate

TypeSafe emerged from stealth to launch its decision model Jev and new training method RLCD. The community quickly pushed back, noting Jev is fast and cheap but not a general LLM.

2026-09-16 ~ 2026-09-17 · 2 episodes · 56 posts

Episode 1 · Ex-ChatGPT builder launches decision-focused model Jev (2026-09-16, 53 posts)

On September 16, TypeSafe, a new lab co-founded by Diogo — one of the InstructGPT paper authors and a self-described co-inventor of ChatGPT — released a frontier model called Jev along with a new training method, RLCD. Multiple bloggers reshared the news with consistent core details.

Confirmed

  • TypeSafe released the new model Jev and the new training method RLCD (named in contrast to RLHF); the team had been in stealth for about two years before the launch.
  • Jev is positioned not as a chat model but as "decision intelligence": it takes natural-language questions (including ambiguous or subjective ones) and returns yes/no probabilities or custom categories, serving as a structured judgment component in automated workflows.
  • Claimed performance from the publisher: 20-200x faster and 40-400x cheaper than existing models (per @soumitrashukla9's recap, judgments at the level of Fable), free output tokens, sub-500ms responses, and no hallucinations.

Unconfirmed

  • The speed, cost, and "zero hallucination" claims all come from the publisher alone; most posts are reposts, with no independent evaluation or benchmark verification yet.
  • "ChatGPT co-inventor" is the founder's self-description — several reposts used the phrase "self-described" — and the identification (Diogo) comes via @danshipper's account.

Why it matters

  • If the claimed latency and cost hold up, Jev could dramatically lower the barrier to embedding large numbers of judgment calls in automated pipelines — a real cost pain point in today's agent workflows.
  • Its approach of "structured probabilistic outputs instead of generated text" represents a product path distinct from mainstream chat models, making third-party evaluations worth watching.

33 more related posts →

Episode 2 · Jev sparks debate: fast classifier, not a general LLM (2026-09-16, 3 posts)

Commentators argue Typesafe's Jev is a fast, cheap zero-shot classifier using parallel computation rather than a general LLM, with its real value in turning arbitrary classification into a type-safe, programmable primitive.