Jev by TypeSafe.AI: a parallel "System 1" model for fast structured decisions

Melvin Vivas · X video post · 2026-09-16 · 2:56 · 1,147 views · Open on X

Topics: Industry Trends & Job Market, LLM Fundamentals · Level: intermediate

Summary

Melvin Vivas reshares the launch video for Jev, which TypeSafe.AI calls the first public "System 1" model. In the video, founder Diego Almeida argues that RLHF-trained chat LLMs are tuned to human preferences. He says this makes them overconfident, unreliable and dependent on humans in the loop. TypeSafe's answer is a model trained with a new method, RLCD (reinforcement learning for calibrated decisions), that answers structured questions all at once and returns decisions with confidence scores instead of free text. This is a launch ad, so all speed, cost and "can't hallucinate" claims come from the company and have not been checked independently.

Key points

• The main criticism: RLHF tunes LLMs to human preferences, which causes mode dropping, overconfidence and poor reliability, so real automation still needs humans in the loop.
• LLMs write one token at a time (autoregression, "the tiny straw"). That suits conversation but is slow for software-to-software decisions.
• TypeSafe's claimed fix has three parts: a new architecture, a new sampler and a new training algorithm, RLCD (Reinforcement Learning for Calibrated Decisions).
• They compare it to transformers replacing RNNs: sequential computation gives way to parallel computation, and many structured questions are answered almost instantly.
• Outputs are decisions with probabilities and confidence, not words. The company says this makes them reliable, self-consistent and type-safe, more like code (its claim, unverified).
• Claimed numbers: about 100x faster and 100x cheaper, input at $42 per billion tokens, output tokens free. The quoted post says 20-200x faster, so the figures don't match.
• Target uses are high-volume background work inside software loops: moderating messages, routing tickets, reviewing documents, with fewer retries, parsers and broken schemas.
• Tagline: "building prod, not god". The product is aimed at automation, not AGI.

Resources mentioned

• Jev · tool · typesafe.ai · paid
  A decision model that makes turn-level forecasts on AI agent conversations using only structural signals (turns, tool calls, workflow stages, timing).
  Also in: Generating a Repo Promo Video with a Claude Skill on Sonnet 5.5 vs Opus 5.5 (Melvin Vivas on X · notes), GLiDE by Fastino Labs: A Post-Trainable Reasoning Decision Model (Melvin Vivas on X · notes), Using Jev as a Reranker to Augment RAG Retrieval (Melvin Vivas on X · notes), Using Jev (TypeSafe AI) as a Decision Model for Email Classification (Melvin Vivas on X · notes) and 12 more
• Diego Almeida · person · x.com · free
  Founder of TypeSafe.AI, who says he co-created ChatGPT and RLHF at OpenAI.
• Jev (TypeSafe AI) · tool · typesafe.ai · free
  TypeSafe's new post-training algorithm for calibrated, confidence-scored decisions, pitched as a replacement for RLHF.
  Also in: JevDev: Open-Source UI Tool for Experimenting with Jev (Typesafe.ai) (Melvin Vivas on X · notes), jev-dev: A UI Tool to Keep Run History When Experimenting with Jev (Melvin Vivas on X · notes), JevDev: Open-Source UI for Experimenting with Jev by Typesafe.ai (Melvin Vivas on X · notes), Building a Jev Session-History Tool with Codex and Astra (Melvin Vivas on X · notes) and 7 more