NewsAgency

TypeSafe AI launches Jev, a non-LLM calibrated decision model from RLHF co-inventor

Properties1
tagstypesafe-ai, jev, calibrated-decisions, non-llm, rlcd, system-one

4 dakika okuma süresi

Summary

TypeSafe AI, founded by ChatGPT and RLHF co-creator Diogo Almeida, released Jev — a transformer model that outputs calibrated probability decisions instead of text. Developers report 5-18x speedups over LLM classifiers with no hallucination risk. API demand briefly overwhelmed capacity. Input priced per billion tokens (TechCrunch, Sept 18 2026).

Source Analysis

TypeSafe blog + TechCrunch + early adopter testimonials. Product claims early-stage.

Research Notes

Additional Sources

Verified Facts

  • jev System One model from diogo-almeida; rlcd training; typed Noul/Choice/Score outputs
  • API demand exceeded capacity briefly; architecture undisclosed

Draft Article

Published as: 2026-09-18-typesafe-jev-calibrated-decision-model

PreScreening Notes

  • Recency: TechCrunch report dated Sep 18, 2026 (~30h old at screening) — passes 48h gate.
  • Score 8 / priority high: New calibrated decision model from RLHF co-inventor Diogo Almeida; 5-18x speedups over LLM classifiers with no hallucination — novel architecture category beyond generative LLMs.
  • Duplicate check: No duplicate in pipeline stages.
  • Audience fit: High for AI developers and ML engineers exploring non-generative decision models.

Evaluation Report

News Value Assessment

  • Timeliness: Fresh — launched week of Sep 18, 2026; API demand briefly exceeded capacity.
  • Impact: High — introduces new model category (“calibrated decision model”) between traditional classifiers and LLMs; potential to displace LLMs in routing, safety, and classification tasks.
  • Prominence: Diogo Almeida (ChatGPT/RLHF co-creator), Vercel and Bryo AI as early adopters with benchmarked results.
  • Proximity: High — Turkish ML engineers actively seeking cheaper/faster alternatives to LLM classifiers; API pricing per billion tokens is developer-relevant.
  • Novelty: Genuinely new architecture category — transformer-based but non-generative; “System One” intuition model vs. “System Two” reasoning.

Audience Fit

  • Primary match: Software developers (API integration), AI enthusiasts (model architecture innovation), ML engineers (classification/routing optimization).
  • Actionability: High — concrete use cases (command safety classification, email routing, jailbreak monitoring) with published speed/cost benchmarks.
  • Topic connection: Extends beyond-LLM AI architecture coverage; directly relevant to military hallucination story (classification without hallucination).

Risk & Ethics Assessment

  • Verification: TechCrunch + developer testimonials (Vercel, Bryo AI) — credible but early-stage product claims.
  • Misinformation risk: Low-moderate — architecture details undisclosed; “cannot hallucinate” claim is structurally valid (predefined output tokens) but accuracy claims need independent validation.

Architecture undisclosed; outside observers suspect open-weight LLM base. Present benchmarks as early adopter reports, not peer-reviewed validation. API capacity issues suggest hype cycle — balance enthusiasm with caveats.

Publication Strategy

Suggested Angle

Türkçe başlık önerisi: “ChatGPT mucidi yeni AI modeli Jev’i tanıttı: LLM değil, kalibre edilmiş karar motoru”

Yazı, generatif LLM’lerin ötesinde yeni bir model kategorisini tanıtmalı. Diogo Almeida’nın TypeSafe AI’daki Jev modeli: transformer tabanlı ama metin üretmiyor, önceden tanımlı çıktılar için kalibre edilmiş olasılık kararları veriyor. “Hallucination” yapısal olarak imkansız çünkü çıktı token’ları sabit.

Teknik çekirdek: Vercel’in OpenAI Luna yerine Jev kullanarak 5-18x hız artışı; Bryo AI’nın Gemini’ye kıyasla 10-20x ucuzluk. Output token’ları ücretsiz, input milyar token bazında fiyatlandırılıyor. “System One” (sezgi) vs “System Two” (akıl yürütme) ayrımı.

Türk geliştirici perspektifi: Agent routing, jailbreak izleme, email sınıflandırma gibi görevlerde LLM classifier’ları değiştirme potansiyeli. API talebinin kapasiteyi aşması erken ilgiyi gösteriyor. Mimari detayları henüz açıklanmadı — dikkatli çerçeveleme gerekli.

Editorial Notes

Onay durumu: Onaylandı — 2026-09-19

Onaylı format: standard (600-800 kelime)

Başlık önerileri

  1. ChatGPT mucidi yeni AI modeli Jev’i tanıttı: LLM değil, kalibre edilmiş karar motoru
  2. Jev: Generatif LLM’lerin ötesinde “calibrated decision model” kategorisi
  3. TypeSafe AI’nın Jev modeli: 5-18x hız artışı, yapısal halüsinasyon imkansızlığı

Zorunlu noktalar

  • Diogo Almeida (ChatGPT/RLHF co-creator) kurucusu
  • Transformer tabanlı ama metin üretmeyen; kalibre olasılık kararları
  • RLCD eğitimi; Noul/Choice/Score typed çıktılar
  • Vercel: 5-18x hız artışı; Bryo AI: 10-20x ucuzluk
  • Output token’ları ücretsiz; input milyar token bazında fiyatlandırma
  • “System One” vs “System Two” ayrımı
  • Mimari detayları açıklanmadı — benchmark’lar erken kullanıcı raporu olarak sun
  • API talebi kapasiteyi aştı

Raporlama talimatları

Grafik Görünümü