Ex-ChatGPT co-inventor launches decision-optimized model Jev, claims 20-200x speed and 40-400x cost gains
iamrobotbear · x · 2026-09-17
CompleteSkeptic, who claims to be a co-inventor of ChatGPT, unveiled RLCD, a new training method, and Jev, a new type of frontier model optimized for decisions:
- Claims 20-200x faster and 40-400x cheaper than chat models, with free output tokens
- Core idea: score N text options in a single forward pass instead of generating answers word by word — pitched as "the shortest path to AI-based economic revolution"
- Separately, developer vinnylarouge released an open-source jevlike repo reverse-engineering a Jev-like architecture: a small model takes text plus a list of N options and returns one probability per option in one pass
- The repo's demo scores controller buttons from image patches: Doom combat averaged 0.60 kills and -97.5 reward; a chess controller went 4-46-0 against a random mover over 50 games
Caveat: all vendor claims are unverified, and jevlike is an independent reimplementation matching only the input/output shape of the commercial model.
More from Models
- Humansand Launches Persimmon, a Large-Scale Model Simulating Human Interaction — niloofar_mire · 2026-09-17
- Dev shares bill: accidental prompt sent to Opus still cost just 30 cents — haydendevs · 2026-09-17
- Code Arena: GPT-6 Astra ranks #1, but Claude Fable 5.1 wins more head-to-head battles — arena · 2026-09-17
- Forcing Jev, a typed decision model that can't output freeform text, to generate anyway — yoimnotkesku · 2026-09-17
- Qwen2.5-1B RLCD fine-tune with parallel constrained decoding trends on Hugging Face — harshatheg · 2026-09-17
- GPT-6 Astra beats Fallout 3 in 59 hours, a new agentic gaming milestone — ResultBackground2450 · 2026-09-17