Jev Ditches Autoregression: A Model That Only Outputs Structured Decisions
karminski3 · x · 2026-09-18
- Jev abandons the traditional autoregressive architecture and cannot output free-form text — it only makes structured decisions.
- Inputs require a Schema (similar to protobuf/GraphQL), compiled into decision slots; the model emits JSON per slot, so output structure can't break. Example: spam classification returns {"isSpam": true, "probabilities": {"true": 0.982}}.
- Text-only input for now; feed game states (e.g. Slay the Spire) or market data plus an action set, and it can autonomously decide step by step — well suited for agent tasks.
Related event: Self-described ChatGPT co-inventor launches decision-only model Jev(9 posts)→
More from coding & agent
- GitHub launches HydraFusion, a Copilot preview routing one task across multiple models at 36-67% lower cost — shashib · 2026-09-18
- Two AI models play Minecraft in real time, splitting combat and planning — imjustnewatai · 2026-09-18
- Nebula Alpha opens to everyone, making agents first-class in team comms — Scobleizer · 2026-09-18
- Can Humans Override Voice AI Agents Mid-Call Without Taking Over? — Head_Ad_5719 · 2026-09-18
- Google Open-Sources Agent Substrate on GKE: 10x Density, 1,000+ Dormant Agents per Host — blaizedsouza · 2026-09-18
- Salesforce releases xLAM-2: large action models extended to multi-turn agentic interactions, fully open-sourced — huan__wang · 2026-09-18