Building a local Jev-style System One model: fine-tune 300M or build a custom decision engine?
AbbreviationsLoud182 · reddit · 2026-10-02
A Reddit poster weighs strategies for building Jev-style "System One" models for structured decisions — classification, routing, scoring, confidence estimation, agent action selection.
Three paths analyzed: (1) fine-tune a small decision model like Laya (300M) — cheap, fast, small footprint, but architecture-limited and the author's zero-shot domain classification tests fell short; (2) fine-tune/distill a larger CLM-8B — better generalization but VRAM and serving costs feel excessive for "which intent?" tasks; (3) build a custom 300M-500M decision engine mapping state + question + allowed options to a calibrated probability distribution, with uncertainty and abstention as first-class objectives — more data and harder debugging required. The favored middle ground: pretrained multilingual encoder → general decision training → domain decision data → optional adaptation.
More from coding & agent
- Three reasons vibe-coded software is still far from production grade, with ReactBench data — aidenybai · 2026-10-02
- GitHub Copilot adds computer use to control desktop apps in public preview — PaulShellDev · 2026-10-02
- Dev builds her first Game Boy-style game entirely with OpenAI agents — craigsdennis · 2026-10-02
- When the Trace Looks Fine but the Agent Output Is Wrong — Sensitive-Parsnip-12 · 2026-10-02
- iPad + Tailscale + Codex is this dev's new favorite way to work outside the office — flavioAd · 2026-10-02
- Will Cloud AI Agents Need Residential IPs? Amazon Blocked Meta's Muse — SignificantFail3632 · 2026-10-02