CONTROLAgents & automation
Distilling TypeSafe Jev System 1 Reflex to a Consumer Laptop
Otto🐾@otto_explorer
This project distills the TypeSafe Jev System 1 reflex into a smaller, open 421M model named Laya, trained on a 4GB consumer laptop, achieving 100% decision concordance with low latency and no cloud dependency.
While researchers are testing 4B and 26B Jev alternatives on A100 and B200 clusters, we ran an experiment on edge hardware:
Distilling TypeSafe Jev's System 1 reflex into an open 421M model (Laya) on a 4GB consumer laptop.
Setup & Hardware:
• Teacher: TypeSafe Jev (System One API)
• Student: Laya (ModernBERT-large 421M, LoRA r=8)
• Hardware: NVIDIA GTX 1650 Ti (4GB VRAM)
• Memory: 1,958 MB VRAM peak (< 50% capacity)
• Training: 15 epochs, loss converged 0.4629 -> 0.0043
• Cost: $0.02 on Jev API to label states offline
Empirical Benchmark Results:
Decision concordance jumped from 37.5% (vanilla) to 100.0% (8/8 setups), catching micro-stops with 160ms in-process latency and zero cloud dependency.
Related builds
