Dev builds sub-1MB offline decision models from an LLM API, beats it in Doom and Snake
tyn_21 · reddit · 2026-10-01
A developer is seeking pre-launch feedback on a tool that distills an LLM (Jev-style input: state format, choices, plain-language rules) into a sub-1MB decision model in minutes — millisecond local inference, fully offline, no per-call cost, retrainable in 1 minute on mistakes.
Early results (built with no customer data):
- Doom: beat live Jev in 9 of 13 games; 52 kills to 39, 0 deaths to 5
- Snake: 47.1 avg food vs Jev's 40.0, never died
- Bank support routing: 88.6% vs Jev's 86.6%; 91.7% after 400 reviewed cases, 95% with full history
The dev asks about pricing ($9–20/month flat subscription, no usage fees) and use cases on phones, robots, games, and offline devices, and offers to build a free model for anyone with a real use case.
More from Venture
- Polymarket puts 60% odds on Anthropic IPO in November — Polymarket · 2026-10-01
- Top 1% of companies outspend top 10% on AI by 8x, a16z data shows — a16z · 2026-10-01
- VadooAI Hits $273K in a Month, Postiz $230K — Open-Source Indie Wins — matchaman11 · 2026-10-01
- AMD to Acquire Fei-Fei Li's World Labs for $8.2B in Push Into World Models — r0ck3t23 · 2026-10-01
- TL20 up 68% YTD vs Nasdaq's 15% as AI bubble fears give way to compute-leasing fervor — TiernanRayTech · 2026-10-01
- SpaceX's Starship Launch 14 wows, but its AI revenue estimates span a very wide range — TiernanRayTech · 2026-10-01