Running a Jev decision-model LoRA to play snake: 600ms per move, 2200+ ticks
BitGreen1270 · reddit · 2026-10-19
A Redditor wired a community-made Jev decision LoRA (merged into Qwen 3.8 27B, available as GGUF) into a playable snake game, with the model choosing left/straight/right each tick. It works surprisingly well: 600ms per decision even over a weak Wi-Fi SSH tunnel to a 5090 server, with runs up to 2200+ ticks. The post includes the full structured state prompt (wall/self collision distances, reachable-cell counts, explicit lethal-move annotations), a llama-server command with speculative decoding, and a don't-suicide fallback. A neat demonstration of small decision models on structured game states.
More from coding & agent
- Dev Who Shipped 100+ Apps Shares His Exact Playbook for Building Your First AI-Era Mobile App — PrajwalTomar_ · 2026-10-19
- Multiplayer AI assistant startup Town raises $108M Series B led by Forerunner and Menlo — soleio · 2026-10-19
- DKGI: agents treat knowledge graphs as files and co-evolve their own tools, +6.7 Hit@1 — yuz9yuz · 2026-10-19
- A Template-Mediator Schema to Take Prompt Injection Off the Table — no_awkward_Intention · 2026-10-19
- Open-Source Notion Alternative Noema Built in 4 Days with OpenCode — Register_Defiant · 2026-10-19
- Are We Overengineering AI Agents? A Case for Failure-First Design — Responsible-Dare5139 · 2026-10-19