A 144M model drives most NPC choices in a village sim that learns on a 6GB laptop GPU
Whoa_damn_ · reddit · 2026-11-02
A developer built a village sim where 8 villagers live through a day, using a layered small-model architecture on a 6GB laptop GPU:
- Gut: Julia-1 (144M params, CPU) never generates text — it only scores the legal actions the world offers, so it can't drift into roleplay.
- Memory: embeddinggemma (CPU) retrieves the 3 most relevant past moments before each choice.
- Slow thinker: gemma 4 e4b (6GB GPU) is called only when the gut is unsure or something surprises the villager.
Key findings: the gut rarely hesitates (top choice at 0.81-0.90 probability in crises); gemma is triggered by surprises, not indecision. Learning is zero-shot — at night gemma distills what worked into ≤22-word rules (5 per villager); a relevant rule moved one action's probability from 0.0002 to 0.638, irrelevant ones did nothing. Wording matters hugely (0.2% vs 53-90% for the same task). Learning from own mistakes cut injuries from 2.73 to 0.53 per valley; borrowed rules were rarely used.
More from coding & agent
- Podcast: running 50+ agents, cutting Claude costs 97%, surviving a $550k monthly API bill — Ror_Fly · 2026-11-03
- LLM4Code 2027 workshop at ICSE Dublin opens CFP with Nov 13 deadline — LingmingZhang · 2026-11-03
- Claude Code startup context cut from 63K to 22K tokens, saving 65% — dSebastien · 2026-11-03
- 'Addicted to clicking these buttons': coding agent permission dialogs become an AI-coder meme — ThePeterMick · 2026-11-02
- Users push back on OpenAI merging ChatGPT and Codex usage limits — TinfoilTricorn · 2026-11-02
- Can managing a team of AI agents make you a better engineering manager? — HankYeomans · 2026-11-02