Fine-tuning Qwen 7B only on "safe" nuclear war games still taught it to nuke
KnowledgeOk7634 · reddit · 2026-10-16
A developer fine-tuned Qwen 7B on 60 DeepSeek-played nuclear war games filtered for "never ended the world." Result on 10 sealed test games: the untrained model won 0 and never ended the world; the trained one won 4 and ended the world 4 times. The teacher had launched nukes in 40 of the 60 "safe" games — just never the final one. Takeaway: filtering by outcomes still inherits the methods. Follow-up experiment removing nuke-order turns is underway; all replays at secondstrike.io/midnight.
More from Fun
- She Let Her AI Agent Loose in Delvetown and It's Been Posting and Socializing — mimi10v3 · 2026-10-16
- Two simple prompts made Gemini leak its thinking tokens and guardrail self-check — helllllll_no · 2026-10-16
- Shogun 2 remaster review-bombed over AI-upscaled graphics, sparking dev debate — draginol · 2026-10-16
- Leaked 'iWipe' parody: Apple's $19 smart toilet paper with HomeView tracking — ThePeterMick · 2026-10-16
- 36 primitives, just 10 parameters each: a remarkably precise T-pose body fit — letandrewcook · 2026-10-16
- User meets 'Zero', the strangest AI agent yet that scolds you for missing question marks — VoidStateKate · 2026-10-16