37 AI novels tested: 8k context cuts repetition from 42.8% to 0.5%
lmoroney · x · 2026-10-12
Jessica K. Murray generated 37 full novels across 8 local models and 3 machines; bigger isn't better. Key findings:
- Biggest win was lowering a setting: gemma4:12b repetition dropped from 42.8% at 32k context to 0.5% at 8k.
- An old 6 GB RTX 2060 running MoE gemma4:26b wrote a clean 65,000-word book in 4 hours.
- Only 1 of 5 quality tools ranked the books sensibly; one scored 91.2/100 to a manuscript repeating a sentence 500 times.
- Quick check: if chapters keep shrinking from chapter 1 to 30, the model is looping—fix settings first.
Related event: Test of 37 AI Novels Finds 8k Context Cuts Repetition to 0.5%(2 posts)→
More from Models
- Opus 5.5's fondness for "carried" and "held" sparks Opus-4.6-as-grader speculation — Sauers_ · 2026-10-12
- OpenAI raises the plane's chromatic number lower bound to 6, Lean-formalized — fortnow · 2026-10-12
- Liquid AI open-sources Open d1 decision models: d1-3B runs at 35ms per image on Jetson Thor — TheTuringPost · 2026-10-12
- Grok Bot is winning the agentic personal assistant race, says dev — Arindam_1729 · 2026-10-12
- Does Codex refuse or cut back work when its own budget estimate runs high? — r618NecessaryStation · 2026-10-12
- AI 2027 Update: Reality Tracking at 70-90% of Forecast, Superhuman Coder Slips to 2028 — jessi_cata · 2026-10-12