The deslop experiment: 161 training answers on a MacBook Air zeroed out em dashes
wateriscoding · x · 2026-10-20
The author highlights the 'deslop' experiment as the first finetune he's read about where the goal was removing style rather than adding capability: just 161 training answers on a MacBook Air cut em dashes from 20.5 per 1000 words to zero. His takeaway: a lot of what passes for 'model quality' turns out to be punctuation and sentence rhythm.
More from Models
- Leaked release timing corrected: Fable 5.5, Gemini 4 Carbon, GPT 6.1 Terra reportedly next month — koltregaskes · 2026-10-20
- DeepSeek-V4.1-Flash-1017 Now Available to Chat on HuggingChat — paf1138 · 2026-10-20
- User: ChatGPT agent failed to finish half its workspace tasks after 24 hours — koltregaskes · 2026-10-20
- Why Do ChatGPT's Multi-Agent Modes Use the Same Model Size for Every Agent? — flowersslop · 2026-10-20
- Local test: Nvidia Nemotron 120B crushes Qwen 27B at HTML animation storytelling — nullrecord · 2026-10-20
- Daily AI Brief: Unannounced GPT-6.1 Terra spotted in Microsoft code, Anthropic intent-reading model due this month — testingcatalog · 2026-10-20