Beyond 'not X but Y': small team says LLMs also overindex on 'content slop'
_akpiper · x · 2026-11-10
The team behind writing model Deft claims a breakthrough: their new training algorithm (invented by partner @rosmine and refined over months) generates creative prose uncorrelated with top LLMs, making Deft the only model that scores 100% human on Pangram 90% of the time.
Their key insight: big LLMs don't just produce "style slop" (like the mocked 'not X, but Y' patterns) but also information slop — overindexing on certain substantive ideas, not just stylistic tics. As one commentator noted, solving AI creativity means also solving AI detection of AI writing.
More from Models
- Gemini 4 Carbon goes live at the same prices as the previous generation — ColbyHawker · 2026-11-11
- SemiAnalysis: Claude Fable 5.2 spends 4.3x more output tokens for the same GDPx score — dylan522p · 2026-11-11
- Unverified: Gemini 4 Carbon reportedly rolling out to developers and consumers — tulseedoshi · 2026-11-11
- Google rolls out Gemini 4 Carbon, a flagship cybersecurity model leading code-security benchmarks — ralucaadapopa · 2026-11-11
- OpenAI Realtime API adds India-specific SIP endpoint after EU data-residency option — juberti · 2026-11-11
- Google DeepMind ships Gemini 4 Carbon, which ports 90s game Theme Park to web in one shot — demishassabis · 2026-11-11