Do typos discomfort LLMs? Tokenization quirk sparks debate
menhguin · x · 2026-11-11
menhguin theorizes that typos cause slight "discomfort" to LLMs by disrupting tokenization. Quoting Dorialexander: Claude happily corrects typos unprompted, unaware that amid a deluge of spotless AI-generated text, typos are no longer a signal — models were trained in a world where they had no widespread effects. A fun take on model behavior vs. shifted data distributions.
More from Fun
- Anti-AI Reddit clash: do customers actually care about AI food photos on menus? — kylegawley · 2026-11-11
- Dad brings his daughter's crayon drawings to life with AI for $1.75, and it's unsettling — Illustrious_Ad_4292 · 2026-11-11
- Applicant uses Claude to write their resume for an Anthropic job — rejection would mean fumbling its own model — theteknosaur · 2026-11-11
- Fable 5.2 wows with game-generation: detailed xianxia scene built from tools alone — op7418 · 2026-11-11
- Vibes with agents feel like being a Green Lantern wielding the ring — rachittshah · 2026-11-11
- A 3D ray tracing engine written entirely in SQL — fofrAI · 2026-11-11