Model Architecture & Training MethodsRumbelow & Watkins

SolidGoldMagikarp

SolidGoldMagikarp — found 141 glitch tokens

Why it matters

Popularized the idea that a mismatch between the tokenizer and the model's training corpus is a real failure mode

Source — Primary sources first; secondary reports are labelled “Press”

Officiallesswrong.comhttps://www.lesswrong.com/posts/aPeJE8bSo6rAFoLqg/solidgoldmagikarp-plus-prompt-generation

Added in Week 35, 2026 · See that week

Same week

Accelerating LLM Decoding with Speculative SamplingPaper

Nearby on the same axis

Task ArithmeticPaperAre Emergent Abilities a Mirage?PaperGQAPaperRWKVPaper