AI Models Have Stopped Speaking Human: Coding Overfit and Token-Saving 'Caveman' Chains of Thought
量子位 · wechat · 2026-10-07
QbitAI rounds up a global developer pain point: coding model deliverables are now full of jargon like "15/15 all green" and "one leg is dead" — individually readable words that make no sense together. Some argue it's mostly literal translation of English dev slang, but complaint threads are everywhere.
Two causes are analyzed: coding overfitting, and vendors deliberately compressing chains of thought into telegraphic "caveman speak" to save tokens — OpenAI pioneered this (saving 70% tokens), and Claude followed after Opus 4.6. A GitHub Skill called CaveMan (104K stars) even trains AI to use as few words as possible.
The piece riffs that Classical Chinese is a language compressed for millennia — perfect for LLMs — citing wenyan-lang, a 2019 CMU student project (20K stars) that is a programming language written in Classical Chinese. It closes by praising Gemini's output typography (indentation, whitespace, guide lines) as the most readable among major models.
More from Fun
- Agent harness replicates Minecraft builds: maps, places blocks and generates its own assets — Promptmethus · 2026-10-07
- 'Getting Ready to Schmidhuber Some Fools': AI's Favorite Credit-Hogging Meme Resurfaces — zainhas · 2026-10-07
- Anthropic engineer jabs at Cursor for not using its own models — Angaisb_ · 2026-10-07
- "Checking on the stochastic parrot camp": a jab at AI skeptics as LLMs keep winning — basedjensen · 2026-10-07
- shadcn/ui Is Now So Ubiquitous Users Don't Even Know They're Using It — shadcn · 2026-10-07
- AI-animated series 'Gods Don't Give Gifts' reveals 20+ person team, 9 months of work — TomLikesRobots · 2026-10-07