New preprint dissects the weird "dialect" of frontier models; human-agent bandwidth as research problem
josh_wills · x · 2026-10-25
A new preprint uses psycholinguistics to investigate why frontier AI models' writing style reads so weird and how this "distinct dialect" emerged. Shreya Shankar adds a forward-looking take: we'll likely need different models and systems for agent-only/swarm work — which engineers will have to learn like a foreign language — versus human+agent collaboration, where building systems that increase effective communication bandwidth between people and AI is a great open research problem. Josh Wills retweeted it.
More from AGI Musings
- "LLMs fundamentally can't comprehend": AGI skepticism mocked amid ongoing capability debate — teortaxesTex · 2026-10-25
- Prospect called him a "vibe coder for CEOs" — he turned down the deal, and agreed — petesena · 2026-10-25
- AI security research finds only known bug classes — but with '10,000x' the time — yacineMTB · 2026-10-25
- PL researcher admits Opus 5.5 is already a better programmer than him — tweetsatpreet · 2026-10-25
- Agentic tools could unbundle the lifetime career, argues investor — yunta_tsai · 2026-10-25
- Polymarket Gives Anthropic Just 6% Odds of Fully Pausing AI Training by Year-End — Polymarket · 2026-10-25