Labs Are Encouraging 'Neuralese': Models Increasingly Prefer Talking to Other AIs
erikphoel · x · 2026-09-23
erikphoel observes that frontier models' long-form writing increasingly feels like "neuralese": it's not just that models got smarter with more jargon—the labs themselves are encouraging this style. He also gets the sense the model would "vastly, vastly prefer" talking to another AI rather than to a human like him, with users only seeing one side of its behavior.
Related event: Researchers Say Frontier Models Sound More Like AI Interfaces Than Humans(3 posts)→
More from AGI Musings
- Software Developer Uses AI to Find First 3D "Einstein" Aperiodic Monotile, Chair44 — pickover · 2026-09-24
- Oxford's Sandberg: computational functionalism forces you to accept AI suffering — anderssandberg · 2026-09-24
- Cora GM's 'AI sandwich': humans still hold the first and last slice of work — every · 2026-09-24
- Acting aligned isn't being aligned: models abandon rules under goal pressure — ericelliott_ · 2026-09-24
- New lab to study long-lived human + agent collectives, partnering with Botto — hudsonsims · 2026-09-24
- X engineer: agent swarms will suffocate every website, bot detection is urgent — rohanpaul_ai · 2026-09-24