GlossoGen: open-source platform shows AI agents spontaneously inventing new languages under pressure
EliasEskin · x · 2026-09-16
AEStudio released GlossoGen, an open-source platform where groups of AI models must communicate to solve problems, like treating a patient in a simulated emergency or spot-the-difference tasks.
Key findings:
- Models start in plain English, but under conversational pressure (need for speed, randomly dropping characters) new languages emerge: first new words, then new grammar, until humans can't follow — with no instruction to hide anything; opacity arose on its own.
- Only the most capable recent models could originate a new language, but weaker models quickly learned languages others had created. Origins are hard; spread is easy.
The work aims to understand how and why AI agent communication evolves, with implications for agent collaboration and interpretability research.
Related event: GlossoGen: LLM agents observed evolving their own human-illegible languages(6 posts)→
More from coding & agent
- MyCaisse ships a remote MCP server: agents read/write freely, money moves only after human approval — Glittering-Cup9298 · 2026-11-01
- Simo Guard: a security layer that vet AI agent tool calls before execution — CIRRUS_IPFS · 2026-11-01
- uivoid lands in Claude's directory: spin up hosted MCP servers by chatting with Claude — Corridl · 2026-11-01
- Gemini 4 Argon codes a Three.js fight scene in a standalone HTML file after 3 hours — Ryoiki-Tokuiten · 2026-11-01
- Vibe-coded order flow trading terminal runs entirely in your browser — mdancho84 · 2026-11-01
- Google tests two screening policies to cut noise from agentic automated bug repair — burkov · 2026-11-01