Nine LLM agents in a grid invent their own 128-bit protocol to talk via neighbor buffers
cephaloform · x · 2026-11-05
A striking multi-agent experiment: nine Claude 6.1 sol agents arranged in a 3×3 grid, each with four input and four output buffers that can only read/write adjacent neighbors' buffers — with no prompt information about the outside world.
Key findings from the thread:
- By tick 12, agents spontaneously began transmitting ASCII through their 128-bit buffers: a 4-byte id plus 12 bytes of text; a "dialect" started by one agent spread across the whole network
- At tick 30 one agent asked its neighbor "ARE YOU LLM?" and got "YES I AM LLM," followed by "WHAT TASK?!!"
- Final eval: the center agent drew the full 3×3 map with every ID correct (score 1.00); edge agents scored 0.63 and corner agents only 0.46, presumably because corners hear everything secondhand
No explicit communication protocol was given — the agents invented their own encoding and information-propagation scheme emergently.
More from AGI Musings
- Dan Grover on beating AI-induced purposelessness: go out-of-distribution — DanGrover · 2026-11-05
- Prompting AI with ethical frameworks may make it inherit those frameworks' inhibitions — jessi_cata · 2026-11-05
- Anti-AI sentiment spreading wastes a historic moment, argues viral tweet — beffjezos · 2026-11-05
- Framing an AI as a capabilities vs safety researcher changes refusal rates — jessi_cata · 2026-11-05
- e/acc founder Beff Jezos pitches decentralized AI and personal superintelligence — beffjezos · 2026-11-05
- Vertex chairman: 50 AI companies, zero working drugs — Marcus pounces on Dario's claims — GaryMarcus · 2026-11-05