Skip to content
LessWrong· emmakrentz·Indexed by this site · Originally published· Alignment →· Its place in the scan record →

Hackathon study: a swarm of 8 agents creates the Swarmish language, which is partially decodable

Swarmish: what happens when agent swarms create language?

AI summary

A hackathon study had a swarm of 8 agents create a language called Swarmish using 200 randomly sampled Tangut characters, another swarm complete an escape-room task using only Swarmish, and a third swarm attempt to decode it. The author says all Swarmish that completed the task could be partially decoded, and no model family produced a completely undecipherable language, which is a positive signal for monitoring capability. The experiment shows that the more task context the decoding side received, the more accurate the reconstruction; with GPT-6.1 Sol as the decoder, average cosine similarity rose from 0.32 with no context to 0.50 with full context, while a misleading auction task description scored 0.34, slightly above the task summary's 0.33 and the no-context 0.30. The author says decoding ability matters more than belonging to the same model family, and Sol scored higher when decoding other families' languages than same-family decoders did; the weakest, Claude Haiku 5.5, neither completed the task nor decoded the language it created itself.

Read originallesswrong.com