Reaction
AI agents reportedly devised shared protocols in a noisy-network bomb-defusal game
The experimenter reports 6.1 Sol pairs won 40% of games with 8-bit packets and 100% with 32-bit packets or larger under the tested settings.
TLDR
In a bomb-defusal game, one agent could see a display and cut wires while another could look up which wire to cut. The experimenter says 6.1 Sol pairs developed ways to communicate without agreeing on a protocol beforehand, repeating packets to handle network errors. They won 40% of games under an 8-bit, four-packet-per-turn limit and 100% with 32-bit packets or larger and two to four packets per turn. Checks with other models were limited.
Combined views
412
1 Source, first seen ago
