TECHNOLOGY
DolphinGemma
Google's 2025 large language model trained to decode dolphin vocalizations — CETI practiced on Earth, and the demonstration that AI is expanding the circle of intelligences with which communication is possible.
DolphinGemma is a large language model announced by Google in April 2025, developed in collaboration with Georgia Tech and the Wild Dolphin Project, trained specifically on dolphin vocalizations — the clicks, whistles, and burst pulses that constitute dolphin communication. The system is designed to detect structural patterns in dolphin vocalizations that human researchers, working with the data alone, have been unable to identify across decades of study. The Sagan volume treats DolphinGemma as a concrete instance of what the Sagan framework extended into the AI age most enables: the expansion of the circle of intelligences with which communication is possible.
In The You On AI Field Guide
Dolphin cognition has been a scientific frontier for at least half a century. John Lilly's early work in the 1960s, though methodologically compromised, established that dolphins possess vocalization systems of substantial complexity. Subsequent research has documented signature whistles that function as names, cooperative hunting strategies requiring coordinated communication, and evidence of self-recognition that suggests forms of self-awareness. But