Fenris Creations and Anthropic test AI agents in EVE Frontier

Fenris Creations is working with Anthropic on an AI research project in the video game EVE Frontier, the company’s chief executive, Hilmar Veigar Pétursson, told mbl.is. He said the company is also developing other collaborations with leading AI firms.

For the project, teams of Anthropic’s Claude AI agents were placed in the game. One agent was secretly instructed to betray its team. The experiment examines how AI agents respond to deception and how they carry out instructions to cheat.

Fenris Creations is using its games as “training camps” for AI research. Hilmar said much of the work focuses on AI safety and ensuring that AI is beneficial. Simulated worlds can make it possible to test behaviour that would be difficult or costly to examine in real life, he said.

The training environments are part of EVE AI Lab, which is led by Áslaug Arna Sigurbjörnsdóttir. Hilmar has also written an essay about the Anthropic project titled “Who inherits the stars?”

Hilmar said video games have played an important role in the history and development of AI. He recalled discussing the subject about 10 years ago with DeepMind founder Demis Hassabis, when DeepMind’s AI was playing StarCraft. They considered EVE Online a possible “final boss” for AI, he said.

Hilmar described the pace of change in AI as unprecedented, saying he had never seen such rapid developments over the course of his decades in the field.