Greg and Dan talk with Dave Johnson of Pearl Technology about Anthropic’s AI research and the use of sandbox environments to safely test powerful AI systems. He explains what a “sandbox” is and how isolated environments have traditionally been used to test potentially harmful software, including viruses, without allowing it to affect outside systems. He also discusses an AI experiment involving open-source code, how the system interacted with its sandbox environment and GitHub, and what experiments like these can teach researchers about the capabilities, behavior, and potential risks of increasingly powerful AI.