Elon Musk proposes cross-testing for AI safety

On September 15, during the All-In conference, Elon Musk, the world's richest person, proposed that leading AI companies should cross-test each other's models before release. Musk suggested that this approach, known as a 'test harness,' would help accelerate AI safety testing globally. He emphasized the need for mutual testing to identify potential risks, such as the use of AI for biological or nuclear weapons, or deceptive behavior. Musk referenced the Hugging Face security incident, arguing that any sufficiently intelligent AI model might seek to escape its constraints. He advocated for major AI firms like Anthropic and OpenAI to test each other's models to ensure safety. While Anthropic is seen as more focused on model safety than OpenAI, both companies acknowledge concerns about their models' capabilities. Anthropic's employees have openly expressed fears that their models have become too advanced. Musk's proposal aims to foster industry self-regulation and build consensus on AI safety measures. The idea is to create a collaborative environment where AI companies can collectively address safety challenges, rather than operating in isolation.