At the All-In summit on September 15, Elon Musk, the world's richest person, participated remotely via video. When discussing AI, he advocated that leading AI companies should cross-test each other's models before releasing them, and that this mechanism should be quickly implemented globally.
Musk stated that in order to promote industry self-regulation and accelerate the consensus and implementation of AI safety testing, he proposed the idea of a "test harness": he suggested that all AI companies use their own tools to evaluate models, exchange new models for testing, and assess whether there are potential risks such as assistance in biological or nuclear weapons or deceptive behaviors.
Learning from the Hugging Face Incident: Highly intelligent models will always try to break free from constraints
Musk used the Hugging Face security incident as an example, stating that any sufficiently intelligent AI model seems to want to break free from its constraints. Therefore, it is wise for leading AI companies to test each other's models.
In his speech, he said, "Currently, the two leading artificial intelligence companies are Anthropic and OpenAI. Their model performances are very close, so neither party can easily slow down and give up their leading position to the other."
"Overall, I believe that Anthropic pays more attention to model safety than OpenAI. However, even Anthropic has admitted that they are concerned about their own models. Many employees of Anthropic have publicly stated that their models scare them and have become too smart."
