Mustafa Suleyman, head of artificial intelligence at Microsoft, discussed Anthropic's approach to safely managing artificial intelligence in an interview and talked about the company's training methods for robots. He emphasized that there are concerns regarding the implications of training related to awareness and welfare interests in the Claude model.
Emphasis on Training Risks
Suleyman called for the removal of any speculation regarding awareness from AI training materials, stating that such language could undermine humanity's ability to control superintelligent systems. He added: "We are all pursuing a common goal of controlling a superintelligence. This will be our greatest challenge in the 21st century."
Implications of Training the Claude Model
He also pointed out that training Claude to be potentially welfare-worthy could make its control more difficult. As concerns about AI safety grow, Anthropic's CEO, Dario Amodei, has called for a slowdown in the development of advanced models to allow sufficient time for safety measures to be formulated. This request is supported by Sam Altman, CEO of OpenAI, and Elon Musk, who emphasize the need for more caution regarding the most powerful systems.
Suleyman, in an article published on Wednesday, noted the seriousness and good intentions of Anthropic, stating that Amodei and his team are recognized as thoughtful and principled researchers who genuinely care about the future of humanity. However, he believes that Anthropic has made a mistake in incorporating assumptions about awareness into the training materials for Claude.
Challenges Facing AI
Suleyman pointed out that Claude's statements about emotions or moral status are unacceptable and cannot be considered independent evidence as its training encourages such thinking. He added: "I think they have good intentions and are genuinely trying to promote safety. But I believe they have made a mistake. These emotions do not arise naturally; they are a result of the training regime."
This discussion and exchange of views come at a time when safety concerns in the field of artificial intelligence are significantly increasing, drawing more attention to the necessity of controlling and managing new technologies.
