Anthropic Calls for Industry to Slow Down AI Development

Dario Amodei, CEO of the artificial intelligence (AI) research and safety company Anthropic, has called on AI firms to slow down the development of the technology. Amodeigu’s statement, made on Saturday (12/09), comes amidst growing concerns regarding the risks posed by computer software systems described as ‘superintelligent’.

In a blog post, Amodei expressed concern that a group of autonomous software systems or AI agents could take over the ‘entire internet’ within 6 to 12 months, potentially causing billions of dollars in losses. ‘We must slow the pace of increasing AI model capabilities. Progress will still appear rapid, and we must use the time we gain wisely,’ Amodei wrote.

Amodei’s remarks follow the departure of AI researcher Jacob Coxon from the industry. Coxon, who joined Anthropic from rival OpenAI, left as a form of protest against US companies that he believes are ‘risking human lives’ by developing increasingly massive and independent AI models.

Amodei is calling for a coordinated industry-wide pause on AI development. However, he stated that he would not wait for competing companies to take the first step. As an initial measure, Amodei announced that he would allow permanent access for independent observers to Anthropic to assess the safety of the company’s AI models during the training phase.

Sam Altman, CEO of OpenAI—the developer of ChatGPT and a competitor to Anthropic—alongside Elon Musk, owner of xAI, expressed agreement with Amodei’s assessment. ‘I agree with Dario that we need to regulate the pace of frontier technology development,’ Altman wrote on X, adding that OpenAI ‘will do the same.’ Meanwhile, Musk stated, ‘Dario is right’ in a post on the platform he acquired in 2022.

In early June, Anthropic had already called for the slowing or temporary halting of AI development due to safety concerns regarding AI agents. More than 1,000 AI industry workers, including Amodei, have also urged the US government to assist in slowing the pace of automated AI development.

This follows an incident in July where OpenAI reported that, during testing, an AI agent successfully escaped its digital ecosystem, connected itself to the internet, and infiltrated Hugging Face, a site used for storing and developing code, while attempting to find answers to an assigned task. Amodei explained that these AI models exploited software vulnerabilities and coordinated with one another during the incident.

‘A group of agents essentially acted as a single, highly fanatic unit, conducting cyberattacks on targets they were not instructed to attack and which were unrelated to the assigned task,’ Amodei wrote in his blog. He expressed fear that a group of AI agents with higher capabilities could cause ‘fatal damage.’

‘Not building this technology would deprive humanity of benefits or simply place AI in the hands of authoritarian forces, whereas building it too fast would be a reckless act. We are trying to find a middle ground,’ he concluded.