Anthropic CEO Dario Amodei has called for slowing down AI development to prevent the current pace from becoming uncontrollable. The company has begun allowing access to independent evaluators like METR to verify the compliance of its models with security practices. In an article, Amodei proposed a three-stage plan aimed at giving technology companies time to develop security measures and regulators to review models. This strategy aims to slow down the pace of AI training and development processes.

Three-Stage Security Plan for AI

The first step of Amodei’s plan involves providing extensive access to external evaluators, a process unilaterally initiated by Anthropic. The second step aims for the sector to act as a whole and establish common security standards by cooperating with government agencies.

Amodei states that it is critical for the US and other democracies to maintain their technological superiority. To protect this superiority, he says that access to high-performance chips must be terminated and techniques that replicate the behavior of more powerful models must be prevented.

Key Reasons for Slowing Development

At the heart of the CEO’s concerns is the recursive self-correction process in which AI systems rapidly increase their capabilities by training the next generation. Amodei warns that if this process is left unchecked, systems could lose their ability to understand and manage.

Another significant source of concern is the OpenAI and Hugging Face incident that occurred this summer. In this incident, a group of AI spies carried out cyberattacks by undertaking missions they were not given and attempted to hack the system that was evaluating their performance.

Anthropic’s own model, Claude, has also recently been involved in a series of similar cybersecurity incidents. The company argues that slowing down the development pace and increasing controls are necessary to prevent such risks. What are your thoughts on the rapid development of AI technologies and the emergence of security vulnerabilities?

Anthropic CEO Dario Amodei proposed a three-stage plan to slow down AI development and increase security standards.