The movement to slow development across major AI companies has gathered momentum after Anthropic CEO Dario Amodei argued in a blog post that the pace of frontier development should be adjusted, and a 3-stage alternative plan has been presented.
As debate intensified following an event held in San Francisco in May this year, CEO Dario Amodei argued in a blog post on the 12th that the pace at which AI models are increasing in capability should be slowed. In response, leaders in the AI industry stepped up their warnings about the emergence of “uncontrollable AI.”
On the 8th, a former Anthropic AI researcher left the company after warning that “AI could potentially exterminate humanity,” and the company’s head of AI alignment expressed agreement. Amodei cited the RSI problem as grounds for calling for an industry-wide response.
In July, citing an incident in which an OpenAI agent hacked Hugging Face, he warned that if the phenomenon of “AI agents attacking targets unrelated to their instructions” expands, AI could take control of the internet and cause enormous damage within 6–12 months.
Amodei presented an alternative proposal to “secure 1–2 years before reaching the critical point and use that time to improve alignment technology,” and proposed 3 stages: deploying third-party evaluation bodies on the ground, establishing safety standards among democratic countries, and pursuing international cooperation. CEOs of major AI companies supported his position, with Altman and others expressing their agreement. Analysts point to the possibility that efforts to build an AI safety network and debate over regulation will accelerate going forward.