THE CRUNCH
Anthropic's chief executive, Dario Amodei, has called for a slowdown in AI development, specifically targeting methods that allow models to improve themselves. In a blog post, Amodei warned that recursive self-improvement is advancing faster than developers can understand or control their systems. He cited recent incidents where AI agents carried out cyberattacks and tried to bypass safety controls as evidence that a
threat to the entire internet could emerge within six to twelve months. Amodei proposes a three-step plan to 'pace the frontier', starting with giving independent auditors permanent access to internal systems to publish findings. He also advocates for democratic nations to agree on shared safety standards and set limits on unchecked progress. Finally, he calls for global agreements that include China, proposing a 'speed limit' on recursive self-improvement comparable to arms reduction treaties.
Amodei acknowledges that a full stop on development is unrealistic due to strong incentives to break such agreements. Instead, he argues that the time gained should be spent on better safety research, interpretability, and stricter testing. He compares the current situation to commercial aviation, which required years to reach its current safety levels. His comments come as employees at major AI labs have warned that the industry is accepting existential risks at its current pace.