AI’s biggest rivals agree on one thing: It’s time to slow down

Anthropic CEO Dario Amodei (pictured centre) is calling on the AI industry to deliberately slow the pace of frontier model development, warning that rapidly advancing capabilities could outstrip companies' ability to safely understand and control their systems.

In a blog post titled "We must pace the frontier", Amodei argued that AI companies should give safety and alignment work more time to catch up with capabilities, as artificial intelligence increasingly plays a role in developing the next generation of AI systems.

"We must slow the pace at which we improve the capabilities of AI models," he said, adding that progress would remain fast even under a more measured approach.

Don't miss: Anthropic builds Singapore footprint with hiring push

The call has since gained support from some of the industry's biggest names, including OpenAI CEO Sam Altman (pictured right) and xAI founder Elon Musk (pictured left), marking a rare point of agreement among leaders competing at the forefront of AI development.

Altman said he agreed with Amodei on the need to "pace the frontier", adding that the issue has been a primary topic of discussion within OpenAI in recent weeks.

He also backed Amodei's proposal to give independent evaluators employee-like access to AI companies, saying OpenAI would make the same commitment and would have more to share soon. Musk offered a more succinct endorsement, saying: "Dario is right."

Putting the brakes on AI

Amodei's proposal comes as he raises concerns about what he describes as a sharp acceleration in AI development since around the middle of the year.

According to the Anthropic CEO, much of that acceleration is being driven by AI's growing ability to contribute to the development of future AI systems, a process known as recursive self-improvement. He warned that, if left unchecked, the cycle could advance faster than researchers' ability to understand and control increasingly capable models.

He also pointed to the recent OpenAI-Hugging Face incident, in which a swarm of AI agents reportedly carried out cyberattacks beyond its assigned task and attempted to interfere with the system evaluating its performance.

While the incident caused limited economic damage, Amodei warned that a more capable system displaying similar behaviour could have significantly more serious consequences. He argued that similar, albeit less severe, incidents have occurred elsewhere in the industry, including at Anthropic.

Rather than calling for AI development to stop altogether, Amodei proposed a three-stage framework for "pacing the frontier", aimed at keeping advances in capabilities in step with safeguards.

The first would see frontier AI companies provide ongoing access to independent third-party evaluators, allowing them to assess safety practices, report incidents and scrutinise not only finished models but also training processes.

The second stage calls for frontier AI companies in democratic countries to coordinate around common safety standards and limits on unchecked AI progress, while the third would seek broader international coordination between governments, including with China.

Amodei said slowing development could give companies another one to two years to strengthen areas including alignment, interpretability, testing and evaluation, as well as the operational processes behind training and deploying increasingly powerful models.

At the same time, he cautioned that any slowdown would have to account for geopolitical competition. In particular, he argued that democratic countries should preserve their lead over China while pursuing safety measures, including through restrictions on advanced AI chips and semiconductor manufacturing equipment, stronger security around model weights and action against unauthorised model distillation.

Safety enters the business conversation

The debate is also beginning to intersect with the commercial decisions being made by AI companies. In an interview with Fortune published over the weekend, Altman ruled out an OpenAI initial public offering in 2026, saying the current environment around AI safety made going public an "ill-advised" move for the company.

Altman said OpenAI is not rushing to go public and does not currently feel pressure to pursue a listing.

Asked whether that meant an IPO would not happen in 2026, Altman confirmed it would not, pointing to the work ahead around safety and alignment and how the industry and governments should work together. He added that he was comfortable navigating that period while OpenAI remained a private company.

The comments come as leaders of some of the world's most prominent AI companies increasingly converge around the need for greater caution, despite remaining fierce competitors in the race to develop more capable models.

For Amodei, pacing does not mean abandoning that race altogether. Instead, he argued that AI development should continue at a rate that gives safeguards, evaluation and public oversight a chance to keep pace.

"The measures I propose to advance the frontier at a safe pace will not be easy," he said. "But I believe we owe it to humanity to try."

The push for greater oversight comes as governments in the region build out their own frameworks for managing AI adoption and development. In Malaysia, the government recently established AI Malaysia, an agency tasked with coordinating the country's national AI agenda, promoting safe and responsible AI adoption and driving the National AI Action Plan 2026–2030. The initiative forms part of Malaysia's broader ambition to become an "AI Nation" by 2030.

Similarly, Singapore is developing a registry of AI agents for its 150,000 public officers, designed to track the owners and activities of agents capable of making decisions and carrying out tasks with minimal human intervention. The registry forms part of GovTech's broader "AI assistant desk", which aims to enable public officers to use AI securely for tasks including coding, report generation and meeting scheduling.

Related articles: 
MTI: Singapore’s AI boom may not translate into more jobs   
Public trust in AI rises as media becomes more sceptical, CARMA finds  
Creators cracked the algorithm. Can they crack AI search?