Anthropic Report on AI Accelerating Its Own Development Alarms Researchers

Artificial intelligence is quickly speeding up research at Anthropic, allowing its models to improve twice as fast as they did a year ago. This rapid progress comes to light even as the company's leaders publicly urge the industry to slow down AI development.

The findings come from a new report by the AI safety research group METR, which spent 11 weeks investigating inside Anthropic. METR researchers stated they are confident the acceleration is not simply the result of adding more computing power. Instead, they concluded that artificial intelligence itself is driving most of the speedup.

Employees reported huge productivity gains while using the company's internal models. Staff members estimated they were about 12 times as productive with AI help than they would have been without it, which is double the boost reported just months earlier.

This acceleration has raised alarms among outside researchers because it points toward recursive self-improvement. This is a scenario where AI systems can build more powerful successor versions on their own with little or no human help. Critics worry such fast advancement could soon outpace human ability to understand and control the technology.

The report also revealed that Anthropic finished training three unreleased models, named Model 2, Model 3, and Model 4, over a six-month period. However, public access to the full findings is limited. Anthropic used its agreement power to heavily redact the final report before its release, keeping much of the underlying evidence secret.

Anthropic co-founder Sam McCandlish defended the redactions in a post on X. He acknowledged that the secrecy makes it harder for others to check the work, but argued that sharing specific details would be commercially sensitive and provide a dangerous roadmap for competitors to speed up.

The heavy editing and the fast pace of internal progress have fueled frustration among safety experts. Critics argue there is a sharp clash between Anthropic’s public calls for safety and responsible pacing, and its own aggressive internal development.

Summary

A new independent investigation reveals that Anthropic's AI models are rapidly accelerating their own development, pushing the technology forward even as the company's executives call for the industry to slow down. The findings highlight growing concerns over recursive self-improvement and the challenges of maintaining transparency in frontier AI research.