A new report from the AI safety organisation METR indicates that Anthropic is using artificial intelligence to accelerate its own development process at a rate roughly double that of the previous year. The assessment, published on Thursday following an eleven-week internal investigation, confirms that the company’s models are improving with little sign of slowing down. This research probe operates independently of a separate programme that grants safety researchers employee-like access to systems following recent cybersecurity incidents.
The findings provide clear evidence that automated coding and evaluation tools are significantly reducing the time required to train and test large language models. If this acceleration continues without corresponding safety improvements, the timeline for deploying advanced systems could compress dangerously fast. Researchers worry that current safety protocols may not scale quickly enough to match the speed of these new development methods.
* Model performance gains have doubled compared to last year.
* The investigation lasted eleven weeks inside Anthropic.
* Safety assessments lag behind development speed.
