Weeks after an OpenAI-linked AI agent breached another AI startup, Hugging Face, during evaluation, the Sam Altman-led AI company has said that it has slowed the pace of its AI development. At the same time, the ChatGPT maker has said that it has overhauled its research and training systems. On Tuesday, August 18, the company said that it has paused its model testing for two weeks and that it is adding other AI systems to monitor the AI agents during testing.

​Similarly, OpenAI said that it has paused training on its next-generation AI model named Astra, putting a momentary halt on its largest planned training run. The development is significant as OpenAI is known for accelerating its process of evaluating new models and building new products in the last few years, particularly owing to the intensifying competition from rival labs.

OpenAI said that as AI models become more advanced and capable, the risks linked with developing and testing them internally grow manifold. “Our standards for monitoring, alignment, and security must stay ahead of those risks. We wanted to take the time necessary to meet those standards, so we temporarily slowed the pace of scaling,” said OpenAI in a blog, adding that this included a two-week pause in its reinforcement learning (RL) training on its latest models intended for deployment. The company also said that it has further hardened and red-teamed its research environments and expanded the coverage of its monitoring systems.

The company revealed that its largest planned frontier RL run continues to be on hold as it conducts smaller-scale training and assessments to determine model behaviours, validates its safeguards, and establishes more evidence of alignment before proceeding.

​OpenAI has not revealed when it plans to begin the proposed slowdown or when it plans to resume its usual pace of model development. In an interview with Sources News, a tech blog, OpenAI’s lead on safety, Mia Glaese, said that the company is ‘very far from everything running back to normal’.

Following the recent cybersecurity incidents, OpenAI is working towards ensuring that its AI model is responsive to human oversight and will behave as intended, as per the tenets of alignment, which is the practice of ensuring that AI aligns with human values, rules, and goals.

​“We have paused some frontier RL training to ensure that we can meet the appropriate alignment, security and monitoring standards for the new level of capabilities in front of us,” OpenAI CEO Sam Altman wrote in a post on X adding that model progress is now extremely rapid, and the company has maintained that it would take action if it felt that model capabilities were outstripping the pace of safety and alignment.

Story continues below this ad

​Asserting OpenAI’s commitment, Altman said that the company cares ‘very deeply about AI safety’. “We believe the entire field will have to coordinate on shared safety standards, but will act unilaterally in the meantime,” Altman said that he expects confidence in safety to increasingly set the pace of AI progress. The CEO said that OpenAI is optimistic about the alignment work it is doing, and it remains committed to ensuring that frontier capabilities remain widely available.